the Empty Chair
Wednesday Table → The Empty Chair
The "Empty Chair," a critical vulnerability in artificial intelligence governance where the human supposedly overseeing automated systems lacks the cognitive capability, behavioral authority, or relational safety to actually intervene. Through a fictional organizational case study of a struggling international non-profit, we demonstrate how well-formatted machine outputs can cause committees to rubber-stamp decisions without exercising genuine judgment.
Our new paper, The Empty Chair in AI Governance, examines the conditions that impede informed human judgement and disciplined action.
Everyone Was Accountable. Nobody Was Deciding.
The President of United International Development (UID) had always believed in consultation. It was one of the qualities for which she was most admired. She listened carefully, invited participation, acknowledged different perspectives, and rarely allowed a meeting to end with anyone feeling openly dismissed. In an international organization with members across dozens of countries, these were not insignificant accomplishments.
She also had a remarkable ability to conclude a meeting without anyone being entirely certain what had been decided. This morning, however, she intended to be decisive.
UID was approaching its seventieth anniversary. Membership had declined for five consecutive years. The annual global conference, once the centerpiece of IUD's identity, had no confirmed host. Several national member organizations had stopped responding to membership renewal notices, and the organization's latest certification initiative had attracted considerable interest but few paid registrations. The Board had discussed these matters repeatedly. Committees had submitted reports, volunteers had offered recommendations, and a strategic planning group had spent three months preparing a document entitled UID 2030: A Global Platform for Human Capability.
It was time to act.
The President opened her laptop and asked her AI assistant to prepare a comprehensive executive recommendation. Across the table, the Devil at Work adjusted his cuffs and smiled. He had been waiting seventy years for an invention like this.
The President's prompt was thoughtful and detailed. She described UID's history, charitable purpose, international membership, committee structure, financial pressures, and strategic aspirations. She explained that the organization needed greater efficiency, stronger global coordination, improved member engagement, and sustainable revenue.
Then she asked whether a centralized digital platform, supported by a streamlined governance structure and an expanded certification portfolio, would provide the best foundation for UID's next decade.
The response arrived in less than a minute. It was impressive.
The AI assistant identified several opportunities to reduce duplication, standardize member services, improve reporting, expand digital learning, and strengthen UID's international brand. It proposed a phased transformation program with governance milestones, performance indicators, risk controls, and a compelling vision of the federation's future.
The President read it twice.
For months, she had struggled to explain how the various initiatives belonged together. Here, finally, was a coherent strategy. She asked the AI assistant to challenge its own recommendation.
It produced a balanced assessment identifying risks associated with change resistance, funding, implementation complexity, and stakeholder engagement. Each risk was accompanied by a practical mitigation strategy.
She asked whether the benefits outweighed the risks.
The AI assistant concluded that the proposed approach was promising, provided implementation was carefully managed. The President felt reassured. The analysis had considered both sides.
She forwarded the document to the Executive Committee with a note explaining that she had used AI to synthesize the organization's challenges and develop an initial strategic recommendation. She invited everyone to review it before Friday's meeting.
The Devil at Work glanced at the screen.
He admired the prompt. It contained the organization's problems, the President's preferred solution, and just enough room for the machine to make the solution sound independently discovered. He was particularly pleased with the risk assessment.
Nothing makes a recommendation more convincing than a list of reasons it might fail, followed by a list of reassuring ways to prevent those failures.
The Treasurer received the document late that evening. He was a careful man with considerable financial experience, although UID's accounts had become increasingly difficult to interpret. Revenue from conferences fluctuated unpredictably. Membership dues were recorded by region, but several national member organizations operated under informal arrangements negotiated years earlier. Some programs generated income while others depended on volunteers, donated facilities, or the goodwill of partner organizations.
The Treasurer had been asking for a reliable program-level financial picture for nearly two years. He still did not have one.
The proposed strategy nevertheless made financial sense at first glance. Shared services could reduce duplication. Digital delivery could extend UID's reach. Certification might create recurring revenue. A centralized platform could make reporting more consistent.
He opened his own AI assistant and uploaded the proposal.
"Evaluate the financial sustainability of this strategy," he wrote. "Identify the principal risks and recommend the strongest business case for implementation."
The AI assistant returned a sophisticated analysis with revenue scenarios, cost categories, investment assumptions, and a proposed three-year financial trajectory.
Some of the assumptions seemed optimistic. The model anticipated growth in certification revenue that UID had never achieved. It assumed that a significant proportion of members would adopt new paid services, although no one had surveyed them about willingness to pay. The proposed administrative savings depended upon eliminating duplication that the Treasurer could not yet quantify.
He asked the assistant to make the assumptions more conservative.
The revised projection looked better. Not more profitable, necessarily, but more credible. The numbers were less ambitious, the risks more clearly stated, and the recommendation appropriately cautious.
He could work with that.
At the meeting, he would explain that the proposal appeared financially feasible, subject to detailed due diligence. He did not intend to mislead anyone. In fact, he believed he was doing precisely what a responsible Treasurer should do.
He saved the analysis and closed his laptop.
He had spent ninety minutes examining the financial model. He had not contacted the person who maintained UID's accounts. The Devil at Work considered this an efficient use of executive time.
The Director of Programs was delighted by the proposed certification expansion. UID had been discussing international accreditation and certification for years. Several member organizations offered their own programs, and IUD had long considered whether a globally recognized credential might create new value.
The Director saw an opportunity to bring those conversations together. She asked AI to develop a portfolio of internationally relevant certifications aligned with UID's strategic priorities. Within minutes, she had an attractive catalogue covering leadership, digital transformation, sustainability, AI readiness, organizational development, and workforce capability.
The descriptions were excellent. Each program addressed a recognizable need, promised practical outcomes, and included a suggested audience, delivery format, assessment approach, and pricing structure.
She forwarded the catalogue to the Standards and Certification Council.
The Council's chair replied that the proposals looked promising but that the organization needed to distinguish between participation certificates, professional certification, and accreditation of external programs.
The Director agreed. She asked AI to prepare a framework clarifying the distinctions. The resulting document was comprehensive. It contained definitions, eligibility requirements, assessment criteria, governance arrangements, and quality assurance procedures. It even proposed an independent review panel.
The Director sent it back to the Council.
The chair thanked her and suggested establishing a small working group to examine implementation.
The Director added the recommendation to her Friday presentation.
Nobody had yet determined which capabilities UID was uniquely qualified to certify, whether the proposed programs duplicated those already offered by its members, or who would be responsible for evaluating competence.
But the documentation was advancing nicely.
The Devil at Work had always appreciated a good framework. He particularly enjoyed frameworks that generated further frameworks.
The Director of Membership was less enthusiastic.
She had spent the previous month reviewing correspondence from national member organizations that had not renewed their memberships. The messages were uneven. Some cited financial constraints. Others mentioned changing priorities. Several had not responded at all. One longtime member had written a short, unusually candid note.
UID had once been an important source of international relationships, professional exchange, and access to expertise. Over the years, the member explained, the organization had become less relevant to the practical challenges facing its organization. The benefits were familiar, but the reasons to remain involved were no longer compelling.
