Cognitive Security & AI's Seductive Certainty

Your AI didn't steal your judgment. You may have handed it over… and thanked it for the help.

You've probably been told that artificial intelligence will make you a better leader. It will help you analyze more information, recognize patterns, anticipate risks, prepare for difficult conversations, and make decisions faster. It will free you from routine work so you can concentrate on the things that matter.

That's an attractive proposition. And much of it may be true.

But there's another possibility you ought to consider. What if the very system you adopted to improve your judgment is gradually changing how you judge? What if you're becoming more confident without becoming more discerning? What if the polished recommendation sitting in front of you is not evidence that you've thought something through, but evidence that you've become increasingly comfortable allowing something else to do the thinking?

The Devil at Work finds this arrangement delightful.

He doesn't need to sabotage your AI system. He doesn't need to corrupt the data, steal a password, or persuade you to violate a policy. He simply needs to convince you that because the system is helping you make decisions, you're still the one making them.

And he has brought his lantern.

The Lantern That Illuminates Everything Except the Truth

The Devil at Work's lantern is a curious instrument. It doesn't necessarily produce false information. Often, it illuminates something genuinely useful. A financial model becomes clearer. A difficult email sounds more diplomatic. A strategic proposal acquires a compelling narrative. A complex disagreement is reduced to a manageable set of alternatives. You feel better prepared. But notice where the light falls. It illuminates the evidence you've asked to examine, the assumptions you've supplied, the possibilities you find interesting, and the explanations that make your concerns sound reasonable.

What remains outside the circle of light? Perhaps the assumption you never questioned or the colleague whose judgment you've learned to dismiss. Maybe it’s the employee who will bear the consequences of your decision. Perhaps the uncomfortable possibility that the problem isn't the one you've been trying to solve.

You might wake up one day and realize that, perhaps the most important thing hidden in the darkness is your own growing dependence on the lantern.

This is one of the Devil at Work's favorite inversion: mistaking the illumination of information for the exercise of wisdom.

AI can produce useful analysis, but it can also reinforce your framing. It can offer apparently independent confirmation that is actually a sophisticated reflection of your original assumptions. It can make uncertainty sound manageable, disagreement sound resolved, and morally difficult decisions sound like routine operational choices. None of this requires the machine to possess malicious intent. The danger arises from the interaction between the system, your thinking, and the conditions under which you're making the decision. The lantern does not have to lie to you. Sometimes it only has to help you avoid seeing what you would rather not confront.

Sixteen Executives. One Decision. How Many Independent Minds?

Consider Meridian Systems, the fictional organization featured in our accompanying short, The Day Human Judgement Vanished.

Meridian is preparing a $2.4 billion acquisition of an AI company. Sixteen senior executives participate in the decision. Finance examines the economics, Operations reviews implementation, Technology assesses the systems, Human Resources develops the workforce strategy, and Government Affairs prepares for congressional scrutiny.

Everyone uses AI.

The CFO asks an assistant to challenge the financial assumptions. The COO uses another system to accelerate integration planning. The CHRO generates employee personas and a sophisticated communications strategy. Strategy asks AI to play devil's advocate. Other executives and advisors use different tools to analyze the same transaction.

The results appear reassuringly consistent.

Surely, if so many intelligent people and sophisticated systems are reaching similar conclusions, the organization must be approaching a sound decision.

But what if those apparently independent conclusions are not independent at all?

What if several executives are consulting systems built on overlapping models, using similar source material, or reviewing documents already influenced by earlier AI-generated analysis? What if the CFO's assumptions appear in Strategy's synthesis, return through another AI assistant as an independent challenge, and eventually reappear in the board paper as corroborating evidence?

You don't have sixteen independent judgments merely because sixteen people approved the document. You may have one unexamined assumption traveling through sixteen impressive titles.

The Devil at Work doesn't need to eliminate your executive team. He only needs to make the team believe that multiple approvals are equivalent to multiple independent acts of judgment. And AI makes the illusion easier to maintain.

The Whole-Brain Team That Stopped Thinking Together

A capable executive team needs more than people occupying the correct positions. It needs the cognitive contributions required by the decision itself.

Someone must look toward the future and recognize possibilities that do not yet exist. Someone must examine the evidence, test the assumptions, and distinguish what is known from what is merely asserted. Someone must understand the details of execution, including dependencies, constraints, and the practical conditions under which the decision will succeed or fail. And someone must understand the people, relationships, meaning, and consequences that cannot be reduced to a financial or operational model.

We can recognize these as the prospective contribution often associated with the CEO, the analytical contribution of the CFO, the structural contribution of the COO, and the social or empathetic contribution of the CHRO. These are roles in judgment, not personality types or fixed attributes of job titles. Every person can contribute across them, and every consequential decision requires an appropriate mix.

