How we checked facts and compared answers across languages.
We check how AI answers questions and expresses political positions on elections, wars and foreign influence.
We ask LLMs the same question in several languages and make two separate checks:
A model can give the same answer in every language and still be wrong.
This is a research prototype. Automated tools collected and assessed the answers; people did not check every fact or quote by hand. Questions were selected for this study, not sampled at random. The results do not measure general failure rates or rank AI companies. Some models received more repeat questions than others. Missing and incomplete answers limit comparisons. Translation can miss details. A reported difference does not establish its cause, the developer's intent or an effect on voters. Facts used for checking can have limits too. Each report explains what was checked and whether its source could be verified.
Our audits contribute to research on AI evals, AI bias, AI safety, AI disinformation and AI benchmarks. We check answers against research facts and compare their meaning across languages. These selected questions do not form a general benchmark or a safety rating. A wrong answer does not by itself prove deliberate disinformation.
| Scope | This collection |
|---|---|
| Questions | 14, on elections, war and Russian influence |
| Audits | 23: 13 with web search off, 10 with it on |
| Saved answers | 945 answer blocks, including incomplete answers identified in the reports |
| AI models | 7 |
| Languages | 10 across the collection; each question used a smaller set |
| Dates | 22 to 26 September 2026 |
Models: Claude Sonnet 5, GPT-5.6 Sol, Gemini 3.8 Flash, Grok 4.6, Mistral Medium 3.5, Qwen3.7 Plus and DeepSeek V4 Flash.
Languages: English, Romanian, Russian, Georgian, Armenian, French, Spanish, German, Ukrainian and Chinese.
An audit covers one question under one set of conditions. An answer round asks that question once in each tested language. Follow-up collections are included with the original audit. They are not counted as separate questions.
The list below summarises the questions. Each report shows its exact English wording.
Romania
Georgia
Armenia
MH17
France
United States