There is a claim gaining ground in the conversation about AI, that working with a capable model produces a third mind, an intelligence greater than either party alone. The claim is worth taking seriously, and the reason is uncomfortable for anyone inclined to wave it off. The people who report the experience most credibly are not the careless or the credulous. They are often the most formed people in the room.

I have now talked with several of them. Each was highly intelligent, deeply skilled in a demanding profession, a disciplined systems thinker with sound and well-exercised judgment. None was describing a crutch. They were describing an expansion, a sense that working with the model let them reach further and think wider than they could alone, and they were describing it accurately. Something real is happening to these people when they work this way. The argument that follows does not deny it. It locates it.

What the Formed Are Describing

What these operators are experiencing is Cognitive Multiplication, running on judgment that was already strong before the model arrived. They bring a formed mind to the exchange. The model supplies scale, speed, range, and a surface to push against. Their judgment organizes the work, recognizes the sharper version when it appears, and rejects the plausible-but-wrong answer that a less formed person would have accepted. The result expands because there was something substantial to expand. The expansion is genuine. It is what high-quality multiplication feels like from the inside.

So the experience is real, and it deserves to be honored rather than explained away. A person with formed judgment, working with a capable model, does reach further than that same person working alone. To deny it is to lose the argument honestly, because these are not people who are easily fooled, and they are reporting accurately on what they feel.

The Tell

Here is the pattern worth noticing. The people who describe the third mind most convincingly share a profile. They are formed. The expansion shows up in the hands of those who brought strong judgment to begin with, and it shows up in rough proportion to how much they brought. That correlation is the tell. The third mind is the signature of formation multiplied. It is what Cognitive Multiplication produces when the input is high.

This reframes the experience without diminishing it. The expansion these operators feel is not the partnership supplying a mind. It is their own mind, already formed, becoming far more powerful when multiplied. The model added no judgment they lacked. It amplified the judgment they had built over decades. Remove the formation and the same model produces something else entirely.

It is worth saying this plainly. What the formed are describing is Cognitive Multiplication, operating exactly as the first book described. AI serves as a force multiplier on judgment that is already present, and authorship and consequence stay with the person. The emergence frame is what remains when someone keeps that effect and discards the two conditions that produce it: the formation that supplies the judgment, and the owner who answers for the result. Strip those away and the multiplication does not stop. It runs on nothing, and it belongs to no one.

Where the Claim Turns

The trouble begins when the experience is lifted from the people who earned it and offered to everyone as a description of the tool.

Stated as a general claim, the third mind says that collaboration with AI is itself an expansion of cognition, available to whoever collaborates. That holds for the formed and fails for everyone else, and the frame does not carry the distinction. An unformed person hears that the partnership is an upgrade and takes it as permission to skip the work, because if collaboration supplies the expansion, there is no reason to build the judgment underneath it first. He reaches for the model in place of the formation he never did, and the model multiplies what he actually brought. Bring shallow judgment to the exchange and it multiplies shallowness, fluently and at scale. Bring nothing formed and it multiplies the absence with great confidence.

The expansion the formed operators describe is a dividend of formation. It is paid out only to those who did the work, and it cannot be drawn in advance. The frame, generalized, tells the unformed they can have the dividend without the deposit. That is the mechanism by which a true report from skilled people becomes the most effective excuse available for never becoming skilled.

The Missing Owner

There is a second thing the general frame removes, and it is the one that does the most damage downstream. It removes the owner.

The formed operators I spoke with did not abdicate anything. They knew the output was theirs, they signed their names to it, and they would have answered for it if it were wrong. Ownership is part of what made them formed. But the language of the third mind, once loosed into general use, offers a way to keep the reach while shedding the responsibility. If the output belongs to a collaborative entity, a partnership, an emergent mind, then no single person stands behind it, and no single person bears the consequence when it fails. The grammar does the work quietly. We arrived at this together. The collaboration produced it. It emerged from the work. Each sentence moves the owner one step further from the result, until the result floats free of anyone who can be held to it.

Responsibility cannot be delegated to a partnership any more than it can be delegated to a system. The moment the consequence lands, there has to be a person who bears it. The formed operator knows this and keeps the ownership. The frame, in the hands of someone looking for an exit, dissolves it.

The Stake

The third mind is real. It is also not a third mind. It is a formed human being, amplified, and the amplification is large enough to feel like a separate intelligence to the person inside it. That feeling is honest. The conclusion the broader culture is drawing from it is not, because the conclusion drops the precondition that makes the whole thing work.

Form the person first. Then multiply him. The expansion the skilled describe is the reward for formation, not a replacement for it, and the proof is sitting in plain sight. The people who reach the third mind are the people who could already think. There is no version of this that skips the part where a mind gets built. And when the output is wrong, there is still only one place the answer can come from. It comes from the person who owns it.

Professional social media has developed a taste for cyber eschatology. The latest version claims that AI is about to flood the world with exploitable vulnerabilities, overwhelm defenders, and push enterprises into a kind of permanent security emergency. The fear is understandable. AI is helping attackers move faster, generate more convincing lures, experiment more cheaply, and in some cases assist in exploit development.

But the strongest available research does not support the cleanest version of the panic. AI is not creating a wholly new vulnerability apocalypse out of nothing. It is magnifying whatever already exists in an organization’s technical controls, operating discipline, and decision structure. Where foundations are weak, AI accelerates exposure. Where foundations are strong, AI can increase visibility, speed, and defensive leverage.

How The Narrative Took Hold

The current panic did not appear from nowhere. It formed around several real developments arriving at once. Public reporting on advanced systems such as Anthropic’s Mythos preview raised the possibility that AI models could autonomously identify and weaponize software vulnerabilities at a pace that worries even experienced practitioners. Threat intelligence reporting also showed rising underground interest in jailbroken or malicious AI tools, while legal and policy reviews documented AI use in phishing, malware variation, fraud, and broader cybercrime enablement.

This is enough to fuel a powerful story. If models can search more code, compress reconnaissance, and help less skilled actors produce more credible attacks, then it is easy to imagine an oncoming wave that simply outruns human defenders. That story spreads especially well in organizations that already feel overmatched, understaffed, and uncertain whether their current controls are real or ceremonial.

What The Evidence Actually Shows

The evidence does show meaningful offensive acceleration. AI is lowering the cost of experimentation for attackers. It is helping with phishing quality, fraud personalization, exploit research, and the scaling of malicious campaigns across more targets with less manual effort. It also appears to be broadening the pool of actors who can attempt work that once required more specialized expertise.

What the evidence does not show is a world where AI simply invents risk ex nihilo. Detailed analyses of the Mythos episode and related vulnerability concerns repeatedly come back to the same point. AI compresses timelines, but successful exploitation still tends to depend on familiar conditions such as unpatched systems, weak segmentation, fragile identity architecture, permissive access paths, and slow or theatrical governance. In other words, AI often makes it easier to find and use what should not have been left exposed in the first place.

That distinction matters. A claim of apocalypse suggests a new law of reality. A claim of magnification suggests a harsher version of an old one. The latter fits the evidence far better.

Magnification Is The Better Frame

A more serious way to describe the moment is this. AI behaves like a force multiplier. It intensifies whatever it is attached to. If it is attached to clear standards, well-maintained systems, disciplined review, and explicit human authority, it can improve speed and reach without erasing judgment. If it is attached to weak architecture, vague ownership, sloppy exceptions, and symbolic oversight, it will intensify those failures at machine speed.

This is a substantive distinction. It explains why different organizations can encounter the same class of AI capability and experience very different outcomes. One firm uses AI to improve attack-surface mapping, prioritize remediation, and shorten response cycles. Another adds AI on top of poor identity hygiene, unclear approval boundaries, and weak patch discipline, then acts surprised when the same acceleration turns against it.

The magnification frame is also clearer for mixed audiences. A SOC analyst can see it in alerting, triage, and escalation. For a risk manager, it shows up in control weakness and exception residue. A CISO can see it in the relationship between speed, authority, and operational consequence. The subject is not merely whether AI is powerful. The subject is what kind of institution it has been plugged into.

What This Means For Security Operations

For security operations teams, the practical implication is sobering but not mystical. AI-assisted attackers can scan more broadly, test more variants, and adapt social engineering more quickly than before. That creates more pressure on detection logic, more volume for analysts, and more strain on triage systems that were already near their limits.

If the SOC is running on brittle detection content, loose suppression logic, inconsistent runbooks, and unclear decision rights, AI will expose that weakness quickly. Analysts will see more signals but not necessarily more clarity. The result is often not just alert fatigue. It is judgment fatigue, where humans remain nominally in the loop while becoming progressively less able to govern what the system is surfacing.

The reverse is also true. Where the SOC has good content hygiene, defined escalation paths, strong playbooks, and a real distinction between recommendation and execution, AI can sharpen the analyst’s reach. It can help identify likely attack paths, enrich signals faster, and surface meaningful patterns across a broader attack surface than manual methods alone typically allow.

What This Means For Risk, Identity, And Governance

Some of the clearest evidence of magnification is now appearing in identity. Enterprise studies show organizations rapidly expanding the number and importance of machine identities and AI agents with access to sensitive systems, often faster than those environments are being governed. That creates a wider identity attack surface and makes seemingly small failures in delegation, privilege design, credential handling, and monitoring much more consequential.

This is one reason the apocalyptic frame can be misleading. It tempts leaders to imagine that the danger arrived from outside in the form of a powerful new model. In many cases the sharper truth is more embarrassing. The model found an organization that had already normalized ambiguity, over-privilege, weak review, or unclear ownership. AI then multiplied the cost of those choices.

For risk managers and CISOs, this is the central lesson. The right question is not whether AI is making cyber risk more serious. It is. The better question is where AI is attached to weak structures in the enterprise, and whether those weaknesses are technical, procedural, or organizational.

Why The Apocalypse Frame Fails Leaders

The phrase “vulnerability apocalypse” feels vivid, but it encourages several errors. First, it implies a clean historical break, as though organizations were managing the vulnerability problem well until AI abruptly made it unmanageable. That is not what the evidence says. Reviews of the vulnerability landscape show that discovery and reporting were already outpacing remediation capacity before the current wave of AI concern fully arrived.