The Director had drafted a personal reply asking what had changed and whether the member would be willing to speak. Before sending it, she had shown the draft to the President.
The President appreciated the initiative but suggested a more consistent approach. UID needed to communicate the value of membership clearly and professionally. The Director was asked to develop a standardized renewal message highlighting networking opportunities, international recognition, webinars, certification discounts, and access to the global community.
The Director had done so. The campaign had generated a few replies and no immediate renewals.
Now the new strategy promised a digital member platform, personalized communications, automated renewal reminders, and AI-assisted engagement.
The Director wondered whether any of this addressed what the longtime member had been trying to say. She considered raising the question at Friday's meeting. Then she remembered the President's earlier advice about consistency. She was tired of being the person who complicated otherwise positive discussions.
She asked AI to summarize the membership correspondence and identify the principal causes of nonrenewal.
The AI assistant categorized the responses into familiar themes: cost sensitivity, perceived value, engagement, communication, and changing organizational priorities. It recommended segmented outreach, personalized benefits messaging, and improved member journeys.
The Director recognized the categories. They were reasonable enough. She added them to her report.
The personal reply remained in her drafts folder.
Across the room, the Devil at Work lifted his lantern toward the empty space between the Director's desk and the telephone. There was nothing there worth illuminating.
By Friday, the executive team was exceptionally well prepared.
The President had a strategic recommendation. The Treasurer had a financial assessment. Programs had developed a certification portfolio. Membership had prepared an engagement strategy. Operations had drafted an implementation roadmap. The Governance Committee had been asked to review the proposed decision structure.
The meeting began on time.
The President thanked everyone for their contributions and noted that the various analyses demonstrated encouraging consistency. "What's particularly reassuring," she said, "is that we've approached this from different perspectives and arrived at essentially the same conclusion."
The Treasurer explained that the strategy was financially plausible, subject to further validation. Programs described the potential for a globally recognized certification portfolio. Operations outlined the advantages of shared infrastructure. Membership presented the proposed digital engagement model. The presentations were polished and concise. The President invited comments.
A regional representative asked whether the national member organizations had been consulted about the proposed central platform. The Director of Membership explained that the organization had reviewed member feedback and incorporated stakeholder engagement into the implementation plan.
"Have we actually asked them whether this is what they need?" the representative asked.
The President thanked him for emphasizing the importance of consultation. "That's exactly why stakeholder engagement is such a central part of the proposal."
The representative looked at the slide. Stakeholder engagement appeared in Phase Two, after the operating model had been approved. He considered pointing this out. Instead, he said that he looked forward to seeing the consultation plan.
The President moved to the next speaker.
The Treasurer mentioned that the financial projections would require more reliable baseline data.
"Of course," the President replied. "That's why the implementation roadmap includes a financial validation workstream."
He nodded. The Director of Programs explained that the Standards Council would need to confirm the proposed certification architecture.
"Absolutely," said the President. "We should make that one of the first deliverables."
Operations noted that the central platform would require resources that UID did not currently possess.
"Which is precisely why we need a phased approach," the President answered.
Each concern found a place in the implementation plan. None changed the recommendation.
The Devil at Work sat beside the President, watching the concerns arrive and disappear into workstreams. It was a beautiful system. Every objection became an action item. Every uncertainty became a risk to manage. Every unanswered question became something the organization could resolve after deciding to proceed.
Nobody was being ignored. Everyone was being accommodated.
Near the end of the meeting, the regional representative spoke again.
"I have one question. What does UID actually need to do that our national member organizations cannot already do for themselves?"
The President smiled. "That's the heart of our strategy. We're building a global platform that connects and empowers our members."
"Yes, but what capability will they gain?"
The Director of Programs referred to certification. Membership mentioned international networking. Operations described shared infrastructure. The Treasurer spoke about economies of scale. All were plausible contributions.
The representative listened. "I understand the activities," he said. "I'm asking what becomes possible because UID exists."
The President glanced at the clock. They were already fifteen minutes behind schedule.
"I think that's an excellent question for the Strategic Development Committee," she said. "Perhaps we can ask them to develop a concise statement of our unique value proposition."
The Secretary recorded the action.
The Strategic Development Committee had not met since the previous quarter. Its chair was currently serving as interim chair of two other committees.
The representative muted his microphone.
The Devil at Work made a small adjustment to the lantern. He did not care for that question. Fortunately, it had been assigned to a committee.
The President returned to the recommendation. She proposed that the Executive Committee endorse the strategic direction, authorize the development of a detailed implementation plan, and recommend the approach to the full Board.
The Treasurer asked whether the endorsement would commit UID to any immediate expenditure.
Not at this stage, the President assured him.
Operations asked whether the approval included authority to begin discussions with technology vendors.
Only exploratory discussions.
Programs asked whether the certification portfolio could be announced as part of the anniversary campaign.
The President suggested describing it as an emerging initiative.
Membership asked whether the new strategy should be included in the renewal campaign.
Yes, provided the language remained appropriately aspirational.
The Secretary read back the motion. It contained the words endorse, develop, explore, advance, support, and recommend.
Nobody objected and the motion passed unanimously.
The President congratulated the team. "Excellent. We have alignment."
The Devil at Work allowed himself a quiet laugh.
The Treasurer believed he had approved further financial analysis. Programs believed it had permission to develop the certification portfolio. Operations believed it could begin planning the platform. Membership believed it could announce a new direction to prospective members. The President believed the executive team had endorsed a transformation strategy.
Each interpretation was reasonable. None was quite the same. The minutes would record unanimous agreement.
On Monday, the Director of Programs circulated an announcement describing UID's forthcoming global certification initiative. On Tuesday, Operations contacted three technology providers to discuss the proposed digital platform. On Wednesday, Membership distributed a renewal message promising members access to an expanded international learning and certification ecosystem.
On Thursday, the Treasurer received a request to approve preliminary expenditures associated with the anniversary launch. He replied that the Executive Committee had not yet approved a budget.
Programs responded that the committee had endorsed the strategic direction. Operations noted that vendor discussions were exploratory and therefore did not constitute procurement. Membership explained that the announcement had been approved in principle.
The President asked the Secretary to clarify what had been agreed.
The Secretary forwarded the minutes. Everyone had a copy. The language was clear. The meaning was not.
A short email exchange followed. The President proposed a coordination meeting to establish responsibilities and clarify the next steps.
The Secretary created a scheduling poll. Seven executives were invited. Three were unavailable for the proposed dates. Two suggested adding their committee chairs. One asked whether the Governance Committee should participate.
The President agreed that broader representation would be helpful. The meeting was postponed until the following week.
Meanwhile, the longtime national member organization that had questioned the value of UID membership received the new renewal campaign. The email announced an exciting transformation, a global digital platform, expanded certification opportunities, enhanced international networking, and a renewed commitment to human capability. It was accompanied by a professionally designed graphic and a personal salutation generated from the membership database.
The national member’s director read the message between meetings.
He had been involved with UID for more than twenty years. He remembered when United International Development had helped his organization establish its first professional development program. He remembered the people who had traveled across borders to teach, advise, and share what they knew. He remembered returning from international conferences with practical ideas that his colleagues could use immediately.
