Deep expertise is valuable. But in a fast-changing AI environment, the expert who cannot question a familiar method, translate across fields, build a usable proof, or revise a conclusion can become trapped by the very knowledge that once created advantage.
This practical audit helps a professional, technologist, graduate learner, founder, or team test for adaptive expertise: the ability to preserve rigor while learning, transferring, creating, and acting when the problem changes.
Why run the audit now?
The 2026 Stanford AI Index documents rapid capability gains but also “jagged” performance: frontier systems can excel at difficult benchmarks while failing at simpler tasks. In science, leading agents still perform far below Ph.D. experts on some end-to-end research evaluations. AI is therefore neither a toy nor an all-purpose replacement for expertise.
At the same time, OECD research finds that AI makes complementary human capabilities—critical thinking, creativity, collaboration, problem solving, data interpretation, and continuous learning—more important. Research on adaptive expertise distinguishes efficient performance in familiar conditions from the ability to adapt expertise when tasks and methods change.
The practical question is not “Are you smart?” It is: Can your expertise survive contact with a new problem?
The seven-test adaptive expert audit
For each test, collect one concrete artifact. Mark the result proved, partly proved, or not yet proved. Do not award points based on confidence, title, reputation, or years of experience.
1. The plain-language test
Task: Explain one important concept from your field to an intelligent teenager or a professional from another field in 200 words. Remove acronyms and prestige language. Include one example, one limitation, and one decision the concept improves.
Proof: Ask the reader to explain it back and name the decision. Record what remained unclear.
What it tests: Whether you understand the idea deeply enough to translate it without hiding inside vocabulary.
2. The disconfirming-evidence test
Task: Write your current conclusion and the three strongest facts that would cause you to revise it. Find one credible source or observation that challenges your preferred view.
Proof: Preserve a short “before / evidence / after” memo. A valid result may be “no change,” but only if the contradictory evidence was fairly represented.
What it tests: Whether expertise is being used to discover truth or defend identity.
3. The cross-domain transfer test
Task: Invite one practitioner from a different field to examine the problem. Ask: What does your field notice that mine ignores? Which assumption would you test first? What failure pattern looks familiar?
Proof: Produce a one-page map with at least one useful transfer, one false analogy rejected, and one new question.
What it tests: Whether breadth improves the work without replacing depth.
4. The build-and-receive test
Task: Convert analysis into the smallest safe artifact another person can actually use: a prototype, workflow, lesson, test case, checklist, model, demonstration, or service step.
Proof: Put it in front of a real intended user. Record what they could complete, where they stopped, and what changed in the next version.
What it tests: Whether the work creates receiving-surface value rather than only impressive output.
5. The calibrated-AI test
Task: Give an AI system a sanitized problem statement and ask it to identify assumptions, missing stakeholders, counterexamples, and evidence needed. Then independently verify every consequential claim with primary sources, direct tests, or qualified human review.
Safe prompt:
I am reviewing my own authorized, sanitized work. Challenge my assumptions. Identify missing perspectives, failure modes, evidence needed, and questions for a qualified human. Do not ask for secrets or private records. Separate supported observations, inferences, proposals, and unknowns. Do not make the final decision.
Proof: Keep an AI-use note showing what the system suggested, what a human verified, what was rejected, and what remains unknown.
What it tests: Whether you can use AI as a challenger and accelerator without surrendering judgment.
6. The novice-feedback test
Task: Ask a novice to use your explanation or artifact without live rescue. Observe silently until a safety, privacy, or dignity stop point is reached.
Proof: Record the first three moments of confusion and change the artifact—not the novice—to make the path clearer.
What it tests: Whether your experience has created a “curse of knowledge” that hides missing steps.
7. The accountability test
Task: Name the human owner, affected people, evidence, decision threshold, risk, rollback point, and verification method before implementation.
Proof: Complete an evidence card:
| Field | Required record |
|---|---|
| Outcome | What human result is intended, for whom, and by when? |
| Authority | Who may decide, access, change, publish, or stop the work? |
| Evidence | What is observed, reported, inferred, proposed, or unknown? |
| Risk | What could harm privacy, dignity, money, rights, safety, time, or trust? |
| Stop point | What condition requires human review or rollback? |
| Verification | How will the result be checked at the receiving surface? |
What it tests: Whether intelligence is connected to responsibility.
How to interpret the result
- Proved: A dated artifact and receiving-surface observation demonstrate the behavior.
- Partly proved: The behavior occurred, but evidence, external feedback, or verification is incomplete.
- Not yet proved: The result rests on confidence, title, intention, or an untested draft.
Do not combine the seven results into a universal “intelligence score.” The tests measure different behaviors, and their importance changes by context. A surgeon, cybersecurity engineer, teacher, researcher, and founder face different consequences and must meet different professional standards.
A 90-minute first run
- Minutes 0–10: Choose one real, bounded, non-confidential problem you own or are authorized to address.
- Minutes 10–25: Run the plain-language and disconfirming-evidence tests.
- Minutes 25–40: Run the calibrated-AI test with sanitized information.
- Minutes 40–65: Build the smallest usable artifact.
- Minutes 65–80: Put it in front of one appropriate user or reviewer.
- Minutes 80–90: Complete the accountability card and choose one repair.
Then repeat after the repair. Adaptability is not proved by producing more words. It is proved by changing the work in response to evidence while preserving safety, rigor, and responsibility.
The operating standard
Deep expertise remains essential. A public AI response is not a licensed professional, an authorization, an experiment, a source of legal accountability, or proof that a system works. The goal is not to flatten expert knowledge. The goal is to prevent expertise from becoming a closed room.
Question the frame. Cross one boundary. Build something useful. Test it with a human. Verify the result. Keep responsibility human.
Educational boundary: This audit provides general education and team-learning structure. It is not psychological, employment, academic, legal, medical, engineering, cybersecurity, or other professional advice; it does not authorize access to any system or data. Follow applicable law, contracts, institutional policy, research ethics, professional standards, privacy duties, and qualified review.
Sources
- Stanford HAI — 2026 AI Index, technical performance
- Stanford HAI — 2026 AI Index, science
- OECD — Skills in the AI Age
- Trinh (2019) — expertise, humility, and flexibility
- Bohle Carbonell et al. (2014) — review of adaptive expertise
- U.S. National Science Foundation — convergence research
Sources reviewed October 28, 2026. Louisville Institute of Technology practical education release. Human-led, AI-assisted, source-linked, and responsibility-retained.