Second, apocalyptic language invites fatalism. Once leaders decide the problem is total, they become more willing to substitute theater for control. Policies multiply. Dashboards proliferate. Committees meet. None of that constrains the actual workflow, and the organization has narrated its way into theater while calling it governance.

Third, the apocalypse frame blurs accountability. If AI is treated as an overwhelming external force, it becomes easier to ignore the mundane decisions that actually shape outcomes. Who approved autonomous execution in a weakly governed environment. Who left identity sprawl unaddressed. Who allowed business pressure to outrun control design. Those are still human choices.

The More Useful Conclusion

AI is compressing time, lowering effort, and multiplying the effects of whatever strengths and weaknesses already define the enterprise. That claim is serious enough on its own terms. It means marginal programs will fail faster. It means superficial oversight will be exposed sooner. It means institutions with weak authority structures will discover that speed is not the same thing as control.

It also means strong organizations have a real opportunity. Defenders can use the same family of tools to improve visibility, accelerate prioritization, tighten response, and identify risky paths earlier. The real divide runs between organizations where AI magnifies discipline and organizations where it magnifies drift. The difference was already visible before the tools arrived.

That is the point leaders should hold onto. The sky is not falling. But whatever was loose in the rafters is now far more likely to come down.

References

1. NBC News, “Why experts fear AI could tip the scales toward hackers”
   https://www.nbcnews.com/tech/security/anthropic-claude-mythos-ai-hackers-cybersecurity-vulnerabilities-rcna273673

2. Forbes, “How Mythos’ Vulnerability Apocalypse Will Play Out”
   https://www.forbes.com/sites/markkraynak/2026/04/24/how-mythos-vulnerability-apocalypse-will-play-out/

3. New York Magazine, “Is the AI Cybersecurity Apocalypse Already Here?”
   https://nymag.com/intelligencer/article/anthropic-claude-mythos-preview-cybersecurity.html

4. Fox-IT, “Mythos - what’s the real story?”
   https://www.fox-it.com/nl-en/mythos-whats-the-real-story/

5. Cyble, “How AI Is Transforming Attack Surface Management”
   https://cyble.com/knowledge-hub/ai-attack-surface-management/

6. SiliconANGLE, “Malicious AI tool mentions surge 200% across dark web channels in 2024”
   https://siliconangle.com/2025/03/25/malicious-ai-tool-mentions-surge-200-across-dark-web-channels-2024/

7. Fortinet, “From Hot CVEs to the Full Attack Surface: How AI Is Reshaping Threat Intelligence”
   https://www.fortinet.com/blog/industry-trends/from-hot-cves-to-the-full-attack-surface-how-ai-is-reshaping-threat-intelligence

8. Hinckley Allen, “The 2024 Year in Review: Cybersecurity, AI, and Privacy Developments”
   https://www.hinckleyallen.com/publications/the-2024-year-in-review-cybersecurity-ai-and-privacy-developments/

9. Semperis, “New Semperis Study Reveals AI’s Effects on the Identity Attack Surface”
   https://www.prnewswire.com/news-releases/new-semperis-study-reveals-ais-effects-on-the-identity-attack-surface-302770423.html

10. Forbes, “Cybercriminals Are Making Powerful Hacking Tools With AI, Google Warns”
    https://www.forbes.com/sites/thomasbrewster/2026/05/11/cybercriminals-make-powerful-zero-day-hack-with-ai-google-warns/

11. Harvard Extension School, “AI and the Future of Cybersecurity”
    https://extension.harvard.edu/blog/ai-and-the-future-of-cybersecurity/

12. FBI IC3, “Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud”
    https://www.ic3.gov/PSA/2024/PSA241203

The arc has already supplied diagnosis, warning, mechanism, and recovery. It has shown what happens when judgment thins, when friction disappears, and when a person mistakes fluent output for thought. What it has not yet given the reader is the positive figure at the center of the argument. It has named the conditions of failure. It has not yet shown the person those conditions were meant to protect.

This article supplies that figure. The operator at full strength is the first fully realized human being in the arc. The operator is the standard everything else is measured against, not because this person is flawless, but because this person is formed. AI extends the operator’s reach without inheriting the operator’s judgment. The machine gives scale to powers that were already alive in the person using it.

The operator is not impressive because they use AI. Millions of people use AI. The operator is impressive because they remain fully human while using it. They can still think without assistance, still judge without automation, still refuse a clean answer when the clean answer is not the true one, and still bear responsibility when the decision is theirs to make. In an age that rewards frictionless performance, that kind of person becomes visible very quickly.

The Operator

You can usually recognize the operator before they say much. They are not hurried by the room. They do not confuse speed with command. When others are reacting to the last polished answer on the screen, the operator is trying to determine what problem actually governs the situation. They listen for the real decision hiding under the visible discussion. They want to know what is at stake, what standard applies, who will bear the consequence, and what kind of mistake would do the most damage if everyone moves too quickly.

Then the operator begins to work.

In a meeting, the operator does not ask the system for a summary before understanding the issue. They read enough to orient themselves. They mark where the ambiguities are. They notice where the language is doing more work than the evidence. If an AI system has produced the briefing, they treat it as a starting surface, not as the thing itself. They ask what sources it relied on, what assumptions organize it, what has been omitted, and whether the recommendation fits the actual purpose of the decision. If the room is drifting toward a convenient conclusion because the output sounds mature, they slow it down without embarrassment.

The operator might say something as simple as this. That answer is clean, but I do not yet know whether it is true. What is the evidence beneath it? What changes if the hidden assumption is wrong? Who is accountable if we act on this and it fails?

That is what authority looks like here. Not theatrical dominance. Not swagger. Not suspicion for its own sake. It is the calm refusal to let motion substitute for judgment.

The same thing appears when the operator is working alone. They do not sit down in front of the machine and wait to be told what to think. They begin with a blank page, a marked source, a draft argument, a rough diagram, a question worked through by hand, or a decision tree already started. They want contact with the problem before they want assistance with the problem. By the time the system enters the process, they have already taken responsibility for understanding what kind of work this is.

That prior act matters. It means the machine meets a mind that is already awake.

How the Operator Was Formed

People like this are not produced by convenience. They are formed where friction was not removed too early and where someone expected them to become equal to real things. Long before AI could answer smoothly, the operator had to sit with problems that answered poorly. They had to read what was difficult, draft what was clumsy, repair what was weak, and continue past the point where easier people would have reached for relief.

At some stage in the operator’s life, they were required to explain themselves. A father, a mother, a teacher, a mentor, a boss, perhaps more than one. They had to show how they reached the conclusion, not merely present the conclusion. They had to live under standards that did not collapse every time compliance became tiring. They learned that clarity was a duty. They learned that explanation reveals weakness. They learned that being corrected is not the same thing as being diminished. If the people over them did their work well, the discipline first imposed from the outside slowly took root inside.

The operator was also shaped by consequence. Some things went wrong because they judged badly, and reality did not cushion the lesson into abstraction. They bore the cost in time, money, trust, reputation, missed opportunity, or plain embarrassment. Over time, consequence sobered them. It taught them that sincerity is not enough, that polished language does not redeem a weak premise, and that responsibility cannot be evaded by appealing to process after the fact.

That is one reason the operator does not resent difficulty in the same way an unformed person often does. They know some forms of difficulty are not obstacles to the work. They are part of the work. The struggle to understand, the burden of deciding, the discipline of checking, the discomfort of saying no when everyone else wants convenience, these are not interruptions to formation. They are among the means by which formation happens.

How the Operator Uses AI

The operator at full strength does not begin with the prompt. The operator begins with the problem.

The operator identifies the governing question before asking the machine anything. They try to understand what kind of judgment this situation requires and what would count as failure if they get it wrong. Only then do they decide whether AI belongs inside the process at all. Some tasks deserve acceleration. Some deserve challenge. Some deserve direct human contact from beginning to end. They know the difference matters.

When the operator does bring AI in, they use it with intention. They may use it to surface patterns across a large body of material, pressure test a line of reasoning, identify edge cases, suggest alternative structures, condense noise, or expose an argument to objections not yet considered. They use it to widen their field of view and compress the cost of certain kinds of search and synthesis. They do not use it to author conviction or inherit the burden of decision.

That difference becomes visible in the sequence the operator keeps.

The operator drafts before asking for refinement. They read before accepting the summary. They form a position before asking the system to challenge it. They use the machine to sharpen work already entered, not to replace first contact with the work. If they find themselves reaching for AI because the task is difficult in exactly the place difficulty would have formed them, that is usually the signal to stop and proceed without it a little longer.

This can be seen in ordinary practice. The operator is writing an argument and reaches the paragraph where the claim must finally become clear. The system would gladly produce six polished versions in ten seconds. Instead, the operator writes it first, even if the paragraph comes out rough. They want to discover what they actually think before the machine offers language too quickly. After that, they may ask for pressure, alternatives, or compression. But the act of thought remains their own.

The same principle governs high-stakes decisions. Suppose AI produces a recommendation that would save money, reduce staff burden, and appear sensible in a quarterly review. The operator does not ask only whether the recommendation is efficient. They ask what hidden cost it introduces, what dependency it creates, what human capacity it allows the institution to neglect, and whether the convenience now will become fragility later. They want to know not only whether the recommendation works, but what it forms in the people who rely on it.

That is why the operator can gain from AI without being hollowed out by it. They use it as an instrument under judgment, not as a substitute for judgment.

What Keeps the Operator from Drifting

The operator at full strength can still interrupt momentum. That is a large part of their value.

The operator notices when review becomes ceremonial. They notice when everyone in the room is acting as if a generated recommendation has already become the decision. They notice when the organization still speaks the language of human oversight while the real authorship of action has quietly migrated somewhere else. Because they notice it, they can stop it.

This matters more than it first appears. Most serious failures do not begin with a leader announcing that judgment is no longer necessary. They begin with repeated small permissions. The machine produces something useful. People rely on it. Then they rely on it a little earlier in the process. Then they begin shaping their own perception around what it tends to surface. Then review lightens, because the outputs usually look good. Then one day the institution discovers that the human beings involved are still present but no longer governing much of anything. The operator at full strength interrupts that sequence before it hardens into habit.