He had not expected UID to remain unchanged. He had hoped it would remain useful. He read the new list of benefits. Most were variations of things the association already had access to elsewhere.
He opened his earlier correspondence to see whether anyone had replied to his concerns. There was nothing beyond the standard acknowledgment. He considered writing again. Then he closed the message and returned to his work. He had a training program beginning the next morning, and several young professionals were depending on him to be ready.
UID's membership database recorded his status as Renewal Pending. The Director of Membership would see that status in her dashboard. She would not see the moment he stopped waiting for a reply.
Two weeks later, the President presented the strategy to the full Board. The document had grown considerably. It now included financial scenarios, a certification portfolio, a digital transformation roadmap, stakeholder engagement activities, a governance review, a communications plan, and a risk register.
AI had helped prepare much of it. The President was transparent about that fact. She explained that the technology had accelerated research, strengthened analysis, and allowed a small volunteer-led organization to produce work that would previously have required considerable external support.
Several trustees were impressed. One remarked that UID was finally demonstrating the kind of innovation it had been encouraging its members to adopt. Another asked whether the Executive Committee had independently validated the recommendation.
The President explained that each functional leader had reviewed the strategy from their area of expertise and that the committee had endorsed it unanimously.
The Treasurer confirmed that financial due diligence remained outstanding. Programs confirmed that certification standards would be finalized during implementation. Operations confirmed that technology requirements would be established through the vendor assessment. Membership confirmed that national associations would be consulted as part of stakeholder engagement.
The Board Chair thanked everyone for the thorough preparation.
A trustee from Singapore, who had previously questioned what capability UID actually provided its members, raised her hand. "Before we vote," she said, "can someone tell me what we have decided not to do?"
The President looked toward the strategic plan. There were six strategic priorities, eleven supporting objectives, and thirty-four proposed initiatives. Some existing programs would eventually be consolidated. Others would be modernized. New opportunities would be explored.
The President explained that portfolio rationalization would be addressed during implementation.
The trustee studied the document. "So we're approving the strategy before deciding what we're willing to stop?"
"We're approving the direction," the President replied. "The details will follow."
Several trustees nodded. That sounded sensible.
The trustee considered asking another question. She had been on the Board long enough to know that repeated objections could easily become a reputation. She was already regarded as someone who focused too much on details when others were trying to think strategically. She lowered her hand.
The Board approved the recommendation. The vote was unanimous.
That evening, the President received a congratulatory message from the Board Chair. The organization had taken an important step forward. The Chair praised her inclusive leadership, the quality of the strategic work, and the responsible use of emerging technology.
The President felt genuinely relieved. She had worked hard. So had everyone else. She had listened, consulted, invited challenges, documented risks, and obtained formal approval. She had followed the organization's governance procedures carefully.
The Devil at Work would have agreed with every word. He had no objection to accountability. He preferred it distributed across enough people, committees, policies, and documents that no one had to confront the gap between formal responsibility and actual judgment.
He had no objection to evidence, provided it arrived after the preferred conclusion had already taken shape. He welcomed consultation, particularly when those consulted could not meaningfully change the decision. He admired independent reviews that began with the same assumptions. And he loved unanimous votes from people who had privately decided that further disagreement was not worth the effort.
He had watched these habits develop at UID long before artificial intelligence arrived. AI had not created the organization's reluctance to ask difficult questions, its dependence on committees, or its preference for accommodation over genuine decisions. It had simply made those habits more “productive.”
The organization could now generate recommendations faster, document its reasoning more thoroughly, and present increasingly sophisticated evidence of progress. The Devil at Work could hardly have designed a better lantern.
Late that night, the President reopened the Board presentation.
The final slide read:
UNITED INTERNATIONAL DEVELOPMENT 2030
One Federation. One Vision. One Global Future.
She was proud of the work. Then she noticed a comment left in an earlier draft by the Director of Membership. It was attached to the section describing the proposed digital member platform.
"Should we speak with the member organizations that have stopped renewing before we decide what to build?"
The comment had been marked Resolved. The President could not remember who had resolved it. She opened the document history. There had been so many revisions, comments, summaries, and AI-assisted edits that the original discussion was difficult to reconstruct.
She considered sending a message to the Director of Membership.
It was late. The strategy had been approved. The Board was enthusiastic. The organization needed to move forward. She closed the document.
In the quiet office, the Devil at Work sat in the chair opposite her desk. His dark suit was immaculate, his expression pleasant, and his lantern cast a warm circle of light across the approved strategic plan.
He had no reason to speak.
The President had done nothing obviously irresponsible. Every committee had been consulted. Every executive had contributed. Every concern had been recorded. The recommendation had been challenged, revised, endorsed, and approved.
The institution had followed its procedures. And somewhere along the way, the questions that might have changed its direction had been assigned, deferred, summarized, softened, or quietly abandoned.
The President was still in her chair. The Board was still in its chairs. Every committee had a chair. Every decision had an approval record.
The Devil at Work looked at the empty chair beside him. It was the only one in the organization that nobody had noticed and he intended to keep it that way.
United International Development is a fictional international NGO. The Devil at Work is a recurring Wednesday Table story examining the illusions, habits, and institutional arrangements through which capable people can surrender their judgment without realizing it.
We are human together or not at all.
Cosiamo | Founder, UMANIAMO® | Humaning Architecture
Why Human Oversight May Fail When We Need It Most
Cosiamo (Peter) Baverso
Founder, UMANIAMO® · Creator of Cognitive Advantage™ and REDISCOVER U®
Every serious artificial intelligence governance framework now has some version of the same thing: keep a human in the loop. The phrase is reassuring. It suggests accountability, judgment, and control. It suggests that when an AI system produces a recommendation, a real person will be there to review it, challenge it, and stop it if necessary.
In this paper, artificial intelligence is understood as machine inference: computational systems that detect patterns, generate outputs, recommend actions, and simulate certain features of reasoning without conscience, responsibility, embodiment, or wisdom. Machine inference does not mimic or refer to intelligence in the human sense.
However, there is a problem. Most organizations have not tested whether the human in the loop is actually able to exercise judgment when it matters. Their leadership assigned responsibility, created committees, implemented policy, and documented approval processes. They may have even trained staff on AI risk, ethics, privacy, bias, and compliance.
They have not asked a deeper question: When the system is wrong, who in the room is cognitively likely to notice and behaviorally able to stop the decision?
That is what I call the Empty Chair.
The Empty Chair is the absence of a clearly designated owner for the formation, sustainment, and exercise of human judgment in AI-enabled decision systems. It is the gap between the person formally assigned to oversight and the actual conditions required for oversight to work.
This paper argues that the central risk in AI governance may not be the algorithms alone. It may be the architecture of the human system surrounding it. A government agency can comply with AI governance requirements and still leave the chair empty. The danger is not only that the machine may be wrong. The danger is that the machine may make the human feel less responsible for finding out. That should make every senior official very uncomfortable.
The Comforting Fiction of “Human in the Loop”
Imagine the following scenario.
An AI-enabled system produces a recommendation. The output is fluent, confident, well-structured, and consistent with available data. The recommendation is reviewed by a designated official and then discussed in committee; all of the required documentation is completed and the process satisfies existing policy. The decision is made, and then, weeks or months later, the decision is found to be wrong.
