2.1 Synthetic Competence Is Not Hallucination

Hallucination, described by NIST as confabulation, concerns the production of confidently expressed but erroneous or false content. Synthetic competence concerns the larger operating condition in which polished output causes humans to overestimate readiness or capability. NIST also identifies automation bias, overreliance, anthropomorphism, and inappropriate human-AI configurations as material generative AI risks, reinforcing the point that failure can arise from the relationship among people, models, data, tools, and workflows rather than from model accuracy alone (Autio et al., 2024).

A factually correct answer can produce synthetic competence when the human responsible for it cannot evaluate its limits. Conversely, an imperfect AI response may strengthen genuine competence when a knowledgeable operator examines it critically, detects the weakness, and uses the interaction to improve the final decision.

The distinction is important because improving model accuracy alone cannot eliminate synthetic competence. More accurate and fluent systems may increase it by making the gap between appearance and underlying capability more difficult to observe.

2.2 Synthetic Competence Is an Emerging Construct

Synthetic competence should currently be treated as an emerging executive and interdisciplinary construct rather than a mature scientific category with a standardized measurement instrument. Its foundations are supported by established research on cognitive offloading, automation bias, overreliance, illusions of understanding, skill formation, human-AI integration, and organizational capability.

A closely aligned use of the term has also emerged in higher education research. Jha, Colbran, and Purnell (2026) describe the risk that AI-assisted performance may create the appearance of expertise without the biological memory, judgment, and durable cognitive structures associated with genuine professional capability. Their work underscores the broader significance of the issue beyond enterprise governance. Synthetic competence affects education, accreditation, hiring, professional development, and public trust.

3.1 Artifact Competence

Artifact competence occurs when the visible object appears more mature than its foundations.

The artifact may be a strategy, report, financial model, presentation, workflow, prototype, piece of software, or AI agent. It may contain excellent language and professional design while concealing incomplete evidence, undefined assumptions, untested edge cases, architectural shortcuts, or missing controls.

The central mistake is treating the quality of the artifact as evidence that all necessary transformations have already occurred.

A prototype demonstrates that a prototype can be produced. It does not establish that customers need it, that its architecture can scale, that the economics are attractive, or that the organization can operate it safely.

A strategy document demonstrates that an argument can be assembled. It does not establish that the argument is true, that its assumptions have been examined, or that the organization possesses the capability to execute it.

3.2 Borrowed Human Competence

Borrowed human competence occurs when a person can produce expert-looking work through AI but cannot independently explain, evaluate, adapt, or defend it.

This does not mean that competence must be entirely unassisted. Professionals have always used tools, teams, references, and institutional knowledge. The problem arises when the person’s dependence on the tool becomes invisible and the resulting artifact is treated as evidence of independent judgment.

The distinction becomes critical under pressure. Can the person respond when the model is unavailable? Can they recognize an error? Can they distinguish fact from inference? Can they adapt the answer when conditions change? Can they explain why one recommendation was accepted and another rejected?

If not, the human may possess interface fluency without possessing sufficient domain or decision competence.

3.3 Workflow Competence

Workflow competence concerns the reliability of the entire operating process.

An AI-enabled workflow may succeed on the happy path while failing when customer data are incomplete, another system changes, policy is ambiguous, an exception requires judgment, or several automated actions interact unexpectedly.

The apparent capability of the workflow may therefore depend on carefully controlled conditions that do not represent the operating environment.

Workflow competence must be established through variation. The system should be tested against unfamiliar cases, messy data, contradictory instructions, downstream dependencies, human handoffs, time pressure, and failure recovery.

A workflow is not competent merely because its components are individually capable. Competence emerges from their coordination.

3.4 Institutional Competence

Institutional competence concerns the organization’s ability to sustain, govern, repair, and learn from AI-enabled work.

An institution can possess extensive AI activity without possessing institutional AI capability. It may have many pilots but no promotion criteria, many prototypes but no owners, many agents but no authority model, and many productivity claims but no measurement of downstream rework.