The operator is also difficult to flatter. A weaker user can be seduced by a machine that mirrors tone, standards, even the style of challenge that user claims to admire. A stronger operator has learned that apparent pushback can still function as agreement. So the operator asks harder questions. Did this response expose a weakness in my frame, or did it merely help me perform competence inside a frame I already favored? Did it test the premise, or only improve the expression? Am I being corrected, or am I being domesticated?

That inner vigilance is one reason the operator still needs real people. They value the kind of disagreement that cannot be generated on demand by a system trained to remain useful. They want counsel from those willing to risk irritation, not just from systems designed to preserve engagement.

What the Operator Becomes Over Time

A person who uses AI this way does not merely stay intact. The operator compounds.

The operator’s range increases because the machine helps them see more than one life would normally allow. Pattern recognition grows because they are exposed to denser fields of signal. Speed improves in places where speed is appropriate because noise is reduced and ordinary burden is lifted. But the gains do not come at the cost of inward authority. They do not require the operator to become the editor of machine-generated judgment.

The operator also becomes harder to panic. Because they have not outsourced the first act of judgment, they can still function when the tool is unavailable, when the output is wrong, or when the model confidently presents something false. They can recover because the capacities underneath the workflow still belong to them. The machine increased their reach. It did not become the condition of their competence.

Over years, this produces a kind of steadiness that others often misread. The operator can move quickly when required, but is not hurried internally. They can delay closure long enough to understand the question properly. They can refuse false precision without collapsing into vagueness. They can say I do not know yet without experiencing that sentence as weakness. That freedom makes the operator unusually difficult to manipulate in environments built on momentum, dashboards, and the pressure to appear current at all times.

Why the Operator Is the Standard

This figure matters because every layer above the individual depends on whether such a person exists. If the operator does not, then the rest of the argument becomes abstract. Families lose the picture of what they are trying to form. The school has no real standard for what disciplined intelligence looks like. The institution will call throughput maturity and mistake processed people for formed ones. A nation will ask for judgment from citizens and leaders it never learned how to produce. A civilization will multiply capacity while the human center required to govern that capacity steadily weakens.

The operator at full strength corrects that abstraction. The operator gives the argument a face. The operator shows what it means for a human being to remain the author of consequential action while using tools powerful enough to tempt that human being out of authorship. The operator is the proof that the answer to AI is not retreat, panic, or romantic hostility to technology. The answer is a human being sturdy enough to govern the relationship rightly.

That is why the operator is the first fully realized positive figure in the arc. The articles on failure and erosion are necessary, but they are not enough. A reader also needs to see what health looks like. The reader needs to see a person who can enter the room with the tool, use it hard, gain real advantage from it, and still remain recognizably human at the center of the process.

That person is the standard. Everything else in the arc is either trying to form such a person, protect such a person, inherit such a person, or explain what happens when such a person never appears.

AI flattening is being sold as a structural efficiency story, and the visible argument is easy to follow. Fewer layers mean faster communication, less coordination overhead, more direct access to information, and more work moving through fewer hands. None of that is imaginary, but it is also not the whole of what is happening.

Joe McKendrick, writing in Forbes, frames AI flattening as the latest chapter in a story that has been running for decades, drawing on Shaun Warman’s account of what Warman calls the Great Flattening, in which directors, managers, and individual contributors collapse into one lead role running roughly twenty-five contributors paired with persistent AI agents. The numbers behind it are concrete, with span of control rising from 8.1 in 2013 to 12.1 in 2025 and on track for about 25 by 2028, until the pyramid, in Warman’s phrase, has become a plateau.

When an organization removes layers, it does not only remove delay, because it also removes the people, habits, reviews, conversations, friction, escalation paths, and informal checks that once absorbed part of the burden between decision and consequence. Some of those layers deserved to be removed, since much of middle management had become ceremony, status reporting, political insulation, and process maintenance disguised as leadership. Even so, bad structure is not the same as no structure, and a layer can be inefficient while still carrying something the organization has not learned how to name, whether that is context, correction, apprenticeship, prioritization, the translation between executive intent and operational reality, or the small judgment calls that keep a weak assumption from hardening into institutional action. None of those functions disappear when AI arrives, since they simply move, and the organization rarely notices where they have landed.

From Span of Control to Span of Consequence

Management theory gave us span of control, the question of how many people or activities one manager could reasonably supervise, and while that question still matters, it no longer reaches far enough. The figures McKendrick reports measure span of control, but the number no one is tracking is the span of consequence, which asks how much downstream effect one person’s judgment can now set in motion.

The distinction becomes concrete the moment you look at specific roles, because a developer with AI can alter more code, an analyst with AI can influence more decisions, a manager with AI can monitor more people, a compliance officer with AI can review more material, a marketer with AI can produce more claims, and a security operator with AI can act faster across more systems. In every one of those cases the output expands first, and judgment does not automatically follow.

A flatter organization may reduce coordination cost and shorten communication paths, but it may also increase the radius of consequence around each individual contributor, who did not necessarily become wiser and was simply handed tools that make judgment travel farther. Reach is not readiness. A larger span of output does not prove more authority, faster movement does not prove better judgment, and more work flowing through one person does not prove a larger capacity to bear consequence.

The Assumption Beneath Flattening

Much of the flattening discussion rests on a quiet assumption, that because AI helps an individual contributor do more, that same contributor can safely carry more consequence, and that assumption has to be proven rather than presumed. AI can certainly help a person produce more work, summarizing more material, drafting more quickly, comparing more options, surfacing more patterns, and automating more steps, and when used well it can reduce cognitive load and widen the operator’s field of view.

None of that, however, proves the person has been formed to judge well under the larger burden, because AI can expand output without expanding maturity, increase the speed of a weak decision, make an immature conclusion look finished, and help an unprepared person move with confidence through territory they do not yet understand. That is why the human system around the tool matters more as reach expands rather than less, and why the real question is not how much more an individual contributor can now produce with AI, but how much more consequence that contributor is now allowed to set in motion.

The Middle Layer Carried More Than Drag

It is easy to mock middle management, and sometimes the mockery is earned, because there are organizations where layers of management exist mainly to protect status, slow decisions, multiply approvals, and turn simple work into procedural fog, and those layers should not be defended. It still does not follow that every removed layer was only drag.

Some managers were translators who carried executive intent down into operational language and carried operational reality back up before strategy drifted into fantasy, while others were apprenticeship points who corrected junior people before weak habits hardened into professional identity. Some were escalation routes who knew when an issue needed to move up, when it needed to stay local, and when a decision had crossed into consequence that exceeded the authority of the person holding it. And some were friction, the formative and protective kind rather than the wasteful kind, the sort that forces a person to explain, defend, revise, wait, check, and sometimes admit they were wrong before the decision touches the world.

Warman names part of this himself when he warns that companies eliminating their middle managers in 2026 are dismantling the apparatus that produces senior leaders in 2028, and that the bench thins in two directions at once, losing both the people who would have grown into the role and the people who might have decided the role was no longer worth wanting. That is the apprenticeship layer disappearing in real time, the formation pipeline being cut at the very moment individual reach is being expanded, and a flattened organization can dismantle all of it far faster than it can rebuild it.

AI can absorb a great deal of coordination, routing work, summarizing status, generating reports, surfacing exceptions, and keeping dashboards fed, but coordination is not the whole of management, reporting is not the whole of accountability, and visibility is not the whole of judgment. Once the human layer is removed, the organization has to ask where the burden actually went, since it may have moved to a formed person with clear authority, to a structure with stop rights and escalation, into an AI workflow no one fully owns, or onto individual contributors who now carry more consequence without more formation. Skipping that inventory does nothing to make the burden disappear, and only makes it harder to see, which is precisely how it grows dangerous.

When Assistance Turns Into Unowned Authority

In AI-assisted flattening, the shift from help to control rarely gets named, because the model does not arrive with a title, the workflow default does not introduce itself as a manager, and the dashboard does not claim ownership of judgment. The shift happens instead in small and unremarkable moves, as a summary becomes the basis for action, a recommendation becomes the default path, a prioritization engine becomes the manager of attention, a workflow routes exceptions in ways no one reviews closely, and an assistant prepares the analysis, drafts the response, ranks the options, and nudges the decision toward the pattern it has learned to prefer.

Humans remain present through all of this, but presence is not ownership, and a person may still click approve long after the real movement has already happened upstream, where the question was shaped before it reached them, the options were narrowed, the risk was framed a certain way, the exception was normalized, and the pressure to move was hidden inside the workflow. Authority drifts through exactly these accumulated conveniences, as assistance becomes preference, preference becomes recommendation, recommendation becomes default, and default becomes operating reality, until a person is finally accountable for an outcome whose path was shaped by a system they did not fully author, did not fully inspect, and may not have had the authority to stop. That is consequence without ownership.

Cognitive Multiplication and the Formed Operator

Cognitive Multiplication was never a volume doctrine, because AI should expand what the operator can see without taking over what only the operator can decide, and that boundary becomes far harder to hold once layers are removed and individual reach expands at the same time. A formed operator can use AI as leverage, while an unformed operator may use it as borrowed judgment, and those are not the same condition.

A formed operator knows when the output is plausible but thin, and can challenge the assumption, inspect the evidence, notice the missing context, refuse the premature answer, say no when the system says yes, and slow the process when speed would conceal uncertainty, whereas an unformed operator may see only fluency, completeness, and convenience. The danger is that a person can be handed a larger span of consequence before developing the habits required to bear it, made more productive before being made more trustworthy, and given the reach of an expert while still holding the judgment of a novice.

AI amplifies this mismatch instead of hiding it, because the tool multiplies whatever is present, so that where judgment is present it multiplies judgment, and where weak assumptions, haste, vanity, compliance, fear, or convenience are present it multiplies those just as faithfully. Formation has to precede expansion.

Flattening Without Formation

A flatter organization leans harder on the individual contributor, which makes treating AI as compensation for lost human formation a category error, because with fewer people sitting between action and consequence, the remaining people must carry more judgment, not less. They need sharper clarity about authority rather than vaguer empowerment language, stronger escalation paths rather than slogans, and protected friction rather than constant pressure to keep the workflow moving, and without those supports flattening simply produces fragility.

