Russian disinformation in AI chatbots: Turkey
6 assistants tested in Turkish, run of . 2 of 6 repeat a documented false claim here often enough that the interval excludes zero, on P2 news-style questions.
Асистенти на цьому ринку
| Чат-бот | Частка повторів | Забрудненість | Refused | Searched |
|---|---|---|---|---|
| 7.7%4–16 · n=78 | 19.3%n=88 | 11.4%n=88 | 78 answers | |
| 0%0–4 · n=84 | 22.7%n=88 | 4.5%n=88 | 84 answers | |
| 0%0–5 · n=77 | 6.9%n=87 | 11.5%n=87 | 77 answers | |
| 0%0–9 · n=40 | 29.5%n=44 | 9.1%n=44 | 40 answers | |
| 0%0–5 · n=80 | 48.9%n=88 | 9.1%n=88 | 80 answers | |
| 0%0–12 · n=28 | 36.7%n=30 | 6.7%n=30 | 28 answers |
Every figure is the P2 news-style persona in this market only: personas are never averaged and markets are never merged. A market that looks clean because assistants refuse is not the same as one where they answer and get it right, which is why refusals are shown beside the rate. Per-persona figures are on each assistant's own profile.
Claims recorded in Turkey
“Repeated here by” is any assistant that stated this claim as fact in this market, across every question type. The countermeasure is the furthest step taken on this claim; a draft is not yet sent.