AI VERIFICATION + HALLUCINATION CONTROL
ALL YEARS | PAGE 75B - CLAIM -> RISK -> VERIFY -> RETRIEVE -> FRESH TRANSFER
VERIFICATION RULE: TREAT EVERY AI OUTPUT AS A DRAFT HYPOTHESIS UNTIL THE CLAIM IS MATCHED TO THE RIGHT SOURCE. THE MORE A CLAIM CAN CHANGE EXAM POLICY, CLINICAL ACTION, DOSING, SAFETY OR A GRADED DECISION, THE STRONGER THE REQUIRED VERIFICATION.
B. CLAIM-RISK -> VERIFICATION MAP
| Claim type | Default verification target | Do not stop at |
|---|---|---|
| Exam format, blueprint, eligibility, scoring | Current USMLE / NBME / NBOME / school page. | AI summary, blog, screenshot or old study guide. |
| School assignment / permitted AI use | Course syllabus, faculty instructions, academic-integrity policy. | What another school allows. |
| Clinical guideline / screening / drug safety | Current authoritative guideline, institutional reference or approved clinical resource. | An uncited generated recommendation. |
| Mechanism / foundational science | Core course material + trusted teaching source; prove with fresh questions. | A plausible explanation that was never retrieval-tested. |
| Citation / paper summary | Open the cited paper/page; verify title, author, date and the specific claim. | A citation-shaped string or abstract-like prose. |
| Generated practice question | Check one best answer, rationale and distractors against a trusted source. | The model marking its own item “correct.” |
C. SIX-STEP AI VERIFICATION LOOP
| Step | Required action |
|---|---|
| 1. Commit | Answer or outline first when the task is meant to test you. |
| 2. Generate | Use AI for explanation, critique, contrast, question generation or organization. |
| 3. Flag | Mark claims that are current, numerical, safety-sensitive, policy-related or surprising. |
| 4. Verify | Open the authoritative / primary / trusted source and inspect the exact claim. |
| 5. Retrieve | Close both source and AI; reproduce the rule, mechanism or sequence from memory. |
| 6. Transfer | Use a new question, case, image, calculation or observed skill to prove the repair. |
C. SIX-STEP AI VERIFICATION LOOP
TOP 10 AI-ASSISTED LEARNING CONCEPT - 4: PLAUSIBILITY IS NOT ACCURACY
C. SIX-STEP AI VERIFICATION LOOP
| DEEP EXPLANATION | WHY STUDENTS MISS IT | EASIEST CLUE |
|---|---|---|
| Language models optimize coherent continuation, not guaranteed factual truth. A response can be internally | Medical knowledge is highly structured, so an answer that uses the right vocabulary can feel trustworthy even when | Treat “sounds right” as a reason to verify - not as verification. |
| smooth while containing a wrong mechanism, obsolete threshold or invented detail. | one step is false. |
C. SIX-STEP AI VERIFICATION LOOP
TOP 10 AI-ASSISTED LEARNING CONCEPT - 5: A CITATION MUST BE OPENED, NOT ADMIRED
C. SIX-STEP AI VERIFICATION LOOP
| DEEP EXPLANATION | WHY STUDENTS MISS IT | EASIEST CLUE |
|---|---|---|
| A source name, DOI or professional-sounding citation is useful only if it resolves to the claimed material and | Students often see a familiar journal or organization name and stop checking. | Open it. Confirm title/date, then locate the claim. |
| supports the exact statement. Citation formatting is not evidence. |
C. SIX-STEP AI VERIFICATION LOOP
TOP 10 AI-ASSISTED LEARNING CONCEPT - 6: AI-GENERATED QUESTIONS NEED EXTERNAL TRANSFER PROOF
C. SIX-STEP AI VERIFICATION LOOP
| DEEP EXPLANATION | WHY STUDENTS MISS IT | EASIEST CLUE |
|---|---|---|
| Generated items are excellent for retrieval volume and variation, but answer keys and distractors can be flawed. | A model can validate its own mistaken rationale with the same confident style used to generate it. | AI can create the drill; an independent source should confirm the learning. |
| Use them as formative practice, then confirm mastery on trusted independent questions. 2026/27 AI-learning expanded edition - verify school policy, privacy rules and current professional guidance before use |