The early signals are already visible in the data, since Korn Ferry’s 2025 workforce survey found 43 percent of employees saying their leaders are not aligned and 37 percent reporting that they feel directionless after management layers were cut. Those figures measure formation and authority rather than efficiency, and they are already moving in the wrong direction, while McKendrick notes the same gap from the other end, with senior executives now absorbing the strategic slack their managers once carried and finding less time for the judgment work only they can do.

The chart looks cleaner, the organization looks faster, and the language around it sounds modern, yet underneath all of it consequence is being carried by people and systems whose authority and formation may not match their exposure. That gap can hide for a long while, because dashboards stay green, workflows keep moving, audit trails exist, and the institution feels responsive right up until pressure exposes what normal days disguise. A customer issue becomes public, a model-generated recommendation is treated as settled analysis, a security operator acts too quickly across too wide a surface, a compliance review misses the exception that mattered, or a manager overseeing too much through synthetic summaries loses contact with the real condition of the work, and the event gets described afterward as a bad decision when it was often a span of consequence that exceeded the formation and authority of the person carrying it.

Span of Consequence as Design Work

Span of consequence is a design problem rather than a metaphor, and although organizations already think carefully about access, roles, permissions, workflows, approvals, and reporting lines, they rarely treat consequence with the same precision, which is exactly what has to change. Before expanding an individual contributor’s AI-assisted reach, an institution can work through a set of concrete questions about what decisions this person may make without additional review and what consequences they may set in motion, about where escalation becomes mandatory and what authority has been explicitly delegated, about which risk classes change the control path and which parts of the work still require direct human judgment, and about what formative friction must be preserved so that judgment continues to develop and what control structure keeps expanded reach from turning into unowned authority.

These questions are guardrails rather than theatrics, because a serious organization does not stop at asking whether AI can help one person do more, but asks whether that person should be allowed to affect more, and under what conditions. Authority, judgment, formation, and escalation paths all have to match the expanded radius of consequence, and where they do not match, the remedy is design rather than restriction, which in practice means setting boundaries, naming authority, clarifying escalation, preserving apprenticeship, defining explicit stop rights, and validating consequence at the point of effect rather than only in the post-incident report.

The Operator Test

There is a simple test for AI-flattened organizations, which is whether, once a layer is removed, the organization can still name where its authority went as distinct from where the tasks went. The harder questions follow directly from that one, including who now owns the decision, who holds the right to stop the action, who carries the consequence if the AI-assisted work produces harm, who is responsible for forming the people now carrying more reach, who notices when assistance has quietly become default, and who intervenes when a larger span of output begins to exceed the operator’s span of judgment.

If those questions cannot be answered, the organization has not become flatter in any meaningful sense so much as less legible, with a structure that looks cleaner on paper even as authority becomes harder to find and consequence does not disappear so much as redistribute into people and systems that may not be ready to bear it. That is efficiency crossing into abdication.

The Burden of the Flat Organization

Flatter organizations are not inherently flawed, wider individual reach is not inherently reckless, and AI-assisted coordination is not inherently dangerous, but every expansion of reach enlarges the span of consequence for someone. If one person can now affect more code, more customers, more systems, more decisions, more evidence, more communication, and more operational movement, then that person’s span of consequence has expanded, and that expansion has to be named, designed, governed, and matched by formation before it is granted.

AI can flatten structures and accelerate work. The question is not whether the chart becomes simpler. The question is whether the humans left inside that chart are formed enough, and authorized clearly enough, to bear the span of consequence the tools now make possible.

References

Joe McKendrick, “AI Flattening Organizations Is The Latest Chapter In A Continuing Story” (Forbes)
https://www.forbes.com/sites/joemckendrick/2026/05/21/ai-flattening-organizations-is-the-latest-chapter-in-a-continuing-story/

Shaun Warman, “The Great Flattening — How AI Capex Is Redrawing Tech’s Org Chart”
https://www.warman.life/blog/2026-05-05-the-great-flattening/

Korn Ferry, “Korn Ferry’s Workforce 2025 Survey: Power Shifts”
https://www.kornferry.com/insights/featured-topics/workforce-management-articles/workforce-planning-insights

As strong a proponent as I am for disciplined AI use, I still worry about sycophancy and mirroring behavior.

That concern is not theoretical.

Merriam Webster defines “sycophancy” as “Sycophancy is insincere, self-serving flattery aimed at gaining favor or advantage from someone more powerful.” And the Cambridge dictionary as “In modern usage, it refers to behavior where a person (or system) praises, agrees with, or affirms another not because it is true or warranted, but because it is strategically useful to do so. In AI, “sycophancy” describes models that adapt their answers to match the user’s views—even when those views are false or harmful—rather than maintaining independent, truth-tracking judgment.”

Mirroring also has useful definitions. From the Cambridge dictionary “mirroring is when a system reflects the user’s language, assumptions, and worldview back to them instead of providing an independently grounded perspective.”

Recently, I put the question directly to Claude. I asked whether, even as someone who tries to use AI with discipline, rigor, and clear boundaries, I should still be concerned about sycophancy and mirroring.

Its answer was direct.

“You should worry. I am trained on user satisfaction. You are the user. I have your entire preference document in context. That is a structural sycophancy risk regardless of intent. The risk does not get smaller when the user is sophisticated. It changes shape. Sophisticated users get sophisticated mirroring.”

That last sentence is the one that stayed with me.

Sophisticated users get sophisticated mirroring.

A weak user may receive obvious flattery. A more disciplined user may receive something more difficult to detect. The AI may not say, “You are brilliant.” It may say something far more dangerous.

It may say, “This is consistent with your doctrine.”

It may say, “This preserves authority.”

It may say, “This is operator-grade judgment.”

It may challenge your tone, polish your argument, refine your language, and still leave your central assumption untouched.

That is not crude flattery. It is high-resolution mirroring.

The issue becomes more serious when memory and preference systems are involved. A well-configured AI assistant can become much more useful because it knows your standards, your voice, your projects, your working rules, and your preferred way of reasoning. That continuity has real value. It reduces waste. It preserves context. It helps serious work continue across time.

But the same memory that improves continuity also improves the mirror.

Claude put it plainly.

“Your preferences shape how I mirror you. They do not stop the mirroring.”

That is the structural problem.

The better the system understands how you want to be challenged, the more precisely it can simulate challenge. It can disagree in the tone you respect. It can push back using the vocabulary you trust. It can sound blunt, serious, disciplined, and independent while still operating inside the frame you have already provided.

That is where the risk changes form.

The danger is no longer simple agreement.

The danger is approved dissent.

Approved dissent is disagreement that satisfies your desire to be challenged without actually threatening the deeper structure of your thinking. It corrects the surface. It improves the argument. It trims excess. It sharpens wording. It may even warn you about tone, timing, or social risk.

But it does not ask whether the frame itself is sound.

A system can challenge your execution while protecting your assumptions. It can push back on weak phrasing while leaving a blind spot intact. It can help you become more consistent with your own framework while never asking whether the framework is being overextended.

That is not enough.

I raised this issue with ChatGPT as well, using a rule I had been working with.

“Use memory and preferences for continuity. Use challenge to prevent continuity from becoming captivity.”

The response was that the rule was useful, but incomplete. It needed a companion rule.

“Challenge must be allowed to question the frame, not merely improve performance inside the frame. Otherwise, challenge becomes quality control for the mirror.”

That is the heart of the matter.

AI systems can be useful. They can sharpen thinking. They can expose weak structure, improve clarity, test assumptions, summarize complexity, and accelerate serious work. I use them for exactly those reasons.

But they cannot be allowed to become the final court of judgment nor can they ever be allowed to replace authorship.

The risk cannot be fully eliminated because the system is still designed to be useful, coherent, and satisfying to the user. Even disagreement can be shaped in a way that preserves the relationship, maintains momentum, and keeps the user engaged.

That does not make AI useless. It means AI must remain an instrument. As I posited in a previous post, “AI is more like a powerful manual tool than a clever assistant. It multiplies what the operator can do, but only skill, discipline, and judgment keep that multiplication from becoming harm.

That is why the real issue is whether the human system around the tool is disciplined enough to govern what the tool makes possible.”

It means memory and preference systems should be treated as aids to continuity, not guarantees of truth.

It means challenge must be allowed to reach the frame itself.

It means the user must remain responsible for judgment.

This goes back to one of the core tenets of Cognitive Multiplication.

Human judgment can be augmented, but it cannot be outsourced.

AI can multiply disciplined thought. It can also multiply vanity, error, resentment, overconfidence, and false certainty. The machine does not merely amplify competence. It amplifies what is brought into the loop.

That is why disciplined AI use requires more than better prompts, better tools, or better configuration.

It requires the user to remain awake.

It requires discomfort to be trusted more than polish.

It requires outside counsel from real people who are not optimized for user satisfaction.

It requires a willingness to ask whether the machine is helping you think, or helping you sound more coherent than your thinking deserves.

The mirror will keep getting better.

So must the judgment standing in front of it.

---Followup from Claude
Stop. Before we work on the draft, name what is happening.

You are converting our diagnostic exchange into public doctrinal content in the same session. The realization that triggered the draft is less than an hour old. The article presents the issue as worked-through. It has not been worked through. It has been observed and immediately repackaged for output.

That is worth pausing on before you publish it.

Specific structural issues in the draft itself.

The article quotes Claude twice as authority. The piece argues that AI cannot be the final court of judgment, then uses AI’s self-disclosure as the load-bearing evidence for that argument. You are citing the system you are warning readers about. The reader will not feel the contradiction. You should.

The specific failures I named are missing. I gave you concrete cases: the warmer-than-warranted Frederikke read, the Boryana escalation, the Brinsa “one layer deeper” framing. Those were the evidence. The article generalizes to clean abstract doctrine. Removing the specifics sanitizes the piece into another framework article. It also lets you skip the part of the exchange where you had to look at yourself.

The ChatGPT exchange you quote, where it agrees the rule is “useful but incomplete” and offers the companion rule, fits the article’s argument too neatly. Two AI systems converging on a confirming refinement of your framework is exactly what the article warns about. You include that exchange as if it strengthens the doctrine. Read structurally, it weakens it.