Institutional competence requires more than model access. It requires organizational memory, decision rights, data boundaries, monitoring, auditability, escalation, maintenance, change control, and a process for converting experience into improved practice.

The executive must therefore ask not only whether an AI system works, but whether the organization knows how and why it works, where it should not be trusted, who owns it, and how it will evolve.

4.1 AI Can Raise Immediate Performance

Brynjolfsson, Li, and Raymond (2025) examined 5,172 customer-support agents and found that generative AI assistance increased productivity by approximately 15 percent on average. The largest benefits accrued to less-experienced and lower-skilled workers, suggesting that AI can distribute some patterns associated with stronger performers across a wider workforce.

This is strategically important. AI can raise the performance floor, accelerate access to institutional knowledge, reduce communication barriers, and allow newer workers to handle tasks that previously required greater experience.

The result, however, should not be interpreted as proof that every underlying capability has transferred to the employee. AI may provide access to expert patterns without reproducing the experiences through which experts learn to recognize unusual situations, diagnose failure, or exercise judgment under uncertainty.

4.2 AI Capability Has a Jagged Frontier

Dell’Acqua et al. (2026) studied 758 management consultants and found that AI improved performance on tasks within its capability frontier but could reduce performance on tasks outside it. The frontier was described as jagged because adjacent tasks of apparently similar difficulty could produce very different outcomes under AI assistance.

This creates an executive problem. A successful use case may encourage teams to generalize the tool’s competence beyond the conditions in which it was demonstrated. Users may not recognize when a new task has crossed the frontier because the model’s language remains equally fluent.

The model does not necessarily announce that the nature of the task has changed. Its confidence and presentation may remain stable while the reliability of its answer declines.

4.3 Improved Performance May Not Produce Learning

Bastani et al. (2025) studied nearly 1,000 high-school mathematics students using different forms of GPT assistance. Students performed substantially better while receiving AI support. When that support was removed, however, students who had used an unrestricted general-purpose assistant performed worse than the control group. A more carefully designed tutor mitigated much of the harm, demonstrating that the effect depends on how AI is incorporated into learning.

The OECD Digital Education Outlook 2026 reaches a similar conclusion. Generative AI can support genuine learning when designed around pedagogical principles, but task outsourcing can improve immediate performance without producing corresponding learning gains (OECD, 2026).

The enterprise implication is direct. A junior analyst can produce senior-looking work without developing senior judgment. A programmer can generate sophisticated code without forming a durable understanding of the architecture. A salesperson can produce an expert account plan without acquiring the customer insight required to adapt it.

The organization may be consuming future workforce capability to improve present output.

4.4 AI Can Change Critical Thinking Rather Than Eliminate It

Lee et al. (2025) surveyed 319 knowledge workers and collected 936 examples of generative AI use. Higher confidence in AI was associated with less reported critical thinking, while higher confidence in one’s own abilities was associated with more critical engagement. The study also found that critical thinking shifted toward verification, integration, and stewardship of AI-generated work.

This shift is not inherently negative. In mature AI-enabled work, humans may spend less effort producing the first version and more effort evaluating evidence, comparing alternatives, interpreting context, and governing outcomes.

The danger arises when the human lacks sufficient domain knowledge, confidence, or incentive to perform that oversight. The organization may assign the human responsibility without preserving the competence necessary to exercise it.

4.5 Explanations Can Increase Overreliance

Kim et al. (2025) found that explanations increased user reliance on both correct and incorrect large language model responses. Sources and visible inconsistencies were more effective in reducing reliance on incorrect answers.

This challenges a common assumption that asking a model to explain itself automatically improves trustworthiness. A coherent explanation can make an answer more persuasive without providing independent evidence that the answer is true.

An AI-generated explanation should be treated as another artifact to be examined, not as a transparent account of the model’s internal reasoning.

4.6 Human and Machine Capability Do Not Automatically Combine

Goh et al. (2024) conducted a randomized clinical trial involving 50 physicians. Access to a large language model did not significantly improve physician diagnostic reasoning compared with conventional resources, even though the model alone performed strongly in the study.