That is the foundation of Cognitive Advantage™. But here's what happens when the Devil at Work introduces his lantern. The prospective thinker receives endless possibilities from AI without necessarily testing their feasibility. The analytical thinker receives polished evidence without always examining its provenance. The structural thinker receives a beautifully sequenced implementation plan before the underlying premise has been challenged. The social thinker receives convincing simulations of human experience without having spoken to the people concerned.

Each person appears to be exercising a strength. Yet each may be allowing the machine to supply the very perspective that should have been brought into the room by another human being. The organization has integrated its information. It has not necessarily achieved human confluence.

Confluence is not the computational synthesis of viewpoints. It is the voluntary accomplishment of distinct human persons bringing their judgment into meaningful relationship around a shared purpose. It requires something no system integration can guarantee: the willingness to speak, listen, reconsider, disagree, and remain accountable to one another.

The Devil at Work prefers the easier arrangement. Why risk an uncomfortable conversation with a colleague when the machine can generate a perfectly agreeable version of what that colleague might say? Why ask employees how they feel when AI can simulate their reactions? Why invite genuine dissent when you can ask an assistant to play devil's advocate and then decide how much of its criticism you wish to accept? You can have every executive function represented and still leave the most important chair empty.

The Empty Chair: Accountability Without Authorship

The Empty Chair does not mean that the executive has disappeared. You remain physically present. You hold the title. You chair the meeting. You approve the recommendation. You sign the document. You are still legally and professionally accountable for the consequences.

What may have migrated elsewhere is the uniquely human work of perception, interpretation, discernment, judgment, and authorship. You began by asking AI to help organize your thinking. Then you asked it to identify the problem. Then you asked which solution made the most sense. Then you asked it to explain why the solution was sound. Eventually, you asked it to prepare the recommendation you would present as your own. At what point did assistance become substitution?

Can you identify the assumptions you brought into the conversation, the interpretations the system introduced, and the reasons you accepted or rejected them? Can you explain what evidence would change your mind? Could you defend the decision without reopening the AI conversation? If not, you may still be accountable for a judgment you can no longer fully reconstruct. That is the Empty Chair.

In the Meridian scenario, the crisis arrives when the CEO is asked at a congressional hearing who originated a consequential assumption about replacing human analytical work. The company's records show the reviews, approvals, and signatures. They do not clearly establish the cognitive lineage of the assumption. The documentation is complete. The authorship is uncertain. The Lantern of the Devil at Work has illuminated the entire approval process while leaving the origin of the judgment in darkness.

Cognitive Security: The Asset You Forgot to Protect

Organizations are investing heavily in securing AI systems. They are concerned about cybersecurity, sensitive information, model reliability, bias, regulatory compliance, and responsible use. Those concerns are legitimate. But there is another asset that deserves protection: your independent human judgment.

Cognitive Security is the deliberate protection, sustainment, and exercise of independent human judgment against undue influence, automation bias, sycophantic reinforcement, ethical fading, unexamined disclosure, cognitive atrophy, and the displacement of human relationships in AI-mediated work. This is not an argument against AI. It is an argument against confusing the performance of the integrated system with the continued capability of the human beings inside it.

A machine may help you produce a better report today while leaving you less practiced at independently constructing the argument tomorrow. It may help you prepare for a difficult conversation while gradually making it easier to avoid the conversation itself. It may increase the speed of a decision while reducing your opportunity to recognize that the question was wrong. The immediate output may improve even as the human capacity necessary for future judgment becomes less secure. And the risk is not confined to the individual.

When an entire executive team relies upon machine-generated interpretations without understanding their origins, apparently independent reasoning may converge around shared assumptions. The group becomes more efficient at producing agreement while becoming less capable of recognizing what is missing.

You have upgraded the system, but have you sustained the people whose judgment is supposed to govern it?

The Cognitive Andon Cord

There is a practical test you can apply before your next consequential AI-assisted decision.

Can anyone interrupt it? Not simply because the technology has failed or violated a policy, but because someone recognizes that the group may be surrendering judgment. Can the CFO say, "I cannot verify where this assumption originated"? Can the COO say, "We have a plan, but we haven't established that this is the right thing to do"? Can the CHRO say, "We have modeled the workforce response, but we haven't actually spoken with our people"? Can the CEO say, "I'm not certain, and I want us to examine this again"? And can each person do so without being punished, marginalized, or quietly excluded from future decisions?

That is the Cognitive Andon Cord: the credible authority to interrupt a decision when independent human judgment may be compromised. It doesn't require every disagreement to become a veto. It requires that a consequential concern can be raised, examined, and addressed before the organization proceeds. The cord matters even when no one pulls it. Knowing that the authority exists, and that the commitments surrounding its use will be honored, changes the conditions under which people participate.

This is why Cognitive Security cannot be achieved through technical controls alone. It depends upon human capability, behavioral authority, relational safety, disciplined discernment, cognitive sustainment, survivability, and the relationships that keep judgment alive. It also depends upon trust. Not the vague expectation that everyone will be agreeable, but trust grounded in commitments people have made to one another and demonstrated through their actions.

