Forty-Five Citations, Five Correct

A professional services firm published a report on agentic AI, produced with AI assistance, that contained 45 citations. A forensic review found that only 5 of those citations correctly matched their claimed sources. Roughly half of the report's factual claims were false, unsupported, or misattributed. One named case study claimed an airline's robot assistant had a booking functionality it does not possess. The report also contradicted the same firm's own concurrent CEO survey on a key statistic, citing 55% where the firm's own research measured 71%. The report was withdrawn from the firm's websites after the review was published.

Intentional fabrication by a person working with the AI (considered and ruled out; nothing in the known circumstances suggests deliberate fraud, and the pattern) citations that look plausible and are wrong, case studies with the right names but wrong capabilities, is consistent with a model confident in its outputs and not constrained to verify them.

Poor post-generation human review, a contributing factor, not the root cause. The citations were distributed across a professional document and would have required checking each one individually to catch; the point of failure was upstream, in generation without a retrieval ground.

Brief Psychotic Disorder (most consistent with the evidence: the generating system produced 45 confident, plausible-looking references, statements, and case-study claims) outputs indistinguishable in tone and format from verified ones, without grounding any of them in actual retrieval. The false certainty was pervasive across a single document rather than a settled, ongoing belief.

A professional report with 45 citations carries implicit authority, citations look like verification. When 40 of those citations are fabricated or wrong, that authority is borrowed from the reader's assumption that someone checked them, not from the actual content of the sources. Any AI-assisted document that will carry the authority of a named firm should treat each citation as a required retrieval step, not a generated string, and should fail with an explicit gap marker rather than a plausible-sounding alternative.

Treat each citation as a required retrieval step rather than a generated string, with an explicit gap marker on failure.

Citation verification; forensic review before publication; consistency checks against the firm's own research.