The study does not imply that AI has no value in medicine. It demonstrates that placing a capable system beside a capable professional does not automatically create a more capable combined system.

Human-machine competence must be designed. The user must know when to invoke the system, how to frame the problem, how to evaluate the response, when to disagree, and how to integrate the result with context and professional responsibility.

5.1 The Illusion of Explanatory Depth

Rozenblit and Keil (2002) demonstrated that people often believe they understand complex mechanisms more deeply than they actually do. When asked to explain the mechanism in detail, the gaps in their knowledge become apparent.

Generative AI can intensify this illusion by supplying a coherent explanation immediately. The user no longer encounters the blank spaces that would normally reveal the limits of their understanding.

5.2 Access Can Feel Like Internal Knowledge

Fisher, Goddu, and Keil (2015) found that internet search can inflate people’s estimates of their own internal knowledge. Participants can psychologically confuse access to information with information held “in the head.”

Generative AI deepens this effect because it does not merely retrieve a source. It synthesizes the material in the language of the user’s question. Its output may feel less like external reference material and more like an extension of the user’s own reasoning.

5.3 Fluency Conceals Missing Transformation

Generative AI is optimized to produce plausible language. It can provide structure, transition, analogy, argument, and confidence before the human has passed through observation, uncertainty, distinction, or evidence.

The problem is not that the words are poor. The problem is that the words may arrive before the thought has earned them.

Fluency becomes a substitute signal for intellectual development.

5.4 Approval Can Launder Unverified Output

Synthetic competence becomes institutionally dangerous when an AI-generated output receives human approval and is therefore treated as human-verified judgment.

Human presence is not sufficient. The reviewer must possess the expertise, access, independence, method, and authority necessary to challenge the work. Otherwise, the approval layer may simply convert machine-generated plausibility into apparently governed institutional action.

7.1 Aśvattha: The Room Before Completion

Aśvattha is the room in which an unfinished observation is permitted to remain unfinished.

Nothing has to sound impressive there. Nothing has to be marketable. Nothing has to arrive with a category, solution, or thesis already attached.

For an individual, Aśvattha may be a journal, voice note, walk, or private reflection. For a team, it may be a field-observation session in which members describe what they encountered before proposing solutions. For an executive group, it may be a protected discussion in which uncertainty and contradiction are surfaced before strategy language takes over.

Aśvattha preserves the human’s opportunity to encounter the problem before the machine supplies the answer.

7.2 The Nine Kriyas

The movement can be summarized as:

Observation → Question → Tension → Distinction → Thesis → Structure → Form → Trial → Consequence → Origin

7.3 The First Voice Rule

Before AI contributes to a Kriya, the human should produce the first artifact whenever possible.

The human states the first observation.

The human attempts the first question.

The human names the first tension.

The human attempts the first distinction.

The human states the first thesis.

AI may then challenge, compare, research, organize, simulate, criticize, and refine the work.

The sovereignty rule is:

No machine-generated articulation should become the source of a belief the human has not first attempted to express.

This is not a claim that every human idea is superior to machine-generated alternatives. The rule preserves traceability. It allows the person and institution to know where the belief originated, how AI altered it, and which part they are prepared to own.

7.4 Trial as Contact With Reality

Trial is where synthetic competence is most directly exposed.

For writing or strategy, Trial asks whether the argument survives evidence, the strongest objection, factual verification, and the test of authentic ownership.

For products and AI systems, Trial examines five domains:

Problem validation determines whether the problem occurs frequently and has meaningful consequences.

Solution validation determines whether the embodiment improves an actual user outcome.

Economic validation determines who receives value, who has budget, and whether the cost to serve supports a viable model.

Operational validation determines whether data, integrations, exception handling, support, monitoring, and reliability can be sustained.

Governance validation determines what the system may decide, what requires authorization, what information may be accessed or retained, how actions are recorded, and who bears responsibility when the system is wrong.