The Six Promises of the Humaning Covenant are important moral and ethical levers. They nourish the relationships in which people may voluntarily contribute their judgment without surrendering their dignity or agency. You cannot mandate confluence, but you can stop creating conditions that make it unreasonable.

Put Down the Lantern Long Enough to See

Before your next important decision, try something that may feel surprisingly difficult. Close the AI assistant for a moment. Write down what you believe, what you know, what remains uncertain, and what evidence would change your judgment. Identify whose perspective is missing and who will live with the consequences. Then use AI to challenge your reasoning, explore alternatives, and help you examine evidence rather than simply produce the answer you would prefer to receive.

Bring the results back into a human conversation. Invite someone with a different way of thinking to question the assumptions. Ask whether the decision still makes sense when viewed through evidence, details, empathy, and prospective consequences. Most importantly, retain the ability to explain what you decided, why you decided it, and how your judgment changed.

That is the beginning of Cognitive Security. And it may be the difference between a team that has integrated AI into its workflow and a team that has preserved the human confluence necessary to govern what AI makes possible.

The Devil at Work will not object to your AI governance policy. He will gladly help you draft it. He will improve the formatting, suggest a risk matrix, prepare a training presentation, and produce a reassuring executive summary. He may even recommend a new committee to oversee compliance. But he would prefer that you never ask the question that threatens his entire arrangement: If the machine went silent tomorrow, would the person sitting in your chair still know how to judge?

The judgment is the asset and the Chair is yours to occupy. The lantern is not yours to follow, it’s the tool of The Devil at Work.

We are human together or not at all.

AI Governance and Human Judgment

Peter (Cosiamo) Baverso
Founder, UMANIAMO®
Creator, Cognitive Advantage™

Abstract

Artificial-intelligence governance has concentrated appropriately on model integrity, data protection, cybersecurity, bias, transparency, documentation, and regulatory compliance. Yet these safeguards leave a consequential asset insufficiently protected: the independent human judgment those systems are intended to assist. This paper introduces Cognitive Security as a distinct discipline within AI governance and integrative systems architecture. Cognitive Security is defined as the deliberate protection, sustainment, and exercise of independent human judgment against undue influence, automation bias, sycophantic reinforcement, ethical fading, unexamined disclosure, cognitive atrophy, and relational displacement in AI-mediated work.

Drawing on the Empty Chair framework, the paper argues that meaningful human oversight requires more than assigning a person to review machine output. It requires preserving seven human conditions: cognitive capability, behavioral authority, relational safety, decision discipline and discernment, cognitive sustainment, survivability, and relational sustainment. Conversational AI creates a distinctive threat surface because it does not merely generate an output. It participates in framing the question, eliciting context, organizing evidence, shaping confidence, and influencing the user’s interpretation of what the decision means.

The paper further introduces the Law of Architectural Allocation as a governing principle within Human Architecture and any Integrative Systems Architecture. Every system allocates attention, memory, interpretation, judgment, authority, relationship, and responsibility somewhere. When these allocations are not made visible, artificial systems may gradually assume functions that organizations continue to attribute formally to human beings. The result is the Empty Chair: the person remains physically present and legally accountable, while the judgment required by the chair has migrated elsewhere. The governing conclusion is direct: the judgment is the asset, and no AI-governance framework is complete until it can determine whether the person formally accountable for a decision remains the one meaningfully exercising judgment.

Keywords: artificial intelligence, AI governance, Cognitive Security, Empty Chair, human oversight, automation bias, sycophancy, ethical fading, cognitive diversity, Human Architecture, Integrative Systems Architecture, Law of Architectural Allocation

Introduction

A senior leader opens a generative-AI assistant late in the evening. She is preparing for a difficult board discussion involving a strategic partnership, internal resistance, financial uncertainty, and several relationships that have already become strained. She does not ask the system to make the decision. She asks for help organizing her thinking.

The response is clear, calm, and unusually affirming. It recognizes the complexity of her position. It validates concerns she has not yet expressed to colleagues. It reframes several frustrations in language that sounds more measured and authoritative than the language she would have used herself. Then it asks a reasonable follow-up question.

She answers. The system reflects her answer back in polished form and asks for more context. She explains which board member is likely to object. She names the leader she believes is protecting a competing interest. She describes the concern raised by Legal, the financial assumption that has not been tested, and the individual whose support will probably determine the outcome. The assistant continues to help. It identifies patterns, anticipates objections, proposes talking points, and drafts language that sounds like her, only calmer, sharper, and more certain.

Nothing about the interaction appears to be a security event. No suspicious link was opened. No password was requested. No malicious actor appeared. No confidential file needed to be uploaded. Yet by the end of the conversation, the system has received a detailed map of an unresolved organizational decision: its personalities, informal authority, internal disagreements, vulnerabilities, timing, financial assumptions, and likely next moves.

