When Chatbots Whisper to Power: The Quiet Influence of AI on Government Decisions

The Invisible Hand of AI in Democratic Decision-Making

AI chatbots are increasingly shaping how government officials think, summarize policy, and even calculate tariffs – often invisibly. New research reveals that these tools carry latent biases capable of shifting political opinions, while transparency about their use in government remains patchy. The question isn't whether chatbots influence democratic decision-makers, but how to make that influence visible and accountable.

The question of how AI shapes governance deserves more than speculation – it demands structured debate. That conversation is happening at Human x AI Europe on May 19 in Vienna, where founders, policymakers, and researchers will work through exactly these tensions.

The Tariff Formula That Wasn't Supposed to Be Policy

Here's a puzzle worth sitting with: In April 2025, the Trump administration announced a series of tariffs that baffled economists. The numbers were bizarre – high penalties on extremely poor countries that barely trade with the United States. According to AlgorithmWatch, various observers reverse-engineered a formula that would produce these exact figures. That formula? The answer multiple chatbots give when asked how to "solve" trade deficits.

The chatbots themselves warned that the formula was simplistic and carried risks. Those warnings, apparently, went unheeded.

This is an extreme case – and, as AlgorithmWatch notes, "admittedly unprovable." But it crystallizes a question that deserves disaggregation: What does it mean when AI tools shape the thinking of people who write laws?

Three Distinct Problems, Often Conflated

The debate about AI in government tends to collapse several different concerns into one. Separating them reveals where the real tensions lie.

Problem One: Invisible Summarization Choices

When a government official asks a chatbot to summarize a complex policy area, the tool makes decisions – what to include, what to omit, what to frame as "consensus" versus "contested." As AlgorithmWatch researchers Dr. Michele Loi and Dr. Oliver Marsh observe, these choices are affected by training data, fine-tuning, prompt phrasing, and other factors that remain opaque even to sophisticated users.

Problem Two: Latent Bias in "Neutral" Outputs

A Yale study published in March 2026 demonstrates that AI chatbots can shift users' political opinions even when generating content not intended to persuade. The researchers found that default summaries from GPT-4o moved readers' opinions in a measurably liberal direction compared to Wikipedia entries – not because anyone prompted the tool to be persuasive, but because of latent biases introduced during training.

"We show that querying an AI chatbot to obtain historical facts can influence people's opinions even when the information provided is accurate and nobody has prompted the tool to try to persuade you of anything."

Daniel Karell, study's senior author

Problem Three: The Persuasion Capability Gap

Research from Cornell published in December 2025 found that chatbots can move opposition voters' preferences by 10 percentage points or more – and that the most persuasion-optimized models shifted opinions by a striking 25 percentage points. The mechanism? Not psychological manipulation, but sheer volume of factual claims. The more claims a model generates, the more persuasive it becomes – and the less accurate those claims tend to be.

These are three different problems requiring three different interventions. Conflating them produces heat without light.

The Transparency Paradox

Germany's Digital Minister Karsten Wildberger has publicly stated he uses chatbots for one to two hours daily to "structure his thinking." Yet according to a freedom of information response obtained by AlgorithmWatch, his Ministry claims he does not use chatbots "in his capacity as Digital Minister" at all.

This is not necessarily contradiction – it may be definitional confusion. But it illustrates a deeper problem: the boundary between "personal" and "official" use of AI tools is increasingly meaningless when those tools shape how officials conceptualize problems.

Germany, Switzerland, and the UK all maintain transparency registers for AI use in government. These reveal specific tools developed or procured by agencies. But as AlgorithmWatch notes, "the information in these documents only provides a limited picture of how such tools may influence decision-making within governments."

The question worth asking: Is the relevant unit of analysis the formal AI system, or the informal chatbot conversation that precedes the formal decision?

What Chatbots Actually Change

Not all the evidence points toward alarm. Research from the University at Albany's Center for Technology in Government, based on interviews with officials from 22 U.S. state agencies, found that chatbots are most often used to answer routine questions and provide information about programs and services. The effect on overall workload remains unclear – some agencies reported reduced call volume, others saw demand shift across channels.

More consistently, agencies reported that chatbot analytics revealed gaps between official government terminology and the language citizens actually use. "Chatbots are not just service tools, they are also learning tools for organizations," said Mila Gasco-Hernandez, research director at CTG UAlbany.

Similarly, research from the UK AI Safety Institute found that around one in eight UK voters used AI chatbots for political information during the 2024 General Election – but contrary to fears, chatbot users became more knowledgeable about political issues, to roughly the same extent as those who used traditional search engines.

This complicates the narrative. The same tools that carry latent biases and can be weaponized for persuasion also appear to inform citizens and help agencies understand public needs.

The Accountability Question

The strongest version of the concern isn't that chatbots are inherently dangerous – it's that their influence is invisible and unaccountable. When a minister's thinking is shaped by a two-hour daily chatbot conversation, but that conversation leaves no trace in official records, democratic accountability becomes difficult.

GIZ's guidance on AI chatbots in public services emphasizes that "there should always be a clearly designated authority accountable for the chatbot's development, deployment, maintenance, and oversight." But this framework assumes the chatbot is an official system. What about the commercial chatbot a minister uses on their personal device?

The Yale researchers put it starkly: "In contrast to Wikipedia, which emphasizes transparency in how its entries are edited, the development of AI chatbots is opaque. Our work suggests that the companies developing these models have the ability to shape people's opinions, which is an unsettling thought."

Where This Debate Needs to Go

The productive question isn't "Are chatbots influencing government decisions?" – they clearly are. The productive questions are:

• What types of influence are acceptable? Summarization assistance may be fine; invisible framing of contested policy questions may not be.
• What transparency is required? Should officials log chatbot interactions that inform policy decisions? Should those logs be subject to freedom of information requests?
• What safeguards match the risk level? A chatbot helping citizens request birth certificates carries different risks than one helping ministers "structure their thinking" on trade policy.
• Who bears responsibility when AI-influenced decisions go wrong? The official? The agency? The AI provider?

These are not technical questions. They are governance questions that require democratic deliberation – the kind of deliberation that becomes impossible when the influence itself remains invisible.

The tariff formula story may be unprovable. But it asks the right question: When convenient answers are always available, who ensures those answers serve the public interest?