AI search said ‘not listed,’ the sanctions list said otherwise: 17 wrong answers in 96 company checks

• Who: SanctionsKit LLC of Denver, Colorado — sanctions screening software; research author Alex Smith.
• What: a study comparing AI-generated search summaries against an official U.S. sanctions list.
• When: searches run October 8, 2026 — one search per company, one search engine, one official list; findings announced October 20, 2026.
• The misses: 17 of 96 summaries stated the wrong sanctions status; 3 said a company was absent from the list when official records showed it was listed.
• The examples: one summary said “no public listing” for a company whose exact name and country appeared in the official file that day; another said a company was still listed while citing the official notice that recorded its removal — researchers checked the current list and the company was absent.

Sanctions screening is one of those jobs where a confident wrong answer is worse than no answer at all. Today, Denver-based SanctionsKit published a study showing how often AI search gives the confident wrong answer: across 96 company searches checked against an official U.S. sanctions list, 17 AI-generated summaries stated the wrong sanctions status — including three that declared a company absent from the list when official records said it was listed.

The announcement, published this morning via GlobeNewswire, lands as compliance teams lean harder on AI summaries for quick answers. SanctionsKit's point is blunt: a search summary is not a screening process, and treating it as one creates exactly the failures the study measured.

The test: 96 companies, one search engine, one official list

SanctionsKit ran one search per company on October 8, 2026, comparing a single search engine's AI-generated summary against one official U.S. sanctions list. The design is deliberately narrow — one engine, one list, one check per company — which makes it a clean measurement of a common habit: typing a company name into search and reading the AI summary instead of checking the source record.

Wrong in both directions

The 17 misses split into the two failure modes compliance teams fear most. Three summaries told the searcher a company was not on the list when the official records said it was. In one example, the summary said there was “no public listing” for a company whose exact name and country appeared in the official file checked that day — and attributed its conclusion to official records that, in fact, showed the company was listed. The remaining errors went the other way: summaries insisting companies were still listed after they had been removed.

A government citation didn't save the wrong answer

One of the more telling cases arrived with its own receipts — and was still wrong. An AI summary said a company was still listed while citing the official notice that recorded its removal. SanctionsKit's researchers followed the citation, checked the current list, and found the company absent. A government citation inside the answer doesn't make the conclusion true; it just makes the error harder to catch.

The risk is in the workflow

SanctionsKit frames the stakes in workflow terms, not model terms. An incorrect “not listed” answer can lead an employee to skip further review of a company that is actually listed. An incorrect “listed” answer can delay customer onboarding or disrupt a supplier relationship for no reason. And an uncertain answer can be mistaken for “no match.” These are the failure modes that appear when employees rely on quick web searches without a defined sanctions screening process. As research author Alex Smith put it: “Employees need a clear process that checks official records, sends possible matches for review, and keeps the evidence behind each decision.”

The study doesn't claim an overall error rate — the 96 companies were selected, not sampled at random. The finding is narrower and harder to dismiss: a quick AI summary is not a screening process. When the answer matters, check the official record, route possible matches for review, and keep the evidence. A summary that cites the source and gets it wrong is still wrong.