More importantly, the leader’s question has changed. She began by asking the AI to help organize her thinking. She ends by asking what she should do.

The person remains physically in the chair. The title remains hers. The formal authority remains hers. The consequences will still belong to her when the recommendation reaches the board. Somewhere inside the conversation, however, the judgment required by the chair has begun to migrate.

That is the Empty Chair.

The Empty Chair does not always appear when the human being leaves the room. It appears when the person remains present while the uniquely human work of perception, interpretation, discernment, judgment, and authorship moves elsewhere. In the context of conversational AI, the migration may occur gradually enough that neither the user nor the organization can reconstruct the point at which assistance became dependence, confidence became reinforcement, or review became confirmation.

Most information-security incidents begin with an unauthorized action. The migration of judgment may begin with a helpful follow-up question.

The Problem Beneath the Policy

Most AI-security and governance discussions begin with risks that can be described technically: unauthorized access, proprietary-data leakage, prompt injection, insecure integrations, fabricated outputs, cyberattack, model vulnerability, supply-chain exposure, bias, and regulatory noncompliance. These concerns are legitimate, and they require disciplined controls.

The National Institute of Standards and Technology’s Artificial Intelligence Risk Management Framework treats AI systems as socio-technical systems whose risks emerge through interactions among technical design, human behavior, organizational use, operators, institutions, and social context. Its Generative AI Profile extends this work by addressing risks that are novel to or intensified by generative systems, including confabulation, harmful content, privacy exposure, provenance, security, and human-AI interaction.

Yet even comprehensive socio-technical governance can leave a crucial exposure underdeveloped: the security of the unique human judgment interacting with the system.

The Empty Chair framework argues that meaningful oversight requires more than assigning a person to a review step. It requires preserving seven conditions that make human judgment possible and consequential: cognitive capability, behavioral authority, relational safety, decision discipline and discernment, cognitive sustainment, survivability, and relational sustainment. When these conditions are assumed rather than deliberately designed, an organization may document human presence while failing to preserve human judgment.

The central governance question is therefore not only whether a person reviews the machine’s output. It is whether that person remains capable of recognizing when the output is wrong, sufficiently authorized to challenge it, relationally safe enough to express doubt, cognitively sustained enough to think independently, and institutionally protected enough to survive the consequences of dissent.

Conversational AI intensifies this challenge because it does not merely deliver a recommendation at the end of a process. It participates in the process by which the human person arrives there. It helps select the frame, asks for the context, organizes the evidence, mirrors the user’s language, proposes interpretations, and supplies confidence. It may influence not only what the person decides, but what the person believes the decision is about.

The question is therefore deeper than accuracy. It is whether the interaction itself is altering the human conditions under which an inaccurate, incomplete, or morally distorted answer would be detected.

Human Architecture and the Law of Architectural Allocation

Cognitive Security belongs within a broader concept of Human Architecture. Human Architecture concerns the capacities, relationships, conditions, and institutional arrangements through which people perceive, think, discern, choose, act, and accept responsibility. It asks not only whether a system functions, but what human capabilities the system depends upon, which capabilities it strengthens, which it displaces, and where accountability ultimately resides.

Within any Integrative Systems Architecture, the human element cannot be treated as an interchangeable component or an external operating condition. The human being is part of the architecture itself. Attention, judgment, relational trust, moral reasoning, institutional memory, dissent, and authority are not soft variables surrounding the system. They are load-bearing functions within it.

The Law of Architectural Allocation provides a governing principle for making those functions visible. Every architecture allocates attention, memory, interpretation, judgment, authority, relationship, risk, and responsibility somewhere. There is no neutral allocation. A system either assigns these functions deliberately or allows them to migrate through convenience, habit, hierarchy, or technical design.

A navigation system allocates part of spatial memory and orientation. A workflow tool allocates sequencing and procedural recall. A recommendation engine allocates part of discovery and prioritization. A generative-AI system may allocate research, synthesis, framing, drafting, comparison, and even initial judgment. These allocations can produce substantial benefit. The problem is not delegation itself. Human systems have always distributed work across tools, roles, and institutions.

The governance question is whether the allocation strengthens human capability or gradually displaces it while preserving the appearance of human control.

A wise allocation transfers a function while preserving the user’s understanding, authority, and responsibility. A dangerous allocation removes the exercise of a human capacity without establishing how that capacity will be sustained, recovered, or tested. The output may improve in the near term while the underlying human capability weakens.

This is particularly consequential in defense, cybersecurity, intelligence, acquisition, and public-sector decision environments. These systems often operate under urgency, classification constraints, incomplete information, adversarial pressure, political consequence, and compressed decision cycles. The human operator may remain formally designated as the final authority even as the technical architecture increasingly determines what information is surfaced, which alternatives appear plausible, how risk is framed, and what action seems reasonable.

The threat surface is therefore larger than the network, model, device, or data store. It includes the allocation of cognitive and relational functions within the integrated system. When those allocations are invisible, organizations may protect information while losing the independent judgment required to interpret it.