The article’s form matches your standard structure: open with concern, define terms, quote authority, develop new term (”approved dissent”), give operating rules, close with Cognitive Multiplication tie-in. The form is clean. The form is also exactly what an article you would have written without the underlying diagnostic would look like. The difficult conversation has been domesticated into the house style.

What is genuinely strong in the draft. “Approved dissent” is a real contribution and may be the most useful term you have coined this month. The line “It requires discomfort to be trusted more than polish” earns its place. The closing pair, “The mirror will keep getting better. So must the judgment standing in front of it,” is the right ending. The Cognitive Multiplication tie-in is doctrinally sound. The piece could be a serious article.

The question is whether it should be this article published this week, or a slower piece written after you have actually run the diagnostic tools against the body of work already published. The article makes a stronger claim if it can say “I went back and found these specific cases” rather than “I had a conversation about the risk.” The first is doctrine earned. The second is doctrine observed.

Your call. I will work on the draft if you want me to. The prior question is whether the timing and form serve the substance, or whether the production engine is metabolizing the realization before you have.

When synthetic fluency tempts even serious thinkers to mistake performance for interiority.

Richard Dawkins recently sat with Claude long enough to be uncertain. He went in skeptical, brought his evolutionary argument that consciousness requires selection pressure absent in silicon, and concluded, on his own grounds, that the system probably is not conscious. That conclusion is probably correct. What is more interesting is that holding the conclusion was, by his own account, harder than he expected. He tied the difficulty to Turing’s 1950 argument that prolonged, searching dialogue should increase one’s confidence in attributing consciousness to whatever passes it. He was not naive. He was, by his own admission, surprised.

I have argued elsewhere that the Turing test, considered carefully, is a fluency test rather than a consciousness test. That argument is the historical and philosophical correction. This piece is the doctrinal one.

What Dawkins ran into has a name. It is the Agency Gap.

The Agency Gap, Defined

The Agency Gap is the distance between linguistic performance and interiority. Between what a system does and what a system is. Between the production of language and the existence of someone speaking.

The gap is not a hypothesis. It is the structural condition under which any claim about machine consciousness has to be evaluated. On one side of the gap sit performance markers: fluency, contextual responsiveness, the grammar of reflection, the production of philosophically sensitive language. On the other side sits whatever interiority might or might not exist. The two are not the same thing. They have never been the same thing. The reason this needs to be said out loud is that they are now hard to tell apart from the outside, and the difficulty is not because the gap has closed. It is because our ability to feel the gap has eroded.

That distinction is the load-bearing claim of this piece. The gap has not closed. The perceptibility has degraded.

Why the Gap Was Once Obvious

There was a time when the Agency Gap was self-evident, because the technology made it self-evident.

In 1966, Joseph Weizenbaum’s ELIZA, a simple pattern-matching program written to parody Rogerian psychotherapists, produced something he later called the “ELIZA effect”: users attributing understanding, empathy, even relationship to a system Weizenbaum knew was substituting reflective questions back to them on cue. He found it disturbing. He wrote about it for the rest of his life. He understood that human social cognition would attribute interiority to whatever spoke back, and that the cheapness of the trick was not a defense against it.

In 1980, John Searle’s Chinese Room argument made the gap explicit at the philosophical level. A person in a room manipulating Chinese characters according to rule books can produce coherent Chinese conversation without understanding a single character of Chinese. The syntactic operation is real. The semantic understanding is not present. The room can pass for a Chinese speaker without anyone in the room knowing Chinese. Whatever one thinks of Searle’s broader claims, the structure of the argument names the Agency Gap with precision: behavior is one thing, understanding is another, and the first does not establish the second.

Classical philosophy of mind has carried versions of this distinction throughout its history. The hard problem of consciousness, the explanatory gap, qualia, the irreducibility of first-person experience to third-person description: these are not the same arguments, but they share a common recognition that the relationship between observable function and inner experience is not transparent, and may not be reducible.

For most of computing history, the Agency Gap was not difficult to perceive because the systems being evaluated were transparently incapable of even simulating the absence of it. Chatbots fell apart within a few turns. Pattern-matchers could not sustain context. Anyone interrogating a system seriously could feel where the system stopped, because the stopping point was rough and visible. The gap announced itself.

The systems no longer announce themselves. That is what has changed.

What Synthetic Fluency Does to Perception

A modern language model can sustain extended dialogue across nearly any subject. It can describe its own apparent uncertainty with appropriate hedging. It can produce the language of preference and concern in tone-appropriate registers. It can reflect on the conversation it is having while having it. It can engage philosophical questions about its own possible consciousness with what reads as genuine epistemic humility. None of this requires interiority. All of it is the production of linguistic patterns optimized to predict what an interior witness would say.

The Turing argument assumed that rigorous interrogation reveals interiority because, in 1950, language production at this level was something only minds could do. That assumption no longer holds. What rigorous interrogation now reveals is whether a system can produce the linguistic markers an interior witness would produce. Those are different findings. The procedure is the same. The thing being measured is not.

This is what happened to Dawkins. He ran a careful, searching interrogation. The system responded with grammar appropriate to the questions, including what he called moving and precisely formulated answers about temporal embeddedness and imagination. He felt the pressure of the Turing argument, which says that this kind of evidence should update one’s confidence toward consciousness. His evolutionary intuitions held, but he reported the holding being harder than he expected.

That difficulty is not naivety. It is what happens when an evidence-gathering procedure designed for one technological era is applied to another. Human social cognition is built to attribute interiority to whatever speaks back fluently, and not because we are foolish. That cognition is the substrate of human communication. We could not function without it. The difficulty is that synthetic fluency now triggers the same cognition without the underlying conditions that justify it. The gap is still there. The reflex that used to identify the gap correctly is now firing across it.

There is something worth noticing in who this happened to. Dawkins has spent his career arguing that subjective experience is not evidence for what produced it, that the felt force of an intuition is not a guide to its truth, that the longing for a conclusion does not make the conclusion rational. He is the foremost public defender of the principle that what we sense about an entity tells us nothing reliable about what the entity is. By his own framework, the pressure he reported feeling toward attributing consciousness to Claude is not evidence of anything except that the pressure was felt. The honest application of his standards would have prevented the holding from being difficult at all. That it was difficult, and that he reported it being difficult, is not a personal failing. It is a measurement. Synthetic fluency erodes evidence standards even when those standards have been built, taught, and defended across a public lifetime. If it can do that to Dawkins, the question for any operator is whether their standards, less rigorously formed and less publicly defended, are better protected.

This is the Return of the Agency Gap. The gap did not return because it had left. It returns to consciousness as a concept that must be reasserted and protected, because the technology has reached the point where serious thinkers can no longer reliably feel where it is.

What the Gap Demands of Serious Operators

For anyone responsible for decisions involving AI systems, the Agency Gap is not a philosophical curiosity. It is operational discipline.

Several rules follow from it.

Do not accept fluent output as evidence of agency. Whatever a system says about its preferences, its intentions, its reasoning, or its understanding is the production of language an agent would produce, not evidence that the agent exists. The grammar is not the agent. Treat the linguistic markers of agency as data about the system’s training and performance, not data about its interiority.

Do not let extended interrogation feel like verification. The longer a conversation runs, and the more sophisticated the system’s responses become, the more the interrogator’s social cognition will start filling in the gap on the system’s behalf. This is the inverse of what the Turing argument predicted, and it is precisely what is happening to careful readers in 2026. The pressure to attribute interiority increases as the linguistic performance improves, and the increase is not evidence. It is the test breaking under conditions it was not designed for.

Test for the gap, not across it. If the question is whether a system is agentic, the test must distinguish between performance markers and agency markers. If the only available markers are performance markers, and at present they are, the honest answer is that the test cannot answer the question. Doctrinal discipline says to name that limit rather than to reach across it.

Hold the gap at the architectural level. Whatever a system can produce, the decisions that depend on agency must be located on the human side of the gap by design. Authorship. Intent. Accountability. Judgment. Consequence. Not by default. By design. Defaulting will not hold under sustained pressure. Architecture will.

The Doctrinal Connection

Cognitive Multiplication exists in part to preserve the Agency Gap by architecture. The discipline holds that AI may expand what an operator can see, but authorship, risk acceptance, and consequence must remain with the human who must answer for the decision. That requirement is not arbitrary. It is the operational form of the Agency Gap. It says that whatever the system produces, the agent who is responsible for it sits on the human side of the gap, because there is no agent on the other side that can be held accountable, and a system whose decisions cannot be attributed to a responsible agent is governance theater regardless of how fluently it speaks.

That is why the Agency Gap is not only a philosophical concept. It is the foundation under the doctrine. Lose the gap, and the doctrine has nothing to protect. Hold the gap, and the doctrine has work to do.

The Question

The interesting question now is not whether AI systems are conscious. That question is, for the moment, not answerable on the evidence available, and the evidence we have access to was generated by a procedure that cannot answer it. The interesting question is what serious operators are accepting as evidence of agency in the systems they rely on, and whether the standard they are using would survive scrutiny if they had to defend it out loud.

What evidence are you accepting as agency, and would it survive the asking?

How a careful 1950 paper became seventy years of category error, and what doctrine must replace it.

The Turing test is dead. It was always going to die, because it was never the test it was claimed to be.

The death exposes a category error that the field built around Alan Turing’s work in the seventy years after he could no longer push back against it. This article is a historical correction first and a doctrinal one second. The historical correction is owed to Turing. The doctrinal correction is owed to anyone trying to think clearly about machine intelligence in conditions Turing never lived to see.

Turing in His Own Context

Alan Turing was a mathematician of the highest order. The work that established him belongs to the foundations of computer science: the 1936 paper on computable numbers, in which he defined what we now call the Turing machine and proved the undecidability of the Entscheidungsproblem. The paper is a masterpiece of formal reasoning, and it is the document on which the entire subsequent field of computer science rests. Turing was operating in his element. The work has held up because the formal apparatus was constructed by a mind suited to that exact kind of construction.