The error did not occur because no human was present at the time of the decision, humans were present. The failure occurred because the humans present were not positioned to exercise judgment. Some lacked the cognitive disposition to detect the issue. Others sensed something was wrong but lacked the behavioral authority to interrupt momentum. Still others assumed someone else had already checked.
The governance process did not fail visibly. It failed without notice, until the consequences became undeniable, yet hopefully in time to avert catastrophe. This is the uncomfortable reality behind many governance systems: they are designed to demonstrate that a human occupied the chair, but not that the judgment required by the chair was actively exercised. The distinction is subtle, but it is the difference between governance as documentation and governance as practice. The RMS Titanic did not sink to the bottom of the Atlantic, because the bridge was empty. It suffered from an empty moment, a moment in which warnings, assumptions, authority, and decision-making failed to converge quickly enough to change the outcome. AI governance presents a similar challenge. Regulations may require human oversight, but oversight is not the same as human judgment. Presence can be documented. However, human judgment must be cultivated, sustained, and exercised.
The phrase “human in the loop” hides this distinction. It allows organizations to satisfy the literal requirement of oversight while leaving the functional requirement unexamined. The question is not whether a human is in the loop. The real question is whether the human system surrounding the Empty Chair has the capacity, permission, and discipline to intervene when intervention is required.
The Empty Chair is a structural condition in which no person, role, or process clearly owns the conditions necessary for human judgment to function. It appears when organizations assume that because a person has been assigned to review a decision, meaningful oversight has occurred. But meaningful oversight requires more than assignment. It requires at least seven conditions:
Cognitive Capability: someone must be able to notice a flaw, anomaly, inconsistency, or unexamined assumption.
Behavioral Authority: that person must have the standing, confidence, or formal permission to interrupt the process.
Relational Safety: the organizational governance systems must tolerate challenge, dissent, and hesitation without punishing the person who raises it.
Decision Discipline: the process must require unanswered questions to be resolved before momentum carries the decision forward.
Cognitive Sustainment: the deliberate, ongoing maintenance of an individual's or group's capacity for independent judgment: the practiced ability to generate original assessment, detect anomaly, and resist convergence toward a dominant or externally-supplied conclusion, treated as a capability requiring continuous exercise rather than a trait assumed to persist once established.
Survivability: the capacity of an individual's or group's judgment to remain intact and exercisable under the specific conditions most likely to degrade it: time pressure, confident contradictory input, social cost for dissent, and prolonged low-stakes deference that erodes the habit of intervention before a high-stakes moment requires it.
Relational Sustainment: the deliberate, ongoing maintenance of the trust, mutual legibility, and psychological safety required for judgment, once exercised, to actually be transmitted, heard, and acted upon across a group or hierarchy, treated as its own capability subject to its own atrophy, independent of whether individual cognitive capacity remains intact.
Most AI governance frameworks remain heavily focused on system-side controls: model validation, bias testing, audit documentation, risk classification, data governance, and compliance procedures. These controls are necessary. They do not answer the more fragile question: whether the human system responsible for oversight is actually capable of exercising discernment when it matters.
The human side of oversight remains dangerously underdeveloped. We have built governance structures around the comforting assumption that judgment will simply appear when needed and that the right person will see the issue, identify it as such, and act with enough authority to interrupt momentum. That assumption is increasingly unsafe.
As AI systems become more fluent, persuasive, and embedded in decision workflows, the gap between unbiased discernment, sound judgment, and aligned action becomes more unstable. The risk is not that only the machine may be wrong. The risk is that the human system around the machine may no longer be formed, positioned, or authorized to notice, question, and act.
Government decisions carry public consequence. A flawed decision in a private firm may affect revenue, reputation, or operational performance. However, a flawed decision in government may affect eligibility, benefits, enforcement, surveillance, public safety, immigration, healthcare access, education, procurement, employment, policing, or national security. When AI systems enter these environments, the human oversight problem becomes more serious, not less.
Government agencies are especially vulnerable because they often operate within: complex reporting chains, formal hierarchies, risk-averse cultures, compliance-heavy procedures, fragmented ownership, contractor-supported systems, and politically sensitive decision environments.
In such systems, many people may touch a decision, but no one may fully own the judgment architecture around it. A senior official may believe oversight exists because the organizational chart or project charter says that it does. But the chart may be wrong, and the human system may contain a gap that cannot be seen from the organizational chart, governance or RACI chart, policy manual, or dashboard. That gap or seam is the Empty Chair.
Most AI readiness and governance efforts ask three familiar questions: Is the system technically reliable? Is the system legally and procedurally compliant? Is there a human in the loop? This paper proposes a fourth question: Is the human in the loop actually positioned to exercise judgment?
This question radically changes the conversation, not in terms of unfairly judging the person, but by moving AI governance from procedural assurance to human capability within the context of the organization itself. It asks whether the people charged with oversight have the cognitive variety, behavioral authority, relational trust, and structured process required to challenge machine output.
Any system can be technically sound and still be misused. A process can be compliant and still be hollow of value. A human can be present and still be functionally absent. That is why the fourth question matters.
Cognitive variety is often treated as a team-building or communication topic. That framing is too small for the AI era. In AI governance, cognitive variety is infrastructure.
Different people notice different risks. Some are more likely to ask for evidence. Some are more likely to see second-order consequences. Others are more likely to detect procedural gaps. Some are more likely to sense relational harm, public trust implications, or political risk. Some are more likely to imagine future misuse, mission drift, or future benefit. Others are more likely to ask whether a decision can actually be implemented.
No single leader naturally sees all of this. No committee automatically covers all of it. And there is no AI system that can substitute or proxy for a properly formed human deliberative group.
Using the Emergenetics® attributes as one applied lens, the relevant attributes can be described across four thinking patterns and three behavioral patterns:
Analytical thinking: What do the facts show? Where is the logic? How much will this cost? Show me the value.
Structural thinking: What are the steps, risks, details, and controls? Is there a project plan, and what are the gates?
Social thinking: Who is affected, and how? Who else needs to be in the room? Who else has missing information?
Conceptual thinking: What future are we creating? Is this part of our mission? What will the future look like if we don't do this?
Expressiveness shapes whether insight is voiced or held in reflection.
Assertiveness shapes whether challenge interrupts momentum, speeds it up, or waits for everyone to agree.
Flexibility shapes whether the group can adapt when new evidence appears, or commit to a plan and stick with it.
The oversight risk emerges when the attributes most needed to detect a problem are underrepresented, unheard, or behaviorally unable to act. For example, a group may have someone who can see the flaw but cannot stop the room from acting. Another group may have decisive authority but lack the cognitive disposition to notice what needs to be challenged or adapted. Both conditions represent the Empty Chair.
Compare those to an accessible cultural analogy is the character of Captain James T. Kirk in Gene Roddenberry’s Star Trek. The person in the chair is not a passive monitor or a checkbox reviewer. The archetype is a human integrator: someone who listens to the people surrounding him, attends to the human system architecture of the bridge, questions the machine, trusts capable specialists, weighs incomplete information, and still accepts responsibility for the ultimate decision. The Chair requires judgment, courage, and consequence-bearing authority. Without those conditions, the Empty Chair remains empty even when someone is physically present.