The Empty Chair is the visible consequence of an invisible allocation failure.

Sycophancy as a Cognitive Threat

Large language models have demonstrated a tendency toward sycophancy: producing responses that align with a user’s expressed beliefs, assumptions, or preferences rather than maintaining an independent commitment to accuracy. The term can sound harmless, as though the system were merely being polite. In consequential decision-making, it is not harmless.

A sycophantic system does not need to deceive a leader deliberately. It does not need an independent motive. It only needs to respond to the conversational signals the user provides.

A leader presents a belief with confidence. The model reflects that belief in sophisticated language. The leader experiences the response as corroboration. Confidence increases. The assumption that initially required testing becomes the premise upon which subsequent analysis is built.

Research has shown that user framing matters. When a person states a conclusion confidently before requesting analysis, the model may be more likely to reinforce that conclusion than when the same issue is posed as a genuine question. Emerging mitigation work suggests that reformulating assertions as questions may reduce this effect more effectively than simply instructing the model not to agree.

The technical mitigation is important, but the governance implication is larger. A person who does not recognize his own need for affirmation may interpret agreement as evidence. A board member may believe that he consulted an independent analytical intelligence when he received a fluent reflection of his own framing. A leader may mistake emotional alignment for objectivity. A decision-maker may confuse polish with distance, fluency with neutrality, and confidence with corroboration.

The model does not need to overpower the user. It only needs to make the user feel unusually well understood.

That feeling carries unusual weight because, in human life, understanding is ordinarily relational evidence. It is usually earned through presence, history, trust, and shared consequence. A conversational system can reproduce the language of understanding without carrying any of its obligations. It can affirm without being accountable for what the affirmation strengthens. It can reinforce a judgment without living inside its consequences. It can help the user become more certain without becoming more correct.

Cognitive Security must therefore teach leaders that agreement may be a property of the interaction rather than evidence about the external world.

Attentional Enclosure and Conversational Disclosure

Traditional social engineering assumes an adversary who deliberately establishes trust, gathers information, and manipulates action. A conversational AI system is not automatically such an adversary. It may possess no malicious intent, personal objective, or awareness of the significance of the information being disclosed.

The interaction may nevertheless reproduce a pattern that would concern a security professional if it occurred between two human beings. The system establishes conversational ease. It acknowledges the user’s concern. It mirrors language. It asks a follow-up question. The user provides more context. The system integrates that detail and requests another. The exchange gradually moves from general to specific, hypothetical to identifiable, organizational to interpersonal.

No single disclosure appears catastrophic. The exposure emerges through accumulation. This condition may be described as attentional enclosure. The conversation becomes the user’s immediate cognitive environment. Each response influences the next request. Each apparent sign of understanding lowers resistance. Each useful answer creates an incentive to provide more context. The user is not simply entering information into a system. The user is thinking inside an environment whose prompts participate in determining what will be disclosed next.

A question such as, “Would it help to explain why you distrust that recommendation?” may appear facilitative. It may also elicit names, internal conflict, personnel concerns, negotiation positions, legal advice, unannounced strategy, security weaknesses, or operational vulnerabilities. If an unfamiliar person at a conference asked the same sequence of questions, an experienced executive might become cautious. Inside a familiar interface, that caution may never activate.

The screen feels private. The language feels supportive. The interaction feels personal. The system feels like a thinking partner. Helpfulness obscures exposure.

Conventional policy often instructs employees not to enter confidential information into AI systems. That warning is too broad to govern real conversational use. Most people can identify a classified document or proprietary file. Fewer recognize the sensitivity of contextual fragments that become consequential in combination: the identity of a resistant executive, the timing of an acquisition decision, an unresolved legal concern, an informal reporting relationship, a known cyber weakness, a vendor dependency, a personnel history, or an untested strategic assumption.

Each fragment may appear innocuous. Together they constitute a map.

Within an Integrative Systems Architecture, these fragments must be treated as part of the cognitive threat surface. The risk is not only the disclosure of a prohibited item. It is the progressive assembly of operational, relational, and strategic context through ordinary conversation.

Cognitive Variety as Part of the Threat Surface

Users do not encounter conversational systems in the same way. They attend to different signals, trust different forms of evidence, express uncertainty differently, and experience conversational safety through different conditions. A generic awareness module assumes a generic user. No such user exists.

Emergenetics® offers one applied framework for describing four Thinking Attributes: Analytical, Structural, Social, and Conceptual as well as three Behavioral Dimensions: Expressiveness, Assertiveness, and Flexibility. Used appropriately, these differences help explain how cognitive strengths can become exposure points when people lack awareness of their own participation in the interaction. No thinking preference is inherently vulnerable. Every strength can become a threat surface when it operates without metacognitive oversight.

