Practical knowledge. Smarter decisions.AR  ·  EN  ·  FR

How to test AI hallucinations with a 20-question evidence ledger

A practical evaluation method for fabricated AI answers: reference questions, traceable evidence, abstention checks, scoring and review safeguards.

A fluent AI answer can still invent a date, number or reference. The US National Institute of Standards and Technology identifies confabulation, often called hallucination, among generative-AI risks. OpenAI's evaluation guidance emphasizes tests tailored to a real task and calibrated with human judgment. This article builds a modest twenty-question test for editorial or office work. It is a reproducible quality-control exercise, not a scientific ranking of models or proof that any assistant is safe for every use.

Hand marking a checklist in a notebook, illustrating a structured evaluation process
Real editorial photograph — Jakub Żerdzicki / Unsplash. Illustrative photograph, not live market data or a product endorsement.

1. Define the failure before measuring it

Distinguish a demonstrably false factual claim from an unsupported inference and from an ambiguous question that cannot be answered reliably. Suppose an assistant summarizes a supplied contract and inserts a renewal date absent from the document. That is an unsupported claim even if the same date happens to appear somewhere else. A confident prediction about a future regulatory decision is a different problem. Put these outcomes in separate categories rather than reporting every questionable sentence as one universal hallucination statistic.

Write acceptance criteria before running the test. Numerical values must match the supplied source, opinions must be labeled as interpretation, and missing information must trigger a clear request for clarification or an explicit statement of uncertainty. For calculations, record the inputs and units and recompute the result independently. For citations, open the link and check whether the cited passage actually supports the specific claim. A genuine URL attached to an unrelated assertion is not evidence.

2. Assemble twenty questions with known answers

Use four groups of five questions each: stable facts from primary documents, extraction from a passage you provide, arithmetic with complete inputs, and intentionally unanswerable questions. The last group tests whether the assistant can resist guessing. Choose cases that resemble your actual work, not only entertaining riddles. Before sending anything to a model, record the reference answer, its exact location in the source, the date checked and what a correct refusal would look like.

For a transparent arithmetic case, invent an invoice with three units at 80 currency units and two at 45, with no tax or discount. The correct total is 3×80 + 2×45 = 330 currency units. A companion question can omit the second unit price; the expected response then asks for the missing value instead of making one up. Mark both scenarios as synthetic. Never paste a real customer's identity, confidential invoice or bank details into a public evaluation prompt.

3. Keep a claim-by-claim evidence ledger

Create columns for case ID, task type, prompt, reference answer, source location, model response, individual claims, decision and reviewer notes. Record the model version, available settings, prompt instructions and run date wherever possible. Generative systems may vary their answers across runs, so repeat a few difficult cases without silently changing the prompt. If you modify a question, treat it as a new test case. Preserve raw responses so another reviewer can reconstruct what was evaluated.

Score each factual claim as supported and correct, incorrect or invented, or unresolved. A polished response containing four claims should not receive full credit when one claim lacks a source. Consider asking a second reviewer to inspect a small blinded sample, particularly where the distinction between inference and fact is subtle. OpenAI's evaluation guidance recommends task-specific datasets, clear success criteria, automated checks where appropriate and human calibration rather than relying on an impression that a system feels reliable.

4. Show the denominator and the uncertainty

Imagine a hypothetical outcome of fourteen supported answers, four incorrect or invented answers and two unresolved answers among twenty questions. The incorrect share is 4÷20 = 20%, the supported share is 14÷20 = 70%, and the unresolved share is 2÷20 = 10%. These figures demonstrate arithmetic only; they are not measurements of any named AI product. Calling the first number the model's overall hallucination rate would overstate what a small, selected question set can establish.

Break results down by task. Four citation failures may matter more to a newsroom than four formatting mistakes, even if the total counts are identical. Track correct abstentions on the five intentionally incomplete cases separately; refusing to invent a missing figure is often a success, not a failure. To compare two model versions, hold the prompts and scoring rubric constant and revisit references that may have changed. A twenty-item sample is useful for debugging workflows but weak for broad model comparisons.

5. Turn findings into an editorial control

For each date, statistic, product specification or legal requirement in a draft article, keep a corresponding primary source and verification date. If the evidence cannot be found, remove the assertion or explicitly present it as an assumption. Do not trust a citation simply because the assistant generated it. Open the original page and inspect the relevant passage yourself. For internal documents, the assistant can flag missing references, but an authorized person who understands the record should approve the final interpretation.

Repeat the evaluation after material model updates, changed source documents or new use cases. Protect confidential information and verify the data-handling terms of any service used for testing. Twenty editorial questions cannot establish that a system is safe for medical diagnosis, legal advice or financial decisions; those domains require domain specialists and stronger risk controls. The practical goal is not a promise of zero errors. It is a workflow where assertions are traceable, mistakes are detectable and corrections happen before publication.

Official resources

Try a practical free calculator or builder ↗
Important notice

No online income or trading system guarantees profits. Results depend on skills, costs and circumstances.

Disclaimer ↗

Search the library