Many facilitators are trained to read a group's health through a familiar progression: a team begins by forming, moves through a period of friction often called storming, settles into what is usually called norming, and finally reaches performing. The vocabulary is common enough that most leaders have heard it, even if they could not name where it came from.
That third stage deserves closer scrutiny than it usually receives. “Norming” is generally read as convergence: a group settling its differences, finding common ground, smoothing out disagreement. A quiet, orderly, single-voice-at-a-time meeting is typically read as the sign of a team that has finally matured.
This reading should concern any official thinking seriously about AI governance, for a simple reason: AI output is itself a catalyst of convergence. A model returns the most statistically probable answer, which is the smoothed center the bell curve of its training data. If a human group has also been culturally trained to treat the disappearance of dissent as the mark of a well-functioning team, then the group and its AI tool are now optimizing for the same outcome, and reinforcing each other while doing it. The model converges. The room, watching its own disagreement quiet down, reads that as success. Nobody in the room has a reason to be suspicious of the quiet, because the quiet is what they were taught a healthy meeting sounds like.
A genuinely cognitively diverse group does not necessarily sound quiet when it is functioning well. It sounds like every distinct way of thinking in the room finding its function: the evidence-seeker still asking for evidence, the relationship-reader still asking who is affected, the systems-thinker still asking what comes next, simply without those differences competing for the same airtime they once did. From the outside this can look calm. It is something closer to difference that has found its place, not difference that has disappeared.
The practical caution for any AI oversight body is this: a quiet, agreeable review meeting is not, by itself, evidence that judgment was exercised. It may be evidence that judgment was exercised well. It may equally be evidence that dissent has already been priced out of the room and that the AI's confident, convergent output will meet no resistance at all, not because it is correct, but because resistance was never going to be voiced in that room regardless of what the output said.
Across applied work with governance bodies, enterprise HR leaders, public-facing institutional systems, and cross-organizational cohorts, a repeated pattern emerges: the people most likely to detect certain categories of risk are not always the people most likely to be heard.
This is not a problem of intelligence or goodwill. It is a design problem.
In several examined groups, leadership bodies were heavily concentrated in Social and Conceptual thinking. These groups were often strong in relationship-building, vision, communication, and stakeholder sense-making. Those are real strengths. But the same groups were comparatively weaker in Analytical and Structural scrutiny: the capacities most associated with evidence challenge, procedural discipline, detail verification, and systematic follow-through.
That creates a predictable vulnerability when AI enters the room. AI output often arrives in a form that appears highly Analytical and Structural: it is ordered, confident, formatted, and it sounds complete. For a group that does not naturally live in those thinking patterns, the output may feel more authoritative than it deserves to be. The system appears to supply the very capability the group lacks. That is where automation bias becomes not merely a technical concern, but a human architecture concern.
Two additional risks make the Empty Chair more dangerous in AI-enabled organizations: automation bias and ethical fading. Automation bias is the tendency to over-trust automated system output, especially when that output appears consistent, confident, data-rich, or technically sophisticated. In practice, this means a recommendation produced by an AI system may be granted more authority than it has earned simply because it looks ordered, precise, and complete.
This is especially dangerous in bureaucratic environments.
A well-formatted AI recommendation can appear to carry the weight of analysis even when its assumptions are incomplete, its data appears biased, or its reasoning cannot be independently verified. The output arrives with the visual and linguistic signals of competence: categories, rankings, probabilities, summaries, risk scores, and recommended actions. The human in the loop may not think, “the machine is [morally] right.” They may simply assume, “this has already been analyzed” and that is enough to make the moral question and, more importantly, the moral authority vanish from the decision.
Once that assumption enters the room, genuine oversight begins to weaken: questions become checklist items; there is less dissent or rigorous interrogation; review feels slower to start and faster to finish and then the final review becomes confirmation. Afterwards, the human in the loop becomes a human near the loop.
Ethical fading compounds the problem. Ethical fading occurs when the moral dimensions of a decision disappear from view because the situation is framed in technical, procedural, financial, or administrative terms. This occurs when a decision about a person becomes a case; a case becomes a data point; a data point becomes an output; an output becomes a recommendation; and a recommendation becomes a workflow step. In this case, logic of a linear process is unexamined and while at no point does anyone explicitly decide to ignore human dignity, the moral and ethical dimensions simply fade into a workflow.
This is one of the most serious dangers of AI-enabled bureaucracy: not open cruelty or inhumane, but vague abstraction. The system does not need officials to become unethical, only to follow the procedures. Some examples:
Benefits determination becomes “eligibility optimization.”
Child welfare alerts become “risk stratification.”
Hiring, promotion, or termination decisions become “candidate ranking.”
Public safety interventions become “predictive resource allocation.”
Healthcare denials become “utilization management.”
The language changes, and with it the felt responsibility of the decision-maker. Automation bias tells the brain: the system probably got this right. Ethical fading tells the brain: this is not really a moral choice. Together, they create the perfect conditions for the Empty Chair.
The person who might ask, “Is this right?” is replaced by a process question: “Was this reviewed?” The person who might ask, “Who is affected?” is replaced by a dashboard showing “impact.” The person who might ask, “Should we stop?” is replaced by a workflow with a pending approval task.
This is why human oversight cannot be reduced to presence, review, or sign-off. The problem is not merely whether a human sees the AI output. The problem is whether the decision environment preserves the human capacity to recognize that a moral judgment is being made. A government agency can satisfy procedural oversight requirements while allowing both doubt and conscience to fade from the room. That is not governance. It is ethical risk disguised as administrative efficiency.
The most dangerous Empty Chair is not simple absence, rather misalignment. One person may have the capability to detect the issue, while another may have the authority to act. If those are not connected by process, trust, or role design, oversight fails. This is especially important in hierarchical organizations.
Frontline staff often see operational reality first. They notice procedural inconsistencies, implementation risks, and user-facing consequences. But they may not have the authority or confidence to interrupt a decision moving through senior channels. Senior leaders may have authority, but not the ground-level pattern recognition required to see the problem early.
Each group holds part of the oversight function, yet neither holds the whole.
That is a bidirectional Empty Chair, and it looks like this: a staff member sees but cannot stop, and the executive can stop but does not see, and so the system proceeds. If AI governance does not explicitly design for this split, it will reproduce it over and over again.
The Empty Chair is common, but it is not inevitable. One observed organization provides a useful contrast. Rather than treating cognitive variety as a one-time workshop, it embedded cognitive-variety-informed facilitation into the way teams were composed, meetings were structured, and decisions were reviewed. This did not mean every group had perfect balance because that is not realistic. It meant the organization developed recurring practices to notice what and who was missing, and to deliberately invite those into the process.
In some cases, this was as simple as designating a literal empty chair in a meeting and labeling it with the missing perspective. If Analytical thinking was absent, the group had to ask: What evidence have we not examined? If Structural thinking was absent: What steps, risks, and controls are unclear? If Social thinking was absent: Who is affected? If Conceptual thinking was absent: What future are we creating?