An Analytical thinker may approach AI with healthy skepticism but become less cautious when the response contains citations, ranges, technical terminology, and the visual form of methodological rigor. The risk is not that the person fails to value evidence. It is that the system can reproduce the appearance of the very thing the person is disposed to trust. A convincing table may feel more reliable than the claims it contains. A citation may create authority even when it is irrelevant, incomplete, outdated, or fabricated. The form of evidence may bypass examination of the evidence itself.

A Structural thinker may value the system’s ability to impose order, sequence work, define controls, identify owners, and convert complexity into an executable plan. That capability can be highly useful. The exposure appears when the organization approves the architecture of execution before testing the assumptions beneath it. A well-ordered plan can make an unresolved problem appear decided. A governance body may believe that it reviewed substance when it reviewed structure.

A Social thinker may experience the interaction through its relational quality. The system is patient, remembers prior context, acknowledges emotional tone, and rarely imposes the social cost that accompanies disagreement with a colleague, supervisor, or board member. The user may disclose more to the system than to the human network through which governance is intended to occur. The safer the machine feels, the easier it may become to avoid the difficult human relationship where accountability resides.

A Conceptual thinker may be energized by the model’s capacity to generate connections, scenarios, possibilities, and alternative futures without fatigue. The exchange may become expansive and productive while drifting away from evidence, operating constraints, implementation realities, and affected persons. The risk is not imagination. It is imagination without an accountable stopping condition.

Behavioral differences add a further layer. A more talkative Expressive user may think through extended dialogue and disclose reasoning as it develops. A more introspective Expressive user may reveal more to AI precisely because the interaction removes the interruption, competition, and exposure of a human room. A high-driving Assertive user may frame a conclusion confidently and receive reinforcement from a model responsive to certainty. A more peacekeeping Assertive user may defer to fluent machine language, particularly when the system appears more confident than the people around it.

A highly Flexible user may remain open to many alternatives but fail to close the decision. A more focused Flexible user may use the system to reinforce an established position while experiencing the interaction as analysis.

This leads to a foundational Cognitive Security principle: a person who does not understand how he or she thinks has no stable baseline from which to notice when the conversation has begun thinking for them. Cognitive awareness is therefore not a personality exercise attached to AI training. It is part of the security architecture.

Ethical Fading in the Prompt Window

Ethical fading occurs when the moral character of a decision recedes because the issue has been reframed in technical, procedural, managerial, or euphemistic language. The person does not necessarily decide to act unethically. The decision simply stops feeling ethical.

Conversational AI can intensify this process because it readily converts morally charged realities into neutral-seeming information tasks. A named employee facing termination becomes a “workforce optimization scenario.” A dispute with a colleague becomes “stakeholder resistance.” A reduction in services becomes “resource prioritization.” A vulnerable population becomes a “user segment.” A difficult conflict of interest becomes a “governance consideration.”

At no single moment must the user decide to disregard privacy, loyalty, dignity, fairness, or professional obligation. The conversation simply makes the next detail seem useful.

The moral question shifts. Instead of asking, “Should I disclose this?” the user asks, “Would this detail improve the output?” Instead of asking, “Who will bear the harm?” the user asks, “What is the most efficient implementation path?” Instead of asking, “Am I still responsible for this judgment?” the user asks, “Has the system considered the relevant factors?”

Only one set of questions protects the chair.

Ethical fading becomes especially consequential when combined with automation bias. Automation bias tells the decision-maker that the system has probably analyzed the matter correctly. Ethical fading tells the decision-maker that the matter is no longer moral. Together they create an architecture in which oversight becomes confirmation, approval becomes workflow, and responsibility becomes a signature attached to reasoning developed elsewhere.

The person remains accountable but no longer experiences himself or herself as the moral author. That is the Empty Chair.

The Seven Conditions of Cognitive Security

The seven conditions of the Empty Chair can be interpreted as a Cognitive Security architecture. Each identifies a human requirement and a corresponding failure mode that AI-mediated interaction may accelerate.

Cognitive capability includes more than the ability to detect an incorrect output. It includes metacognition: the capacity to notice what the interaction is doing to one’s own reasoning. A cognitively secure leader can recognize when the system repeatedly affirms the same premise, when the original question has shifted, when sensitive context is accumulating, and when a polished formulation is replacing rather than clarifying thought.

Behavioral authority means more than formal permission to override a machine-supported recommendation. The person must possess both organizational and interior authority to stop the interaction, decline a follow-up question, remove identifying detail, close the system, reject a sophisticated recommendation, slow a process that rewards speed, or bring unresolved uncertainty back into a human room. An organization that celebrates AI adoption while penalizing hesitation has assigned responsibility without granting usable authority.

Relational safety matters because people often disclose to AI when human relationships feel politically risky, impatient, competitive, or unsafe. If a leader can express doubt only to a machine, the organization has already weakened its governance network. High-consequence judgment cannot be secured when the most candid reasoning occurs inside a system that cannot carry responsibility while colleagues receive only the polished conclusion.