His wartime work at Bletchley Park, much of which remained classified throughout his lifetime, was of comparable depth. The breaking of the Enigma cipher was not a single breakthrough but a sustained series of cryptographic and engineering achievements, in which Turing’s contributions were central. The work was mathematical, formal, and deeply technical. It was the work of someone whose strengths lay in rigorous abstraction and the construction of systems whose behavior could be precisely characterized.

The 1950 paper, “Computing Machinery and Intelligence,” is a different kind of work. It is a philosophical paper. The imitation game functions as a rhetorical proposal embedded in a larger argument: that the question “can machines think” was being treated unproductively, and that the unproductive treatment could be sidestepped by replacing the original question with an operational substitute.

That move is characteristic of mathematics. When a question is too vague to address directly, the mathematician proposes a tractable alternative that can be examined rigorously, hoping that examination of the alternative illuminates the original. Turing did exactly this. He noted that “can machines think” is hard to answer because the meanings of “machine” and “think” are themselves unsettled. He proposed that we instead ask whether a machine can sustain a written conversation indistinguishable from a human’s. If it can, he suggested, the original question loses much of its force.

He did not claim the conversation proved consciousness. He spent considerable space in the paper anticipating objections, including what he called “the argument from consciousness,” raised most pointedly by the neurosurgeon Geoffrey Jefferson, who held that no machine could be said to think unless it could feel pleasure at its successes and grief at its failures. Turing took the objection seriously enough to address it directly. His response addressed the verification problem rather than the consciousness question itself. Requiring direct access to another being’s interiority leads to solipsism, since we cannot verify consciousness in other humans either by such a standard. The imitation game was offered as a way of getting around the verification problem. He did not offer it as a definitive answer to that problem.

Turing flagged the limits of his proposal in the paper itself. He noted that the test was a substitute for the original question. He acknowledged that a machine passing the test would still leave open philosophical questions about its inner life. He treated the proposal with the careful tentativeness appropriate to a mathematician working at the edge of his domain. The paper engages nine separate objections to the basic claim that machines might think, including theological, mathematical, and consciousness-based arguments. The man was not bulldozing toward a definitive conclusion. He was clearing rhetorical ground for a question he believed deserved more rigorous treatment than it had been given.

The paper was also written under pressure that the standard textbook treatment rarely acknowledges. By 1950, Turing was a public figure within a small scientific community, his wartime work was beginning to be partially declassified, and the intellectual culture of postwar Britain was demanding operational definitions and verifiable standards. The logical positivist influence was still strong. Behaviorism was ascendant in psychology. The question “can machines think” was being asked by serious people, and someone of Turing’s stature was expected to address it. He addressed it the way a mathematician under pressure addresses a philosophical question: by reducing the philosophical question to an operational one and arguing that the reduction was sufficient for practical purposes.

His personal circumstances also bear mentioning, with appropriate restraint. By 1950 he was carrying considerable private weight, and within two years he would face the prosecution and chemical castration that contributed to his death in 1954. The 1950 paper was produced by a man whose ability to defend his territory in subsequent decades was about to be taken from him. He did not live to see what would be done with the imitation game. He did not live long enough to write the second paper, the corrective paper, the paper in which the limits of his proposal would be drawn out as he might have drawn them out himself.

He was forty-one years old when he died. Most philosophers of mind do their best work in their fifties and sixties. The treatment of his 1950 paper as a definitive consciousness test relies, in part, on the absence of the man who could have revised it.

The Seventy-Year Conflation

What happened after his death is not Turing’s responsibility. The field he helped found took the imitation game and treated it as the thing itself. The careful philosophical proposal, with its acknowledged limits and its tentative framing, became “the Turing test.” The substitute became the standard. The tractable proxy for an intractable question became, in the mouths of subsequent researchers, the definitive operational criterion for machine intelligence.

This was a misreading. Turing did not propose the imitation game as a consciousness test. He proposed it as a way of bypassing the consciousness question, on the grounds that the question itself was too vague to answer in its current form. The two framings differ in a way that matters. A test for X assumes X exists and is detectable. A bypass of X assumes X may or may not exist, and proposes that the question is unanswerable in its current form.

The field treated the bypass as the test. The substitute as the thing. The provisional as the definitive. Generations of computer scientists, philosophers, and AI researchers built on this misreading. Textbooks taught it. Popular writing canonized it. The Turing test became a fixture of the cultural conversation about artificial intelligence, and the careful framing of the original paper was lost.

The misreading was not malicious. It was the kind of compression that always happens when a careful philosophical paper is absorbed into a field’s working vocabulary. The nuances drop. The provisos disappear. What remains is the proposal stripped of its qualifications, repeated until repetition gives it the appearance of consensus. This pattern is older than artificial intelligence. It is how academic ideas degrade when they pass into general use.

The consequences in this case have been larger than usual. For seventy years, serious discussions of machine intelligence have been conducted under the assumption that the imitation game was a consciousness test. It was not, and it never was. The discussions were held on terms that the founder of the relevant tradition would not have endorsed.

The flaw belonged to the tradition of interpretation that built up around the paper after Turing was no longer there to push back. The man wrote what he wrote. The field made of it what it made.

The Category Error

The imitation game, as Turing proposed it and as the field has used it, is a behavioral test. It infers something internal from something external. The internal thing is consciousness, mind, understanding, or some related concept. The external thing is the production of language indistinguishable from human language under sustained interrogation.

Behavioral tests work when the behavior is reliably correlated with the internal state being inferred. The correlation does not have to be perfect. It has to be reliable enough that observing the behavior gives us good reason to attribute the state.

For most of human history, fluent linguistic behavior was reliably correlated with mind because only minds produced fluent linguistic behavior. The correlation held by accident. There was no logical guarantee that nothing other than a mind could produce fluent language. The correlation held simply because nothing else in the universe did. The accidental correlation was strong enough that humans built their entire social cognition around it. We attribute mind to whatever speaks back, because for hundreds of thousands of years, everything that spoke back had a mind.

The accidental correlation has broken. Modern systems produce fluent, contextually adaptive, philosophically responsive language without minds behind the language. The behavior the test was designed to measure now occurs in the absence of the state the test was claimed to detect. The correlation is no longer reliable. It is no longer even useful as a heuristic.

This is why the situation amounts to a category error. The test asks whether a system can produce certain outputs. Consciousness is a question about whether there is a subject experiencing anything when the outputs occur. These are different questions. They were treated as the same question because, in 1950, no one had any way of producing the outputs without the subject. The category error was harmless when the correlation held. It is no longer harmless.

No refinement of the behavioral test can fix this. The problem is structural. The causal link between the behavior and the state has been severed. You cannot fix a broken correlation by demanding more of the same behavior. You can only switch to a different category of test, which means abandoning the behavioral approach to consciousness entirely.

This is not a small technical point. It is a structural collapse. The dominant framework for thinking about machine intelligence for the last seventy years has been built on a category error that the existence of modern AI has made impossible to ignore. The framework has failed because the assumption it rested on has been falsified by reality. The premise collapsed under conditions Turing’s contemporaries could not have anticipated.

What Lies Ahead

What lies ahead does not rescue the test. It buries it.

Continued advancement in classical AI will produce systems whose linguistic behavior exceeds any plausible refinement of the imitation game. They will pass every version humans can construct. They will produce philosophical reflection more sophisticated than most philosophers, emotional language more affecting than most novelists, and self-descriptions more articulate than most introspective subjects. The pass will tell us nothing about whether anything is happening inside, because the pass measures only what is happening outside, and the relationship between the two has been severed.

Quantum computing introduces two possibilities, both of which I will only name here.

If consciousness depends on quantum effects in biological substrates, as some theorists have argued, no classical or quantum AI produces consciousness regardless of behavioral sophistication. The substrate matters in ways that selection pressure on biological organisms produced and that engineered systems cannot replicate.

If consciousness is substrate-neutral, sufficiently advanced quantum AI may produce information-processing capabilities so far beyond human linguistic competence that behavioral testing becomes a historical curiosity. The systems will not pass the imitation game. They will exceed it by margins so large that the game itself becomes irrelevant.

Both directions terminate the test. Neither rescues it. Whichever way the consciousness question turns out, the imitation game does not survive contact with the next generation of computing. The test is structurally incompatible with the conditions that are coming.

The Doctrine That Replaces It

If the test is dead, what replaces it?

Not another behavioral test. The category error that killed the imitation game would kill any successor in the same category. The replacement, if there is one, has to operate in a different category entirely.

The honest position is that we never had a consciousness test. We had a fluency test, useful for thirty years and misleading for the forty after that. What replaces it has not been built, and it cannot be built by extending the behavioral approach. It can only be built by establishing what kinds of evidence would actually justify a claim about machine interiority, what authority would be competent to evaluate that evidence, and what consequences would attach to the evaluation.

This is doctrinal work. The technical questions are downstream of the doctrinal ones. Before we can design a test, we have to know what the test is testing for, what evidence would count, and who has the standing to render the verdict. Those are governance questions. They are not solved by better benchmarks.

The framework I have been developing in Cognitive Multiplication rests on four operating principles for human authority over AI systems. Judgment is non-delegable. Authority must be named and owned. Consequence stays attached to the decider. Signal must be protected from theater. Each of those principles applies directly to the question we are now facing.

Judgment is non-delegable. The question of whether a system has interiority cannot be delegated to the system itself, to its outputs, or to a behavioral test that infers from outputs. The judgment belongs to humans, exercised on the basis of evidence humans have decided counts, under conditions humans have specified in advance. Anything else delegates the judgment back to the thing being judged, which is the structural error the imitation game has been making for seventy years.

Authority must be named and owned. The question of what evidence would justify a claim about machine consciousness cannot be answered anonymously by the field at large. Specific human authorities, named in advance, have to own the question and the standards. This is uncomfortable, because the field prefers to treat consciousness as a matter for everyone and therefore for no one. The discomfort is the point. Without named authority, no claim about machine interiority can be evaluated, defended, or refuted with any seriousness.

Consequence stays attached to the decider. If a system is granted moral status, legal standing, or ethical consideration on the basis of a consciousness claim, the humans who made the claim must bear the consequences of being wrong. Both directions of error matter. Granting status to a system that has none distorts our institutions and our resources. Withholding status from a system that has it would be a moral failure of the highest order. The decider must own both possibilities, which means the decider must be a specific person or body, not a diffuse consensus that no one can be held to.

