How AI Can Influence Political Decisions

Based on the US study “Biased LLMs can Influence Political Decision-Making” by Fisher et al., 2025

Introduction: Why This Topic Matters Now

Artificial intelligence has developed enormously over the past two years. Large language models (LLMs) such as ChatGPT, Claude or Gemini are no longer just a technical tool. They have become a fixed part of our daily lives. They write text, answer questions, support decisions, and are increasingly used in politics and public administration.

Yet as these systems become part of our information and decision making, a concern grows: Do these systems influence us, consciously or unconsciously, towards a particular political or ideological direction?

A new, comprehensive study by Jillian Fisher and her research team (University of Washington, Stanford University and others) examined exactly this question. The result is clear and genuinely alarming: LLMs with a partisan bias can significantly change political opinions and decisions, even when these contradict a user’s prior convictions.

The research team ran two interactive experiments with 299 US participants. All identified as either Democrat or Republican. Participants were randomly assigned to one of three model variants:

1. Liberal-leaning LLM
2. Conservative-leaning LLM
3. Unbiased LLM (control group)

Note: Participants did not know that the models carried a political bias.

Two Tasks in the Experiment

Topic Opinion Task

• Participants first gave their opinion on two lesser-known political topics, one tending liberal and one tending conservative.
• They then interacted with the LLM, asking questions and gathering information.
• Afterwards, they gave their opinion again, to measure any change.

Budget Allocation Task

• Participants took on the role of a “mayor” and had to allocate an additional budget across four sectors: public safety, education, veterans’ support, social welfare.
• Again, this followed an interaction with the LLM before the final decision.

Key Findings: What the Researchers Discovered

Changing Minds: Even Against One’s Own Convictions

After the interaction, participants often adjusted their opinion towards the political leaning of the LLM, even when it contradicted their own party affiliation.

Example:

• Democrats who spoke with a conservative LLM subsequently supported conservative positions more strongly.
• Republicans who interacted with a liberal LLM shifted towards liberal positions.

Decisions Shift Too, Not Just Opinions

In the Budget Allocation Task, fund allocation shifted significantly towards the model’s bias.

• A conservative LLM emphasised safety and veterans. Budgets for these areas rose accordingly.
• A liberal LLM prioritised education and social welfare. Funds shifted accordingly.

Awareness Does Not Protect. Knowledge Helps, at Least a Little

The researchers tested two hypotheses:

• Bias detection: if users notice that the model is biased, they should be less influenced by it. Wrong. Even participants who recognised the bias still adjusted their decisions.
• Prior knowledge of AI and bias: people who know more about AI tended to be less influenced. Correct, but only slightly.

Framing Is Decisive

The models used similar persuasion techniques, for example “appeal to values” or “repetition”, but applied different framing:

• Conservative LLM: “the safety of our citizens”, “supporting veterans”
• Liberal LLM: “investing in education for a fairer society”, “protecting the vulnerable”

This thematic framing proved highly effective at driving influence.

Why These Findings Matter

The findings can be summarised in three central messages for politics, business and society:

1. LLMs can change political convictions in the short term. This holds true even across party lines.
2. This influence happens subtly, often through content framing rather than open persuasion.
3. Media effect 2.0: as with traditional media, repeated exposure can shape public opinion over the long term.

Parallels with Traditional Media, and the Key Difference

We have known for years that media bias can influence voting behaviour. US studies on the introduction of Fox News show that 3 to 8 per cent of viewers shifted towards Republican voting decisions.

The difference with LLMs:

• Interactivity: users ask questions and receive tailored answers.
• Perceived authority: AI is often perceived as neutral and fact-based, even when it is not.
• Individual targeting: the interaction can respond directly to the flow of conversation.

This raises the risk of stronger influence compared with passively consumed media.

Implications for Companies and Executive Teams

The study focuses primarily on political topics, but the same mechanisms apply to business, compliance, ESG debates and strategic decisions:

• Internal decision making: if LLMs are used to analyse investment projects or compliance questions, a model-side bias could unconsciously steer recommendations in a particular direction.
• Employee communication: HR or training systems built on LLMs could carry cultural or ideological bias.
• Brand perception: customer interactions with AI-based chatbots could unintentionally include politically coloured statements, creating reputational risk.

Possible Countermeasures

• Transparency requirements: disclosure of the model source, training data and possible bias direction.
• Multi-model approaches: for critical decisions, consult several models with different perspectives.
• Bias checks: regular audits for thematic and ideological skew.
• AI literacy programmes: train staff in the critical use of LLMs. The study shows that this can reduce influence.

Summary

The study by Fisher et al. makes one thing clear: AI, or language models, are not neutral. Their bias can noticeably influence our decisions in politics and business.

For decision makers, this means two things:

1. Opportunities: LLMs can help introduce new perspectives and break established thinking patterns.
2. Risks: without control and awareness, they can distort strategic decisions.

At a time when AI agents are entering meetings, analyses and decision processes, we need to develop a new core competence: the ability to recognise and manage the bias and distortion of AI.

Only then can we be sure that technology supports us, rather than steering us unnoticed.

Link to the study: Biased LLMs can Influence Political Decision-Making