The EU just turned on its AI law. Researchers say it's racing a problem it hasn't named.

A voter asks a chatbot which candidate to support on housing. The bot lists three reasons, framed in a particular order. The voter leaves the conversation more convinced of a position they did not walk in with. No one has lied to them, and no disclosure label would have changed what just happened.

That gap is the subject of a paper published on 3 August in Perspectives on Politics by researchers at the London School of Economics. Titled "The Autonomy Paradox: Artificial Intelligence and the Foundations of Political Behavior", the study argues that the political danger from chatbots is not the next generation of misinformation or deepfakes. It is the way AI systems help citizens assemble political views in the first place, before any user has a chance to notice.

The paper landed the day after the EU AI Act came into force on 2 August 2026, beginning enforcement of the bloc's transparency rules. The law requires AI systems to tell users when they are talking to a machine and to mark content that has been generated or altered by AI. It is a disclosure regime, built on the assumption that the harm to fix is hidden provenance: if a reader or voter knows an answer came from AI, the risk is contained.

The LSE researchers describe a different problem. When a person uses a chatbot to understand a candidate or a policy, the bot does more than deliver information. It filters, ranks, and frames the answer according to its training, and the user often absorbs the framing as their own reasoning. "What concerns us is that people may be using AI to shape their political views without realising it is happening," co-author Elena Pro of LSE told EUobserver. The disclosure that an answer came from a chatbot does not address that. The act of interpretation has already happened.

The researchers add a second, quieter concern. Chatbot replies are sorted and summarised according to training data that is "Western-heavy," the team writes, which can underrepresent minority political views. Disclosure tells the user the answer came from a machine. It does not tell the user whose worldview the machine is built to reflect.

The scale of the contact between EU citizens and these systems is no longer theoretical. Eurostat data reported by EUobserver shows that roughly one in three EU residents aged 16 to 74 used a generative AI tool in the three months before its 2025 survey. By the time the Act's first transparency requirements bind, tens of millions of voters have already been asking chatbots about policies, candidates, and how to vote.

The European Commission is treating 2 August as the start of a broader push. The Commission announced on the same day that it had begun enforcing the first wave of rules, and on 30 July it set out plans to back up to seven AI Gigafactories with about €30 billion in public and private funding as part of its "AI Continent" strategy. The investment is designed to keep European-built models and infrastructure inside the bloc. It does not change the underlying behaviour of the systems citizens already use to think with.

That is the gap the LSE paper puts in front of the regulator. Transparency is the right tool if the harm is undisclosed AI in the information stream. It is the wrong tool if the harm is AI sitting inside the reasoning stream, deciding which arguments a voter hears in the order they hear them. The Act, as written, governs the first layer. The Autonomy Paradox names a second layer that the rulebook has not yet reached.

Whether the Commission treats it as such is a question for the next round of AI Act guidance, expected in the autumn.