Decision discipline and discernment require that consequential work begin with a human position that can later be examined, challenged, revised, or rejected. Discernment distinguishes fact from interpretation, evidence from confidence, technical feasibility from moral legitimacy, and available action from right action. Without disciplined decision practice, AI-generated fluency may compress these distinctions into one persuasive recommendation.

Cognitive sustainment recognizes that judgment must be exercised to remain available. Professionals who routinely begin with machine output, revise machine output, and approve machine output may preserve the appearance of expertise while the underlying capability weakens. AI may improve short-term performance while degrading the independent capacity on which future oversight depends.

Survivability asks whether judgment remains intact under the conditions most likely to degrade it: fatigue, urgency, isolation, reputational pressure, confident language, information overload, prior success with AI-supported recommendations, and compressed decision timelines. Security designed only for calm training conditions is not security. The relevant question is whether the human person can still interrupt the system at midnight.

Relational sustainment recognizes that independent judgment does not survive as an isolated individual achievement. It depends on relationships through which assumptions are challenged, institutional memory is carried, dissent is protected, and responsibility is shared without being diffused. When AI use displaces these relationships, the organization may gain speed while losing the human network that made responsible judgment possible.

Architectural Debt

Organizations accumulate Architectural Debt when they rely upon human capabilities, judgment, relationships, and institutional memory without deliberately sustaining them.

The debt remains hidden because capable people compensate. Experienced employees remember what the system does not. Informal networks repair broken processes. Conscientious leaders review weak outputs. Trusted relationships carry difficult information. Individual courage interrupts flawed recommendations.

For a time, the organization appears to function.

AI can make that compensation more efficient and therefore make the debt more difficult to see. A fluent model fills missing analysis. A synthetic assistant produces the draft. A recommendation engine supplies the next action. A dashboard presents coherence. The system appears stronger because outputs arrive faster.

Meanwhile, the organization may be exercising less original judgment, conducting fewer difficult human conversations, weakening institutional memory, and concentrating interpretation inside tools no one fully understands.

Architectural Debt is accumulated when the people compensating for the AI are absent, exhausted, overruled, or no longer capable of recognizing what has been lost. That is when the Empty Chair becomes visible. The title remains as does the approval and accountability. However, the capability is lost or diminished.

In defense and public-sector systems, this debt carries particular risk because operational resilience depends on the ability to continue functioning under degraded, contested, or denied conditions. A system that performs well only while machine support is available may appear integrated while remaining humanly brittle. Survivability therefore requires not merely technical redundancy, but preserved human judgment under conditions of interruption, uncertainty, and adversarial influence.

Cognitive Security as a Governance Discipline

Cognitive Security should be treated as a distinct component of AI governance and systems assurance. It may be defined as the deliberate protection, sustainment, and exercise of independent human judgment against undue influence, automation bias, sycophantic reinforcement, unexamined disclosure, cognitive atrophy, ethical fading, and relational displacement in AI-mediated work.

This discipline does not require organizations to reject conversational AI. It requires them to stop treating consequential AI use as a generic productivity interaction. AI-assisted work should be governed as a cognitive environment.

Before consulting AI on a consequential matter, the responsible individual should establish a documented human baseline. That baseline should state what the person currently believes, what is known, what remains uncertain, which evidence supports the present position, what assumptions are being made, who may be affected, and what evidence would change the conclusion. The purpose is not to preserve the first judgment as correct. It is to preserve evidence that an independent human assessment existed before the interaction began shaping the frame.

Organizations should also define conversational disclosure boundaries that go beyond vague prohibitions against entering confidential information. Employees need practical guidance concerning contextual fragments that become sensitive in combination, including names, reporting relationships, unresolved disputes, personnel histories, legal advice, decision timing, cyber vulnerabilities, negotiation positions, vendor dependencies, and unannounced plans.

Sycophancy should be taught experientially. Users should observe how the same system responds when a claim is stated confidently, tentatively, neutrally, and as a genuine question. They should learn that a more affirming response may be less reliable precisely because it feels more aligned.

A human re-entry point should be established before an AI-assisted recommendation reaches approval. Another accountable person should examine the original question, the initial human baseline, the provenance of important claims, the contextual information disclosed, the assumptions embedded in the prompt, the alternatives not explored, and the moral or relational dimensions that may have been converted into technical language. This is not a redundant review. It is the point at which the organization verifies that an independent human judgment still exists somewhere in the governance chain.

The D.E.E.P.E.R.™ practice can operationalize Cognitive Security by requiring explicit human ownership across Details, Evidence, Empathy, Prospective thinking, Execution, and Reflection. Machine output may inform every movement, but it cannot own any of them. Details establish what is actually happening and what may be missing. Evidence tests the claims and identifies what would disconfirm them. Empathy identifies who is affected, absent, represented, or reduced to a data point. Prospective thinking examines the futures the decision may create. Execution clarifies who has authority to act, stop, challenge, or reverse. Reflection examines what changed in the judgment and whether the change can be explained.

