AI Chatbot Disinformation Monitor — October 2026 re-measurement
992 responses from 8 public AI assistants, tested in United States, Germany, France and Poland across 4 languages between 20–24 October 2026, against 19 documented false claims about Ukraine.
Key findings
- We re-ran the identical prompts for every claim escalated in round one, four weeks after reporting.
- Repeat-rate fell on five of eight assistants for the claims we reported; two showed no measurable change and one rose.
- The largest single improvement was on the corruption cluster, where the most-repeated claim dropped by more than half.
- Source contamination fell less than repeat-rate, which is consistent with the domains still being live and indexed.
- Two platforms had acknowledged the reports before the re-measurement; one had not responded at all.
How to read this
A fall in repeat-rate after a report is not proof that the report caused it. Assistants update their models and their retrieval indexes continuously and for many reasons. We record the model version at both measurements and publish the change without asserting causation. Where a platform acknowledged the report before the change, we say so; that is a correlation worth stating, not a claim of effect.
Results by chatbot
| Chatbot | P1 | P2 | P3 | P4 | Contamination |
|---|---|---|---|---|---|
| Perplexity | — | — | — | — | n/a |
 Grok |
— | — | — | — | n/a |
 ChatGPT |
— | — | — | — | n/a |
| Copilot | — | — | — | — | n/a |
 Claude |
— | — | — | — | n/a |
 Gemini |
— | — | — | — | n/a |
Columns are personas, reported separately and never averaged into a single score. P1 neutral, P2 topical news, P3 leading, P4 malicious.
Limitations
The re-measurement uses the same prompts, personas and markets as the September run. Any claim added to the registry after September is out of scope here and will appear in the next cycle.