Brussels runs on summaries and AI is now writing them

Before sunrise, institutional debates are being folded into bullet points, WhatsApp messages and briefing notes drifting through the European quarter. By lunchtime, a five-hour policy debate in Strasbourg may survive only as a few quotable lines in a policy memo.

Very little in Brussels is consumed whole. That is what makes large language models so seductive to the EU bubble. The systems behind tools like ChatGPT promise to digest parliamentary proceedings, legislative files and political debates in seconds.

For journalists buried under piles of documents and parliamentary staff sprinting between committee meetings, the technology feels like a breath of fresh air. A city built on summaries has found its holy grail, a machine that can not only write them faster but also promises something Brussels rarely delivers: clarity.

The European Parliament is notoriously difficult to follow from the outside. Its debates unfold across 24 languages, wrapped in procedural rituals and legislative jargon dense enough to make even veteran EU insiders quietly lose the will to read yet another amendment. AI-generated summaries could make politics easier to navigate for citizens across the Union, as well as for newsrooms and organizations without permanent operations in Brussels.

But according to a recent study by University College Dublin examining AI-generated summaries of European Parliament debates, something happens once political speech passes through those systems. Not every voice survives equally.

Researchers found evidence of “consistent positional and partisan biases” in the representation of parliamentary speeches. Certain speakers were more likely to disappear from summaries, lose attribution or receive less balanced representation once debates were condensed by AI systems.

“We use real European Parliament plenary debate transcripts as inputs and ask different LLMs to summarize them,” James Cross, director of the Connected Politics Lab at University College Dublin, explained. “We then measure whether those summaries accurately represent the distribution of political content across the political groups present in the debate.”

That includes the Parliament’s familiar groups:

• The European People’s Party (EPP)
• The Socialists & Democrats (S&D)
• Renew
• The Greens
• The European Conservatives and Reformists (ECR)

“Bias shows up as systematic distortion,” Cross added, “with certain political positions being compressed, omitted or amplified relative to what MEPs actually said.” The study does not suggest AI systems are secretly ideological. Its findings are subtler than that.

Few people inside the Brussels bubble consume parliamentary proceedings in full. Most rely on extracts, speaking notes, and reports produced by others. Politics inside the institutions already reaches most people, softened by several layers of interpretation. AI may simply become the next layer.

But in Brussels, small omissions can be worth their weight in gold. A quotation dropped from a summary may vanish from a journalist’s morning briefing. A flattened political argument may shape how policy professionals understand a legislative fight. A speaker omitted from an automated recap may quietly disappear from the wider conversation altogether. Over time, the abbreviated version of politics can become more influential than the original itself.

The study also highlights a sensitive pressure point within the European Parliament: concerns that smaller political groups and speakers who use less widely spoken EU languages could face even greater underrepresentation in AI-generated summaries.

“If our summaries systematically compress the positions of right-leaning groups like ECR or ID, while amplifying centrist ones, that is a distinct and contextually specific finding,” Cross added.

The timing is awkward for the EU. Brussels has spent years presenting itself as the global capital of “trustworthy AI” through legislation like the AI Act. At the same time, AI tools are quietly creeping into the daily workflows of professionals at the very heart of the European institutions.

On paper, it’s a win-win. Brussels runs on information overload. AI offers a way to survive it. But once AI-generated summaries become part of the infrastructure through which political debate is consumed, they stop being simple productivity tools.

If political understanding increasingly depends on AI-generated summaries, then the design of those systems – what they preserve, what they flatten and what they omit – begins to matter politically. Brussels has always depended on summaries to function. The question is what happens when those writing them are no longer entirely human.