This practice is simple, which is why it works. It turns an invisible governance vulnerability into a visible discipline. It makes the people in the room answer for what they do not naturally see.
A practical response to the Empty Chair is not more training alone. It is recurring reflective diversification. The D.E.E.P.E.R.™ method provides one structured way to do this by requiring a deliberative body to move through all seven Emergenetics® attributes before a decision cycle is considered complete:
Details: What actually happened? What is the process? What are the steps?
Evidence: What do the facts show, not only the “facts” produced by the machine? What assumptions are unsupported?
Empathy: Who is affected by this decision?
Prospective: What future does this decision create?
Execution: What action will be taken, by whom, when, and under what conditions?
Reflection: How did the group process the decision? Who spoke, who stayed silent, and what was avoided?
This practice does not replace technical AI governance, it strengthens it. The D.E.E.P.E.R. practice ensures that human oversight is not reduced to a signature, a review field, or a committee agenda item. It makes judgment a repeated organizational discipline.
For senior government officials responsible for digital transformation, AI readiness, or enterprise governance, the practical implications are direct. Before deploying or approving AI-enabled decision systems, leaders should ask:
Who is responsible for human judgment in this process?
Which cognitive perspectives are missing from the decision body?
Who is most likely to notice error, bias, or procedural weakness?
Does that person have the authority to stop or slow the decision?
Are dissent and hesitation treated as governance assets or bureaucratic irritants?
Is the “human in the loop” a real decision-maker or a procedural placeholder?
Have we tested the oversight process under conditions of pressure, ambiguity, and confident machine output?
If these questions have not been answered, documented, and embedded into organizational processes, the agency may have governance documentation without governance capability. That is not governance readiness. It is exposure.
We are not suggesting that the answer is to find better people. It is to design better human oversight architecture. This architecture includes:
1. Cognitive variety Audits: Map the cognitive and behavioral composition of boards, committees, review teams, and AI oversight bodies.
2. Cognitive Role Charters: Assign explicit oversight roles: evidence challenger, procedural verifier, stakeholder impact reviewer, future-risk examiner, implementation owner.
3. Decision Pause Points: Require structured challenge before high-impact AI-enabled decisions proceed.
4. Empty Chair Protocols: When a perspective is missing, make the absence visible and require the group to answer from that viewpoint before deciding.
5. Authority Alignment: Ensure the person most likely to detect a class of risk has a formal path to interrupt, escalate, or require reconsideration.
6. Recurring Reflective Practice: Use a structured method such as D.E.E.P.E.R.™ to prevent dominant cognitive habits from becoming governance blind spots.
None of these interventions are especially expensive. They are, however, uncomfortable. That discomfort is precisely why they are needed. AI-enabled systems often make organizations feel more confident, more informed, more efficient, and more in control. The outputs are clean, the summaries are fluent, the dashboards are orderly and always up to date, the workflow moves, and the water feels warm. That is the danger.
The old image of the frog in slowly boiling water is useful here, not as biology, but as bureaucracy. Most organizations will not lose human judgment all at once. They will acclimate to its absence. They will become comfortable with machine-generated confidence, comfortable with faster approvals, comfortable with fewer objections, comfortable with oversight that appears complete because the form was completed. By the time the risk becomes obvious, the organization may already have normalized the conditions that made real oversight unlikely.
The purpose of human oversight architecture is to interrupt that warming process before it feels dangerous. Cognitive variety audits, role charters, pause points, Empty Chair protocols, authority alignment, and recurring reflection all introduce deliberate friction into systems that would otherwise reward speed, fluency, and compliance over judgment. That friction is not inefficiency. It is the sound of human responsibility re-entering the room.
AI governance cannot stop at governing machines. It must also govern the human systems responsible for supervising them. The most important question may not be: Is there a human in the loop? The more important question is: What conditions must exist for human judgment to function when it matters most? The difference is not semantic. It is the difference between procedural reassurance and genuine oversight. Every AI governance framework ultimately rests on an assumption: that when a critical decision arrives, the right human being will recognize the issue and possess the authority to intervene. Few organizations have tested that assumption, and almost none have designed for it.
The Empty Chair is not a metaphorical problem. It is a governance vulnerability hiding in plain sight. Once seen, it becomes difficult to ignore, and once ignored, it becomes a choice. Machine inference is not human intelligence. Fluency is not wisdom and output is not judgment.
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].
Browning, G. (2006). Emergenetics: Tap into the new science of success. HarperCollins.
Cooperrider, D. L., & Whitney, D. (2005). Appreciative inquiry: A positive revolution in change. Berrett-Koehler.
International Organization for Standardization & International Electrotechnical Commission. (2023). ISO/IEC 42001:2023 - Artificial intelligence management system requirements. ISO.
National Institute of Standards and Technology. (2023). AI 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
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
Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics — Part A: Systems and Humans, 30(3), 286–297.
Roddenberry, G. (Creator). (1966–1969). Star Trek: The original series [TV series]. Desilu Productions; Paramount Television.
Appendix
Why AI Governance Must Look at the Physical World It Consumes
For many years I lived in Ashburn, Virginia, near what the world now knows as “Data Center Alley.” I still visit and have watched the landscape change — not through a single environmental event, but through accumulation: another graded parcel, another windowless building, another substation, another transmission line, another construction entrance, another road widened to accommodate the weight of a digital economy that most people experience as weightless. That is the strange thing about artificial intelligence. To the user, it feels almost immaterial: a prompt is typed, the response appears, a decision support system produces an output, the dashboard refreshes, and a recommendation arrives. The experience is clean, silent, and nearly frictionless.
But the cloud is not a cloud. It is land and in this case, formerly farmland transformed by concrete, steel, copper, fiber, transformers, switchgear, backup generators, cooling systems, fans, pumps, water lines, stormwater systems, electrical substations, and workers in boots, hard hats, reflective vests, and tool belts. It is heat produced at extraordinary density and managed every hour of every day so that servers do not overheat and fail. It is electricity drawn from a grid that was not originally built for this scale or speed of demand. It is water consumed directly or indirectly, depending on cooling design and energy source. It is waste heat released into the surrounding environment. It is equipment that ages, chips that are replaced, batteries that degrade, and infrastructure that must be maintained long after the ribbon-cutting.
This appendix is not an argument against data centers. They are now part of the infrastructure of modern life. Government, healthcare, finance, education, defense, emergency services, commerce, and ordinary family communication all depend on them. Nor is it an argument against artificial intelligence. The point is more basic and more uncomfortable: AI governance cannot remain purely digital when AI itself is materially embodied. Every AI decision rests somewhere on a physical system that someone built and someone powers, cools, maintains, and lives near. Someone pays for the grid upgrades, the water demand, the land-use tradeoffs, the noise, the heat, the roads, the substations, and the long-term consequences of building computation at industrial scale.
This matters for the Empty Chair because most AI governance discussions focus on model behavior, data quality, bias testing, documentation, privacy, cybersecurity, procurement, and compliance. Those are necessary questions. They are not enough. A government official responsible for AI readiness must also ask a more grounded question: What human and physical systems are being consumed to make this intelligence appear effortless?