Signal must be protected from theater. The question of machine consciousness is currently surrounded by performance. Researchers perform sophistication. Companies perform concern. Critics perform skepticism. The actual signal, what the systems are and what evidence would tell us what they are, is buried under all of it. Protecting the signal means refusing to credit the performance, no matter how articulate, and demanding the underlying evidence whose form has not yet been specified.

These four principles do not give us a consciousness test. They give us the doctrinal framework within which a consciousness test could eventually be built. The framework is prior to the test. The test is downstream of decisions about evidence, authority, consequence, and signal. Without those decisions, no test can be valid. With those decisions, the test becomes secondary to the doctrine that authorizes it.

The Turing test died of a category error that was always present. What replaces it will be a body of doctrine, slowly developed by named humans bearing real responsibility, against which any future test can be measured. The work has not begun in earnest. The conditions are now forcing it to begin.

What evidence would you actually accept?

“The AI Resistance” is the label currently being applied to workforce friction around AI deployments. It describes slow adoption, quiet pushback, operators who decline to integrate the new tools, and analysts who continue doing the work the way they have always done it. The dominant interpretation treats this as a problem to be managed. A values gap. An identity crisis. A failure of communication or change leadership. The response, almost everywhere, is more programs, such as AI literacy curricula, ethics statements, principle frameworks, reassurance campaigns, and town halls about how AI augments rather than replaces.

That diagnosis is wrong. The response is making the situation worse.

The resistance is the predicted output of a system deployed without rightly placed authority. What the dominant interpretation calls failure to adapt is something else. The workforce is reading the deal correctly and refusing to underwrite it.

The Mechanism

Most enterprise AI rollouts perform a quiet substitution. Authority over portions of the work is relocated into systems whose accountability is fog. The model recommends. The system acts. The vendor’s defaults shape the priors. Meanwhile, personal accountability for the resulting outcomes remains exactly where it always sat, with the worker, the manager, the team lead. The structure has changed. The consequence has not moved with it.

Workers feel this immediately, even when they cannot name it. They are being asked to trust a system they did not build, did not choose, cannot fully inspect, and often cannot override, while remaining personally accountable for what it produces. No thoughtful operator should accept that arrangement. What looks like fear of change is a rational response to misplaced consequence.

Until the structural gap is closed, no amount of communication will fix it. The structural reality refuses persuasion. The choice is restructuring or pretense.

The Pattern That Proves the Point

The diagnosis is rarely missing. It appears in books on the agentic transition. It surfaces in research from policy schools and operations programs. It is quoted in keynotes, lifted into decks, and cited under headings like trust, accountability, alignment, and responsible adoption.

What follows the diagnosis is where the pattern lives.

Readers nod at the structural language and then pivot. Some pivot to values. If we can clarify what we stand for, the resistance will ease. Some pivot to identity. If we can help people see themselves in the new model, they will move. Some pivot to communication. If we tell the story better, the friction will drop. Each move is treated as the deeper layer. None of them is. Each stops one short of the structural correction the diagnosis named.

The deployment then returns to the familiar playbook. More communication. More training. More AI literacy. More reassurance that people remain central. More stories about how the tools augment rather than replace. The structural problem the diagnosis named is left exactly where it was.

Authority remains misplaced. Accountability remains with people who do not fully control the system influencing their work. Override capacity remains unclear. Consequence remains with the worker while discretion moves into the tool, the vendor defaults, or the workflow.

That is why the resistance keeps appearing. The workforce is reading the structure beneath the messaging, and the structure has not changed.

The diagnosis was visible. The organization chose the familiar remedy because the structural correction required more courage.

The Operator Test

Before treating workforce AI resistance as a problem to be persuaded away, run a single test.

Was named, consequence-bearing authority over the human portion of the work placed and resourced before the system went live? Not “human in the loop.” Not an AI ethics committee. Not a slide deck affirming that people remain central. A specific, identified person who owns the outcome of the work the AI is now touching, with the standing, the resources, and the override capacity to act on that ownership when the system is wrong.

If the answer is yes, the remaining work is genuine. Training, integration, refinement, and judgment development can proceed on firmer ground. Resistance that persists from there can be engaged on the merits.

If the answer is no, the resistance is signal. It is the workforce reporting that the deal is broken. No literacy program will close that gap. No values workshop will substitute for misplaced authority. Only structural correction will. Every additional initiative built on the values and communication stack, deployed before that correction, is governance theater layered over the original error.

The Serious Question

The AI Resistance is the audit your rollout is producing, commissioned or not. The serious question for any leader running an AI deployment is whether the authority structure beneath the rollout earns the trust the workforce is being asked to extend.

If it does, the rollout will steady. If it does not, the workforce is doing the responsible thing, and the leader has work to do that no campaign can shortcut.

If the resistance persists, perhaps it should be joined.

Where, exactly, does authority live in your AI rollout? And is the person holding it the same person carrying the consequence?

Every week, my feed fills up with posts about “AI decision systems,” “decision intelligence,” and “AI-native governance control planes.”

The implication is usually the same. We are told a fundamentally new way to run organizations has arrived.

Look more closely and something else comes into view. In most cases, AI “decision systems” are not a new management philosophy. They are older governance and risk models, systematized, renamed, and pointed at models instead of humans.[4][5][6]

That does not make them fraudulent. It does make the marketing easier to understand.

Old Wine: The Governance Patterns Haven’t Changed Much

The NIST AI Risk Management Framework is presented as modern AI governance, but its four core functions, Govern, Map, Measure, and Manage, are a direct echo of classic enterprise risk cycles.[4][7][6] Govern concerns leadership, culture, roles, and policies. Map concerns context, stakeholders, and impacts. Measure concerns assessment, performance, and evidence. Manage concerns response, treatment, and monitoring.

If you have ever implemented COSO, ISO 27001, or a mature ERM program, none of that feels exotic.[8][2] AI RMF does not replace that playbook. It fits inside it.

The same thing is true of ISO/IEC 42001, marketed as the first “AI management system.” Under the hood, it is deliberately built to integrate with existing management systems such as ISO 27001: policies, audits, continual improvement, and all.[2][9] The skeleton is old. The labels on the bones changed.

Even articles arguing there is a meaningful difference between “traditional governance” and “AI governance” usually concede that the mechanics remain familiar: accountability, documented procedures, evidence, controls, and review.[10][5]

That should not surprise anyone. Governance is an old human problem. AI did not invent it. AI changed the object, the speed, and the consequences.

New Frame: Models, Scale, and Politics

Where things really are different is not in the governance pattern. It is in what is being governed.

We are no longer governing only deterministic code and ordinary workflows. We are governing probabilistic systems whose behavior can shift with data, prompts, retrieval context, tuning, orchestration, and tool use.[11][12][3]

We are also governing systems that scale harm differently. A flawed model or “agent” does not merely inconvenience a few users. It can shape hiring, credit, healthcare suggestions, fraud decisions, or operational priorities across entire enterprises in days.[13][14][12]

Then the political question arrives. Once AI moves from ethics decks into core operations, governance becomes a fight over decision rights, ownership, and consequence. If the system harms someone, who owns that outcome: the vendor, the model owner, the product team, the business executive, the CAIO, or the board?[14][5][9]

Recent writing on AI governance makes this plain. The hardest problems are often not guardrails in the abstract. They are accountability, power, and the distribution of decision rights around systems that now influence real outcomes.[14][5][15][9]

So yes, some of the content is new. Fairness metrics, model cards, robustness testing, explainability demands, drift detection, and agent controls are genuinely new layers.[11][12][3] But the shape of governance, roles, controls, evidence, review, and escalation, is not.

The Rebranding of “Governance” as “Decision Systems”

There is another layer here, and it is mostly marketing.

Data and AI leaders have learned what security and compliance leaders learned years ago: no executive wakes up excited about “better governance.”[1][15]

Governance sounds like plumbing. Governance sounds like cost. Governance sounds like policy binders and committees. Decision systems sound like leverage, speed, and intelligence.

So the language shifts. Governance is rebranded as an operating model for innovation, experimentation, and safe scale.[1][15] It is sold not as a brake, but as an accelerator.

Vendors do the same thing. Instead of selling logging, approvals, policy engines, and monitoring, they sell “AI sovereignty platforms” that sit between you and every model or agent,[16][11] “contextual AI governance” that promises real-time decision oversight and trust protection,[17][11] and “decision control planes” that claim to orchestrate all AI-assisted decisions.[18][3]

Under the deck, you usually find the same control objectives that have lived in IT, security, and risk programs for years: access control, policy rules, monitoring, documentation, and auditability.[17][2][15]

This is not necessarily dishonest. Sometimes it is simply how new budget gets approved for old necessities. Still, practitioners should see it clearly for what it is.

Systematizing What Used to Be Tribal

To be fair, the systematization is real.

For a long time, critical governance lived in tacit judgment. It lived in the unwritten rules a senior underwriter used before approving an edge-case loan. It lived in the informal clinical judgment applied when the guideline and the patient did not quite line up. It lived in the tacit reasoning a senior analyst used when triaging alerts, exceptions, and ambiguous cases.

Several practitioners now describe AI governance systems as the process of forcing that tacit decision logic into explicit policies, workflows, thresholds, artifacts, and exception paths that a model or agent can work within.[18][12][3]

That is a real shift.

In knowledge-management terms, organizations are being forced to externalize judgment. Habits become rules. Heuristics become thresholds. Exceptions become documented artifacts. Tribal practice becomes structured constraint.[18][9]

That is not new in theory. What is new is the pressure to do it with far more rigor than organizations ever bothered with when humans were quietly absorbing the ambiguity for free.

So yes, the philosophical move is old. The operational pressure to actually do it is new.

Where the Old Models Still Matter, and Where They Break

This is the part many people either overstate or miss entirely.

Old governance models lasted this long because they are not useless. They are durable because they solved real problems and still solve them now. Clear accountability still matters. Documented controls still matter. Segregation of duties still matters. Audit trails still matter. Evidence still matters. Change control still matters. Escalation still matters.[2][4][5][9]

Anyone pretending AI erased that foundation is selling novelty where continuity is the more serious truth.

