A confident answer is not a source.
Check an AI-generated claim by tracing the reference, reading the relevant passage and matching the evidence to the conclusion. A practical verification example.
Documented failure mode · Edited by PsyBorg · Updated 2026-11-16
Start with the claim you plan to use
An AI answer may arrive with a publication title, a link and a convincing explanation. Before repeating its conclusion, identify the exact claim you need to check. Write it in one sentence, including any conditions that matter to the decision.
NIST’s Generative AI Profile describes confabulation as confidently presented erroneous or false content. It includes fabricated citations among the ways an answer can mislead a reader. The profile also suggests checking generated sources and citations during evaluation and ongoing monitoring. [1]
A reference can exist while the conclusion remains unsupported
Consider a hypothetical example. An AI assistant says that a study proves a workplace tool reduces errors. You locate a real paper with the title it supplied. On reading the methods, you find that the paper tested transcription accuracy on a controlled set of recordings. The generated summary presents that result as evidence about all office work.
There are now two questions to resolve. Did the system identify the right paper? Does that paper support the broader conclusion you were given? Finding the publication answers only the first question.
PSYBORG’s suggested correction is to describe the tested task and conditions accurately, then leave the broader workplace claim unresolved until appropriate evidence is available. This example is illustrative and does not describe a reported incident.
Check the passage and its scope
PSYBORG suggests the following reading method.
- Find the original publication through its publisher, DOI or responsible institution. Compare its author, title, date and version with the generated reference.
- Locate the passage, table or result that is supposed to support the claim. Read the surrounding qualifications and methods.
- Compare what was measured with what the answer claims. Check the task, participants, operating conditions and uncertainty. Record any gap between them.
- Keep a short note linking the claim to the source and relevant passage. If you cannot verify support, mark it unresolved and seek more evidence before using it for an important decision.
Match the check to the consequence
A useful evidence note can contain the proposed claim, the original source, the supporting passage, the limits you found and the decision still to be made. Someone else should be able to follow that trail.
For decisions with serious consequences, involve appropriate expertise and the process responsible for the decision. An AI-generated explanation can help organize questions, while accepting the conclusion requires a deliberate evidence check.
Sources and scope
[1] NIST AI 600-1, Generative Artificial Intelligence Profile, July 2024. See section 2.2, Confabulation, and suggested action MS-2.5-003. Source checked 16 November 2026.
Read source (nvlpubs.nist.gov)
The verification method, evidence-note format and workplace example are PSYBORG editorial analysis. The example is hypothetical. This briefing supplies no estimate of how often confabulation occurs.
Continue reading
How to read an AI system card
Read source (psyborg369.com)
A benchmark score needs a context.
Read source (psyborg369.com)
NIST Generative AI Profile
Read source (psyborg369.com)
READING ROOM ↗