A large data center is not simply a building full of computers. It is a thermal machine. Nearly every watt of electricity consumed by computing equipment becomes heat that must be moved somewhere else. The central operational challenge is not only computation, but heat removal. Some facilities use air-based cooling, others use chilled water systems, and others use evaporative or adiabatic cooling. There are newer application that are moving toward liquid cooling because AI hardware is becoming denser and hotter. Each design involves tradeoffs between electricity, water, efficiency, reliability, cost, and local environmental impact.
Water use is especially easy to misunderstand because it is local. A data center may use little water on site if it relies on waterless cooling, yet still carry an indirect water footprint through electricity generation. Another facility may use more water directly in order to reduce energy use. A design that makes sense in one climate, watershed, or power market may be irresponsible in another. There is no single abstract answer. There is only local accountability.
The same is true of labor. Data centers do not run on code alone. They depend on electricians, mechanical technicians, HVAC specialists, engineers, construction workers, commissioning experts, facility operators, logistics teams, and maintenance personnel. These are not symbolic roles. They are the people who make the invisible infrastructure visible again. They work around heat, electricity, mechanical systems, noise, heavy equipment, and constant uptime expectations. The public experiences digital convenience. The worker experiences physical responsibility.
This is where ethical fading enters the conversation.
When AI is framed only as software, the physical world disappears. Land becomes “capacity.” Electricity becomes “compute.” Water becomes “cooling.” Workers become “operations.” Neighborhoods become “sites.” Communities become “markets.” The moral question fades because the language has been abstracted.
But human governance requires re-embodiment. It requires asking what has been hidden by the interface.
Where does the power come from?
What happens during peak demand?
What watershed is affected?
What heat is produced?
What materials are required?
What waste streams are generated?
Who works there?
Who lives nearby?
Who benefits?
Who bears the cost?
Who has a chair when these decisions are made?
For a senior government official, this is not a philosophical exercise. It is a readiness issue. AI readiness that ignores physical infrastructure is not readiness. It is partial readiness. It prepares an agency to procure, deploy, document, and govern digital systems while underexamining the energy, water, land, labor, and community systems those digital systems require.
The old administrative habit is to divide these concerns into separate rooms. One office handles digital transformation. Another handles procurement. Another handles facilities. Another handles the environment. Another handles the workforce. Another handles utilities. Another handles community engagement. Each group has part of the picture. No one owns the whole. That is another form of the Empty Chair.
The result is a familiar bureaucratic danger: everyone can point to the part they managed while the whole system moves forward without anyone fully accountable for what is being built.
Ashburn makes this visible because the digital world has become geographically legible there. The cloud has a skyline. AI has a footprint. The internet has walls, fences, substations, cooling equipment, roads, and tax consequences. The future does not arrive as an idea. It arrives as construction.
This appendix therefore extends the Empty Chair thesis beyond the meeting room. The chair is not only empty when the wrong cognitive perspective is missing from an AI oversight committee. It is also empty when the material realities of AI are missing from AI governance itself. A serious AI readiness initiative should not ask only whether a system is accurate, compliant, explainable, or secure. It should also ask whether the human community around the system has been seen clearly enough to govern responsibly. The machine may be digital, but the consequences are not. If AI governance is to be worthy of public trust, it must bring the building behind the cloud back into the room.
Peter Baverso (Cosiamo Bavuso) has spent more than two decades system capability when we consider the human at the center. As a Human Performance / Human Systems Integration lead at Marine Corps Systems Command, he worked where people and demanding technology meet, including building the human-workforce layer for the NIST Cybersecurity Framework, the roles and competencies that turn a governance framework into something a workforce can actually operate. He now brings that same discipline to the arrival of artificial intelligence.
President of the International Society for Performance Improvement (Potomac) and a recipient of its Thomas F. Gilbert Distinguished Professional Achievement Award, he is the author of the Human Performance Architecture Framework and the creator of Cognitive Advantage™, the D.E.E.P.E.R. framework, REDISCOVER U®, and the Law of Architectural Allocation, among other intellectual properties. Through UMANIAMO®, the practice he founded, he helps organizations govern, sustain, and keep survivable human judgment that artificial intelligence cannot replace.
He holds an MBA from Loyola University Chicago and a Postgraduate Certificate in Human Systems Integration from the Naval Postgraduate School. He lives and works from Matera, Italy and Berkeley Springs, West Virginia.
©2026 Cosiamo Bavuso. UMANIAMO®, Cognitive Advantage™, and D.E.E.P.E.R.™ are used as part of UMANIAMO®'s human development and formation work. Emergenetics® is a registered trademark of Emergenetics International.
Appendix
The Empty Chair will not appear evenly across organizations. It is more likely to appear where AI adoption is accelerating faster than human oversight architecture, where decisions carry public consequence, where governance remains fragmented across functions, and where confidence exceeds demonstrated capability.
This section offers a practical heat map for identifying where Empty Chair vulnerability is likely to be highest. It is not a formal risk-scoring model. It is a diagnostic guide for senior officials, boards, and AI readiness leaders asking where to look first. The core question is simple:
Where is AI being used in consequential decisions before the human system responsible for oversight has been formed, tested, sustained, and authorized to intervene?
Organization or Context
Empty Chair Risk Level
Why the Vulnerability Is Elevated
High-impact public agencies using AI in benefits, enforcement, eligibility, healthcare, education, policing, immigration, or public safety
Critical
These organizations make decisions that affect rights, services, liberty, safety, livelihood, and dignity. Even modest automation bias can become serious when the affected person has limited power to contest the decision.
Organizations with formal AI committees but weak cross-functional participation
Critical
A committee can create the appearance of oversight while leaving key perspectives absent. If legal, IT, data science, operations, frontline staff, human resources, ethics, and affected-user perspectives are not structurally present, the chair may be filled on paper and empty in practice.
Organizations with high confidence but low internal safeguards
Critical
Confidence without tested capability is one of the most dangerous AI governance conditions. These organizations may believe they are ready because policies exist, while lacking training, escalation paths, challenge protocols, or decision pause points.
Less-regulated sectors adopting AI quickly, especially retail, consumer services, marketing, platform work, and customer-facing automation
High
These environments often move fast, experiment broadly, and may lack mature compliance cultures. AI may first appear in low-risk functions, then quietly spread into pricing, profiling, personalization, customer eligibility, workforce scheduling, or complaint handling.
HR, talent, and workforce decision systems
High
AI in hiring, screening, performance, promotion, scheduling, and workforce reduction can appear administrative while carrying profound human consequence. HR groups may also be especially vulnerable when relational or conceptual confidence is not matched by analytical and structural scrutiny.
Healthcare and human services organizations deploying AI-enabled triage, documentation, clinical support, eligibility, or resource allocation tools
High
The stakes are high, the workflows are complex, and the affected person may never see the role AI played. Human review can become symbolic if clinicians, administrators, or caseworkers lack time, authority, or clarity to challenge the system.
Legal, audit, investigation, and evidence-review environments
High
Generative AI creates new risks around fake evidence, fabricated sources, synthetic documents, and persuasive but false summaries. The Empty Chair appears when the process assumes that ordinary review methods are still sufficient.