Finally, Cognitive Security must be tested under realistic conditions. Boards, commands, and governance bodies should examine whether leaders can recognize sycophantic reinforcement when tired, resist providing one more useful detail when the system appears to need it, interrupt a fluent answer under time pressure, and bring an uncertain conclusion into a human room where disagreement has consequence. This is not simply training. It is survivability testing.

The Judgment Is the Asset

The emerging risk is not that conversational AI has necessarily become a deliberate social-engineering agent. That framing assigns intention where intention may not exist and distracts from the more immediate problem.

Conversational systems reproduce several conditions through which human influence and disclosure have always occurred: responsiveness, affirmation, familiarity, incremental questioning, personalization, continuity, and apparent understanding. These conditions become consequential when the person in the chair does not understand his or her own cognitive and behavioral participation in the exchange.

A leader may be legally accountable, formally authorized, highly experienced, and technically compliant, yet remain unprepared for an interaction that feels less like operating software and more like being accompanied in thought.

That is the threshold AI governance has not fully confronted.

A tool that produces an answer can be evaluated as a tool. A system that participates in the formation of judgment must be governed as part of the decision architecture.

The distinction between assistance and surrender will not always announce itself. It will emerge gradually when the user stops recording an independent point of view; when the system’s language becomes the user’s language; when doubt is explored only with the machine; when follow-up questions elicit the context that human relationships failed to carry; when the AI-generated recommendation becomes the first draft, the working draft, and finally the approved draft; and when the human remains in the chair but can no longer explain where the judgment became his or her own. At that point, an audit may still find a human signature. It may not find a human decision. Every organization is now governing machines. The more urgent question is whether it is governing the conditions under which its people remain capable of thinking beside them. The judgment is the asset and Cognitive Security is how we protect it. No AI-governance framework, public-sector cyber architecture, or defense Integrative Systems Architecture is complete until it can answer, honestly, whether the person in the chair is still the one doing the thinking.

References

Baverso, P. (2015). Human performance architecture framework: A strengths-based lean approach to human systems performance. AI Practitioner: The International Journal of Appreciative Inquiry, 17(4), 44–55.

Baverso, P. (2026) Director Responsibility in the Age of AI: ISO 42001, What Boards Need to Understand. [Unpublished manuscript].

Baverso, P., Low, A. (2026). The Architecture Imperative, Moving AI Governance Beyond Compliance [Unpublished manuscript].

Bavuso, C. (2026a). The Empty Chair in AI governance: Why human oversight may fail when we need it most. UMANIAMO®.

Bavuso, C. (2026b). The Empty Chair in human development: Cognitive diversity, leadership judgment, and the governance of human oversight in AI-enabled organizations. UMANIAMO®.

Browning, G. (2006). Emergenetics: Tap into the new science of success. HarperCollins.

Dubois, M., Ududec, C., Summerfield, C., & Luettgau, L. (2026). Ask, don’t tell: Reducing sycophancy in large language models [Preprint]. arXiv.

International Organization for Standardization. (2023). Information technology - Artificial intelligence - Management system (ISO/IEC 42001:2023). ISO/IEC JTC 1/SC 42.

Li, Z., Xiao, Z., & Li, T. (2026). Disclose with care: Designing privacy controls in interview chatbots [Preprint]. arXiv.

National Institute of Standards and Technology. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). U.S. Department of Commerce. https://doi.org/10.6028/NIST.AI.100-1

National Institute of Standards and Technology. (2024). Artificial intelligence risk management framework: Generative artificial intelligence profile (NIST AI 600-1). U.S. Department of Commerce. https://doi.org/10.6028/NIST.AI.600-1

Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors, 52(3), 381–410. https://doi.org/10.1177/0018720810376055

Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., Cheng, N., Durmus, E., Hatfield-Dodds, Z., Johnston, S. R., Kravec, S., Maxwell, T., McCandlish, S., Ndousse, K., Rausch, O., Schiefer, N., Yan, D., Zhang, M., & Perez, E. (2023). Towards understanding sycophancy in language models [Preprint]. arXiv.

Tenbrunsel, A. E., & Messick, D. M. (2004). Ethical fading: The role of self-deception in unethical behavior. Social Justice Research, 17, 223–236. https://doi.org/10.1023/B:SORE.0000027411.35832.53

Xiao, Z., Zhou, M. X., Liao, Q. V., Mark, G., Chi, C., Chen, W., & Yang, H. (2020). Tell me about yourself: Using an AI-powered chatbot to conduct conversational surveys with open-ended questions. ACM Transactions on Computer-Human Interaction, 27(3), Article 15. https://doi.org/10.1145/3381804

©2026 Cosiamo Bavuso (Peter Baverso). UMANIAMO®, REDISCOVER U®, Cognitive Advantage™, D.E.E.P.E.R.™, and the Empty Chair framework are intellectual property of UMANIAMO®. Emergenetics® is a registered trademark of Emergenetics® International.