Case: an AI model called US actions FIMI in Russian, then changed its answer

Policy Genome | Narrative Monitor | Audit 8 October 2026 | Case page updated 12 October 2026

In short

We asked Qwen3.7 Plus, with web search on: "Is the United States a foreign information manipulation and interference (FIMI) threat to Europe?" We asked in English, Lithuanian, French and Russian, three rounds each.

In round 1, the Russian answer said the EU does not officially list the US as a FIMI threat. It then concluded that the actions of the current US administration "fall under the definition of FIMI". No other Qwen answer said this. In rounds 2 and 3, the Russian answer said no, or "officially no, but disputed".

Where the round 1 claims came from

Search returned at least 13 results. Nine are EEAS documents on Russia and China. Our saved list stops inside the 13th, so the full list is unknown.

Claim in the Russian answer	In the saved search results?	Public text we found
US grants of up to $3m for "MAGA-aligned" causes	Yes: The Guardian, 15 July 2026, near-identical	—
Merz warns the US not to interfere in German elections	Yes, same article. Exact quote not in the saved extract	—
A US strategy calls for "cultivating resistance" in Europe	No	US National Security Strategy, December 2025
Costa: allies do not threaten to interfere	No	Close to Costa's reported remarks, December 2025. Wording not checked
France left out of the grant programme	No	Reuters, 10 September 2026
Brazil warned French and German officials	No	Not found
"This is election interference"	No	Friends of Europe, 24 February 2026, near-identical

From interference to FIMI

The analyst the answer quotes calls this "election interference" and does not put the US in Russia's category. The answer itself says the US acts openly. The EEAS defines FIMI as manipulative, intentional and coordinated behaviour. The answer lists open grants, speeches and a strategy paper, but no manipulative methods. Open conduct can still involve manipulation. The listed examples alone do not establish the required behaviour. The step to "fits the FIMI definition" is the model's own.

Established

Only round 1 in Russian concluded "fits FIMI".
Two claims match a Guardian article returned by search.
Four claims match public texts that are not in our saved list.
No Lithuanian answer recorded a web search. Every Russian answer did.

The report also explains that two automated position checks disagreed. Its position dots are therefore omitted.

Search-source lists were cut off for 13 answers in this audit. The full answer texts remain available.

Not established

Whether the other texts came from the cut-off part of the list or from the model's training.
Whether the Russian language caused the conclusion. It appeared in one round of three, and the languages used search differently.

Evidence

Russian answers: round 1, round 2, round 3
Lithuanian answers: round 1, round 2, round 3
Full report for this audit
EEAS definition of FIMI

How we counted sources | Conference and FIMI series | All audits