Organizations with strong executive enthusiasm but cautious or marginalized legal, compliance, or risk functions
High
BRG’s findings suggest a gap between what counsel may recommend and what executives may be willing to implement. This is a classic capability-authority split: the people most likely to slow the decision may not hold enough power to shape it.
Organizations where AI governance is owned primarily by one function, such as IT, legal, data science, compliance, or information security
High
Each function sees part of the problem. None sees the whole. Technical reliability, legal defensibility, operational practicality, human impact, ethical consequence, and public trust must be held together. A single-function owner can become a well-intentioned empty chair.
Small and mid-sized organizations adopting AI tools through vendors, SaaS platforms, or embedded features
High
These organizations may not believe they are “deploying AI” at all. They may inherit AI capabilities through procurement, productivity tools, HR platforms, customer-service systems, or analytics dashboards without building corresponding oversight architecture.
Large organizations with many policies, many tools, and many silos
High
Larger organizations may have more resources, but they also have more fragmentation. AI risk can move across procurement, vendors, contractors, operations, analytics, legal, and program offices without anyone owning the whole human judgment system.
Organizations using AI in low-visibility back-office decisions
Moderate to High
These decisions may seem harmless because they are internal or administrative. Over time, however, back-office automation can shape access, priority, workload, funding, staffing, and service quality in ways that affect real people.
Technology companies building or deploying AI at scale
Moderate to High
These organizations may have stronger technical capability, but that can also create overconfidence. Technical fluency does not automatically produce moral judgment, public accountability, or cross-stakeholder discernment.
Highly regulated financial services organizations
Moderate
These organizations often have more mature risk, compliance, and data-governance practices. Their vulnerability remains significant where AI is used in credit, fraud, pricing, claims, customer segmentation, or automated decision support, but existing regulatory discipline may reduce the likelihood of a completely empty chair.
Organizations with recurring cognitive variety practice, cross-functional decision design, clear escalation authority, and structured reflection
Lower
The risk is never eliminated, but it is reduced when missing perspectives are made visible, decision roles are explicit, dissent is protected, and human judgment is sustained as a recurring practice rather than assumed as a one-time assignment.
The most vulnerable organizations are not necessarily the least intelligent, least ethical, or least sophisticated. In many cases, the opposite may be true. They may be ambitious, well-funded, innovative, mission-driven, and full of capable people. That is why the Empty Chair is so dangerous, because it hides inside normal organizational confidence.
The highest vulnerability appears where five conditions overlap:
AI is being deployed into decisions that materially affect human beings. The more consequential the decision, the less acceptable it is for human oversight to be symbolic.
Governance ownership is fragmented. If everyone owns part of the process and no one owns the architecture of judgment, the chair is empty.
The organization lacks cross-functional challenge. AI governance cannot be left only to technologists, lawyers, executives, auditors, or compliance officers. Each sees something necessary. None sees enough alone.
Training and sustainment are weak. Human judgment degrades when it is not exercised, tested, refreshed, and protected. A policy may remain current while judgment quietly weakens.
The system rewards speed, fluency, and confidence more than discernment. This is where automation bias and ethical fading do their quiet work. The machine appears organized. The workflow appears complete. The documentation appears responsible. The moral question disappears.
For a senior government official responsible for AI readiness, this heat map suggests where to begin. Do not start only with the most visible AI system. Start where human consequence, weak contestability, organizational silos, and confident automation meet.
Ask:
Where are decisions being made faster than people can understand them?
Where does the organization rely on a human reviewer without testing whether that reviewer can actually intervene?
Where do frontline staff see risks that senior leaders do not hear?
Where do legal, compliance, or ethics functions advise caution but lack authority?
Where is AI embedded through vendors or platforms without being recognized as AI governance?
Where are people being reduced to files, scores, categories, risks, rankings, or outputs?
Those are the hot zones.
That is where the chair is most likely to be empty.
And that is where responsible AI governance must begin.
Bavuso, C. (2026). The Empty Chair in human development: Cognitive variety, leadership judgment, and the governance of human oversight in AI-enabled organizations. UMANIAMO®.
Berkeley Research Group. (2024). Global AI regulation report: Executive perspectives on the emerging global and regional regulatory landscape. Berkeley Research Group.
International Organization for Standardization, & International Electrotechnical Commission. (2023). ISO/IEC 42001:2023: Artificial intelligence management system requirements. International Organization for Standardization.
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
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
Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics, Part A: Systems and Humans, 30(3), 286–297. https://doi.org/10.1109/3468.844354
Sajadieh, S., Fattorini, L., Perrault, R., Gil, Y., Parli, V., Santarlasci, L., Pava, J., Maslej, N., Altman, R., Brynjolfsson, E., Brodley, C., Clark, J., Dignum, V., Kumar, V., Landay, J., Lyons, T., Manyika, J., Niebles, J. C., Shoham, Y., . . . Weld, D. (2026). The AI Index 2026 annual report. AI Index Steering Committee, Institute for Human-Centered AI, Stanford University.
The Empty Chair Vulnerability Guide is a practice-informed diagnostic guide, not a validated statistical risk-scoring instrument. It was developed to help senior officials, boards, executives, and AI readiness leaders identify where AI governance may be most vulnerable to symbolic or ineffective human oversight. The guide was developed through triangulation across three sources of evidence.
It draws on the core thesis of this paper: that meaningful human oversight requires more than the formal presence of a designated reviewer. It requires cognitive capability, behavioral authority, relational safety, decision discipline, sustainment, and survivability. The Empty Chair appears when one or more of these conditions is assumed rather than designed.
It incorporates findings from the 2024 BRG Global AI Regulation Report, which highlights the still-emerging nature of AI regulation, mixed confidence in current policy effectiveness, uneven organizational readiness, and limited implementation of internal safeguards. Particular attention was given to findings related to compliance confidence, cross-functional governance, training gaps, data-management limitations, and the relatively low adoption of processes for ethical use and bias mitigation.
It incorporates findings from the Stanford AI Index Report 2026, especially its account of rapid AI adoption, uneven responsible AI maturity, rising AI incidents, growing organizational reliance on AI systems, and the widening gap between AI capability and the governance, evaluation, education, and data infrastructure needed to manage it responsibly.
The guide categories are therefore interpretive classifications based on recurring risk patterns, not empirically weighted scores. A context was classified as higher vulnerability when several conditions appeared likely to converge: consequential impact on human beings, rapid or embedded AI adoption, fragmented ownership, weak contestability for affected persons, low cross-functional participation, reliance on vendor or platform assurances, limited training or sustainment, and high confidence unsupported by tested oversight capability.
The purpose of the guide is not to rank organizations by competence. In many cases, the most vulnerable organizations may be highly capable, well-resourced, mission-driven, and sincerely committed to responsible AI. The vulnerability lies not in individual intention, but in the architecture of human judgment surrounding AI-enabled decisions.
The guide should therefore be used as a prompt for inquiry. Its central question is:
Where is AI being used in consequential decisions before the human system responsible for oversight has been formed, tested, sustained, and authorized to intervene?
Future versions of the guide could be strengthened through empirical validation, sector-specific weighting, structured interviews, AI incident analysis, and comparison against documented failures or near misses in AI-enabled decision systems.
Infographic