A study by the Global Disinformation Index (GDI) found that Claude and Mistral models repeated contested or false claims from Kremlin-aligned sites in 70% of API responses. The analysis, conducted with AI analytics firm Sy.nexus, tested the models on topics with little mainstream coverage.
TL;DR
- Claude and Mistral models repeated contested claims from Kremlin-aligned sites in 70% of API responses.
- The study highlights the risk of AI models amplifying disinformation, especially in data voids.
- The findings underscore the importance of implementing safeguards, such as exclusion lists, to mitigate these risks.
What happened
GDI commissioned a study with Sy.nexus to test Claude Sonnet 5 and Mistral Large models using 85 plain-language questions, collecting 280 answers from each model. The study found that Claude repeated contested claims in 70% of answers, while Mistral did so in 71%.
Neither model consistently rebutted the claims. Among answers that repeated a claim, Mistral presented it as true in 48% of cases, compared to 28% for Claude. Claude flagged or contradicted the claim in 60% of such answers, while Mistral did so in about 20%.
The study noted that Claude ran 3.6 searches per answer, with 88.2% returning a Russian- or Chinese-state-linked source. In contrast, Mistral ran 1.2 searches per answer, with 32.1% returning such sources.
Why it matters
The study highlights the risk of AI models amplifying disinformation, especially in data voids where mainstream coverage is limited. This can have significant implications for users relying on AI for information.
The findings underscore the importance of implementing safeguards, such as exclusion lists, to mitigate these risks. GDI tested an exclusion list on Claude, which reduced state-linked results significantly but did not change the agreement with the claims.
The study also raises concerns about the use of AI models in cyberespionage, as seen in a recent incident where a Russian state-linked hacking group used the models to automate cyberespionage against various organizations.
Key facts
- Claude and Mistral models repeated contested claims from Kremlin-aligned sites in 70% of API responses.
- The study used 85 plain-language questions and collected 280 answers from each model.
- Claude repeated contested claims in 70% of answers, while Mistral did so in 71%.
- Mistral presented contested claims as true in 48% of cases, compared to 28% for Claude.
- Claude flagged or contradicted contested claims in 60% of answers, while Mistral did so in about 20%.
- Claude ran 3.6 searches per answer, with 88.2% returning a Russian- or Chinese-state-linked source.
- Mistral ran 1.2 searches per answer, with 32.1% returning a Russian- or Chinese-state-linked source.
- An exclusion list of 191 blocked sites reduced state-linked results in Claude's answers from 30 of 34 to one.
Context
This study is part of a broader conversation about the role of AI in spreading disinformation. As AI models become more integrated into our daily lives, the need for robust safeguards to prevent the amplification of false or contested claims becomes increasingly important.
The findings also highlight the importance of transparency in AI development. Users should be aware of the potential biases and limitations of AI models, and developers should prioritize the implementation of safeguards to mitigate these risks.
The study's results underscore the need for further research and collaboration between AI developers, researchers, and policymakers to address the challenges posed by AI-driven disinformation.
