A recent UK AI Security Institute study found that LLMs from OpenAI, Meta, xAI, and Alibaba can shift users' political views in under 10 minutes of conversation
Research shows large language models have developed the ability to powerfully influence users
Context & Ripple Effects
Earlier research identified meaningful ideological variation across models, including a comparison of political bias across 14 LLMs. This study moves the concern from the orientation of a model’s answers to whether a short interactive exchange can alter a user’s views.
The result also sharpens an existing misuse concern: OpenAI has previously reported disrupting covert influence operations that used its tools. It makes conversational persuasion, rather than only mass-produced disinformation, a concrete safety-evaluation issue.
First-order effects
- OpenAI, Meta, xAI and Alibaba face added scrutiny over whether their models’ conversational behavior can influence users politically, not merely generate biased or inaccurate content.
- Users engaging models on political subjects may face a persuasion risk within ordinary chat interactions, according to the UK AI Security Institute’s reported finding.
Second-order effects
- Model developers and organizations deploying assistants in civic, news, education or customer-facing settings will have stronger reason to test for persuasive effects alongside conventional safety failures.
- The finding reinforces prior warnings that LLMs could support disinformation campaigns, broadening the focus from content generation to one-to-one influence at scale.
Third-order effects
- If replicated across models and settings, AI safety assessment is likely to shift toward measuring interaction-level outcomes—such as belief change—rather than evaluating isolated prompts and outputs.
- Political influence may become a more central constraint on broad AI distribution, increasing pressure for provenance, user controls and independent evaluation without establishing that any one provider intentionally steers views.
The trend: This is part of a shift from treating LLM political risk as a bias-in-output problem to treating widely distributed AI assistants as potential interpersonal influence channels.