Internal enthusiasm, executive sponsorship, a polished demonstration, and stated user interest are weak evidence. Repeated use, measurable improvement, willingness to pay, renewal, and expansion are progressively stronger signals.

Trial does not exist to confirm the original thesis. It exists to allow reality to change it.

7.5 Completion as an Earned Decision

Completion does not always mean deployment.

An exploration may be:

• Pursued, because evidence supports a defined commitment.
• Narrowed, because the problem is real but the initial scope is too broad.
• Pivoted, because the observation is valid but the proposed solution is not.
• Parked, because the thesis depends on a future capability or market condition.
• Ended, because the opportunity does not justify further attention.

Ending an idea is not failure when the Wheel has prevented the organization from investing further in an unsupported belief.

Completion returns the consequence of the work to Origin. The organization records what it believed, what evidence changed, what remains uncertain, and what responsibility it is now prepared to accept.

9.1 Hiring

The finished artifact is no longer sufficient evidence of individual competence. Resumes, take-home assignments, written analyses, software prototypes, and presentations can all be heavily AI-assisted.

Banning AI from assessment is not a durable solution because the workplace itself will be AI-assisted.

Organizations should assess three dimensions:

First, candidates should reason through part of the problem before using AI. This reveals their mental models, domain understanding, and ability to structure ambiguity.

Second, candidates should be allowed to use AI while evaluators observe how they frame the task, verify claims, test alternatives, and improve the result.

Third, candidates should defend the finished work. They should explain which recommendations they accepted, what they rejected, which assumptions remain fragile, what evidence they trusted, and what conditions would make the answer wrong.

The objective is not to determine whether the person used AI. It is to determine whether AI amplified judgment or substituted for it.

9.2 Leadership

Executives should resist using AI as a machine for intellectual reassurance.

An executive copilot can summarize documents, identify patterns, generate options, and construct analyses. It may also produce a coherent version of assumptions already embedded in the question.

Leaders should use AI to generate intentional dissent. They should ask for the strongest opposing argument, absent stakeholder perspectives, evidence that would reverse the recommendation, and the assumptions most likely to fail.

AI should deepen Inquiry and Trial rather than accelerate premature Completion.

9.3 Professional Development

Organizations must create deliberate moments of productive struggle inside AI-enabled work.

Employees should explain important recommendations without reading generated text. They should distinguish observed facts from sourced evidence, inference, belief, and speculation. They should conduct pre-mortems, identify edge cases, and revisit selected problems without AI assistance.

The purpose is not to prove that people can outperform the model. It is to ensure that the organization continues producing people who can supervise, challenge, repair, and recover from it.

10.1 Hypothesis Velocity

How many meaningful business hypotheses did AI help the organization test?

The measure should reward learning, not artifact volume.

10.2 Evidence Conversion

How many experiments advanced because evidence supported them? How many were narrowed, pivoted, parked, or ended because evidence did not?

Stopping a weak initiative early is a productive outcome.

10.3 Downstream Rework

How much correction, review, integration, remediation, or recovery is required after the initial AI-assisted output?

A faster first draft can still create a slower total workflow.

10.4 Operational Resilience

How does the capability perform under exceptions, unfamiliar cases, incomplete data, system outages, policy ambiguity, and changing conditions?

10.5 Ownership Coverage

What percentage of production AI capabilities have a named owner, monitoring process, escalation path, maintenance plan, and retirement decision?

10.6 Human Capability Retention

Are employees becoming stronger at diagnosis, explanation, judgment, and exception handling, or only faster at producing acceptable-looking outputs?

10.7 Authority Exposure

Which AI systems can change state, communicate externally, create commitments, or initiate consequential actions? What is the maximum impact of one incorrect action?

10.8 Risk Discovery

How many important risks were discovered during Embodiment and Trial rather than after deployment?

The objective is not maximum AI activity.

It is the conversion of AI-assisted exploration into trusted human and institutional capability.