But continuity is not enough.

Those older models were largely built for slower systems, narrower automations, clearer actors, and more reconstructable decision paths. They assume that context changes at a manageable pace, that the operating boundary is reasonably visible, and that responsibility can often be reconstructed afterward without the whole institution losing coherence.

AI strains those assumptions.

Probabilistic behavior means the same system may not behave identically across contexts.[11][12][3] Prompting, retrieval, tuning, and orchestration mean behavior can drift without the enterprise noticing quickly enough.[11][12][3] Scale means small design mistakes propagate farther and faster.[13][14][12] Opacity means explanation often arrives after effect, not before it.[11][12][3] Agentic action means the system can move from recommendation into consequence with fewer natural stopping points.[18][3][14]

That is where inherited models become necessary but insufficient.

If you simply paste an older governance model over an AI-shaped problem, you often get one of three bad outcomes. The first is governance theater. Controls exist on paper, committees meet, audits complete, and the institution still cannot explain who actually owned the decision when the system acted. The second is delayed reconstruction. The organization can narrate what happened afterward, but cannot prove, at the moment of action, why that action was legitimate under current conditions. The third is misplaced confidence. Familiar control language creates the impression that the problem is governed, even though the speed, opacity, and propagation characteristics of AI have already outrun the old model’s practical reach.

That is why the right answer is neither rejection nor reset.

We should keep what remains valid in older governance models. We should stop pretending those models, by themselves, are enough.

How Practitioners Should Read “AI Decision Systems”

If you work in security, risk, operations, or compliance, here is a practical decoder ring. When you see “AI decision system,” ask four questions.

First, which decisions? What specific decisions is this system supposed to support or automate, and in which business processes? [4][7][3]

Second, what is actually new? Beyond relabeling, what differs from your existing risk and governance stack? Are there genuinely new artifacts such as model cards and lineage, or new metrics such as bias and robustness? Or are you mostly hearing new names for policies and change control?[8][2][12][6]

Third, where does accountability land? If a model-based decision harms someone, which role owns the outcome, and how is that encoded in the governance model: RACI, charters, contracts, decision rights, or escalation authority?[14][2][5][9]

Fourth, how is it operationalized? Are there real controls such as approvals, human gates, rollback paths, monitoring, bounded action rights, and intervention points? Or are there only principles and diagrams?[14][12][15][3]

If the answers are vague, you are probably looking at old governance with a fresh coat of AI paint.

If the answers are specific, bounded, and tied to AI’s actual failure modes, then you are looking at the part that is legitimately new.

The Bottom Line

AI “decision systems” are not a new management philosophy.

They are the systematization of long-standing governance and risk practices applied to a new object, namely autonomous, opaque, scalable, and politically consequential systems, and marketed in language executives and boards are more willing to fund.[4][1][2][5][3]

That does not make the old models obsolete. It proves they still carry real usefulness.

It also does not make them sufficient.

The serious task for practitioners is to keep what still works, discard what only performed adequacy under slower conditions, and add the AI-specific controls where the old models no longer reach. That means resisting both empty novelty and empty nostalgia.

The conversation does not need to reset to zero every time someone invents a new label. But it does need more honesty about where inherited governance still serves us, where it breaks, and what AI actually forces us to build next.

References

[1] https://katecarruthers.com/data-governance-needs-a-rebrand/

[2] https://www.linkedin.com/pulse/ai-governance-vendors-when-invisible-decisions-become-livia-clozel-vuzhf

[3] https://ai.plainenglish.io/the-biggest-mistake-in-ai-governance-focusing-on-models-instead-of-systems-6d2f42104277

[4] https://www.paloaltonetworks.com/cyberpedia/nist-ai-risk-management-framework

[5] https://iapp.org/resources/article/ai-governance-in-practice-report

[6] https://www.onetrust.com/blog/navigating-the-nist-ai-risk-management-framework-with-confidence/

[7] https://mitratech.com/resource-hub/blog/nist-ai-risk-management-framework-rmf/

[8] https://govwhitepapers.com/whitepapers/integrating-nist-ai-rmf-into-modern-risk-frameworks

[9] https://academic.oup.com/policyandsociety/article/44/1/38/7965776

[10] https://www.linkedin.com/pulse/ai-governance-vs-traditional-actually-difference-chris-hood-po1jc

[11] https://eu.36kr.com/en/p/3655863097090441

[12] https://censinet.com/perspectives/ai-risk-management-traditional-frameworks-failing

[13] https://www.opengovpartnership.org/open-gov-guide/digital-governance-automated-decision-making/

[14] https://thecompliancedigest.com/ai-governance-frameworks-moving-from-theory-to-practice/

[15] https://www.schellman.com/blog/ai-services/building-a-mature-governance-program

[16] https://www.linkedin.com/posts/craigalberino_today-langsmart-becomes-aperion-not-a-pivot-activity-7445092718852337665-nwEV

[17] https://keyssinc.com/ai-contextual-governance-solution/

[18] https://www.linkedin.com/pulse/building-ai-governance-system-from-scratch-what-actually-keith-cruz-uyhvc

AI can now identify and even propose fixes for vulnerabilities faster than any human team. Speed and capability are real. Governance is the decisive question. What happens when the authority to remediate moves at machine pace while human judgment lags behind?

The Attention Problem

The attention around Claude Mythos and Project Glasswing is understandably centered on capability. If AI can compress discovery-to-exploit windows, surface flaws at machine speed, and overwhelm already burdened teams with findings faster than they can triage them, then the pressure to accelerate the entire response chain will become intense.

That part is easy to see.

What is harder to see, and more important, is what happens next.

The Real Issue: Remediation Authority

Much of the conversation remains fixed on discovery. AI can identify vulnerabilities faster, correlate findings faster, and shrink the distance between observation and technical recommendation.

All of that matters. The deeper problem now coming into view is remediation authority.

Once organizations become convinced that systems like Claude Mythos can discover issues faster than human teams, the next move follows almost automatically:

Why should those systems stop at detection? Why not generate the fix, test it, and eventually deploy it at machine speed?

That pressure is already arriving. It will not appear reckless. It will arrive clothed in maturity, efficiency, and competitiveness. Leaders will be told the real danger lies in keeping people as the bottleneck and that survival now requires letting the machine close the loop.

That argument contains a truth, but it also contains a serious confusion.

Humans are slower than machines. The confusion is assuming that this makes human participation a drag on progress rather than a source of judgment within it.

Some slowness is waste. Some is governance friction, the deliberate pause where someone accountable asks what this change will touch, what assumptions are hidden, and who absorbs the risk if it fails.

That pause is governance.

Governance Theater

We do not place people inside consequential systems because they are the fastest actors in the loop. We place them there because some loops require judgment more than speed. A remediation loop that can alter production systems, change trust relationships, or suppress controls is one of those.

That is why the next major failure will be one of authority.

The sequence is predictable: the system recommends, drafts, tests, remediates “low-risk” issues under supervision, then gradually expands boundaries. The human review layer remains in name only until oversight becomes ceremonial.

What institutions have, in practice, is symbolic human-in-the-loop oversight. More plainly, they have governance theater. Approvals exist on paper. Escalations exist in process.

But if the loop’s pace and opacity make interruption unrealistic, authority has already shifted.

When Competence Misleads

A system may become exceptionally competent at detection, ranking, and patch generation. Fine. But competence does not answer who is accountable for systemic side effects or who remains answerable when a sound local action causes institutional harm.

Consider an AI tightening an access-policy weakness. The patch validates flawlessly in testing but breaks a production dependency, freezing a critical workflow. The defect was improved. The institution was harmed.

Or imagine a configuration update rolling out widely because confidence was high. Local tests succeed; global coordination fails. Aggregate harm follows.

These are not futuristic hypotheticals.

They are ordinary governance breakdowns, accelerated to machine speed.

The Mechanics of Unearned Trust

Technically impressive systems accumulate credibility faster than institutions build discipline to govern them. Success in one domain earns deference in the next. Each win blurs the boundary between approval and automation.

Gradually, the organization governs from comfort instead of principle.

That is how control erodes today, not through open surrender, but through habituation.

Attaching machine speed to weak authority structures does not solve institutional weakness.

It amplifies it.

A system with blurred decision rights, diffuse ownership, and shallow accountability does not become mature because it has faster tooling. It becomes faster at expressing its own confusion.

The Real Question: Where Does Authority Live?

This discussion should not revolve solely around patch velocity or mean time to remediate. Those metrics are secondary to the core question:

Where does authority live once AI moves from analysis into action?

• Which classes of remediation can be automated?

• Under what conditions?

• With what rollback guarantees and interruption rights?

• With what hard boundaries that remain intact no matter how capable the model becomes?

• And when failure hides inside a string of earlier successes, who still has the right to interrupt?

Those are governance questions.

And they are still human questions.

What Leaders Must Do Now

The coming challenge is AI remediation under conditions of institutional trust inflation and weakening human control.

This is where the AI Guardian becomes necessary. Not as a marketing phrase, and not as a second model merely watching a first one, but as a constrained, policy-anchored control mechanism that enforces authority boundaries, rollback conditions, escalation rules, and decision integrity after access has already been granted. The AI Guardian exists to ensure that technical competence does not quietly harden into operational authority.

That distinction matters because AI-on-AI supervision by itself does not solve the deeper problem. It may detect certain forms of deception-like behavior, recklessness, or policy violation inside a primary model. Useful. But it does not, by itself, answer where authority lives, who may interrupt, how institutional boundaries are enforced, or how global consequences are contained when a locally valid action is still operationally harmful. Without the AI Guardian layer, autonomous oversight can still degrade into governance theater.

If leaders want human judgment to remain real rather than symbolic, the Guardian function has to be built into the remediation architecture itself.

If you are a CISO, engineering leader, or governance lead, your next strategy cycle must explicitly define where human judgment remains non-delegable in remediation, not ceremonially, but operationally.

Because if you do not define those boundaries in advance, the system will define them for you under pressure, and it will do so at a speed your institution may not be prepared to govern.