AI That Predicts Your Votes on Democracy Sites May Distort Results

Imagine an online town hall where thousands of people are asked to vote on hundreds of ideas. Nobody can read them all, so many votes are left blank. Platforms like Polis and Remesh now use AI to fill in those gaps — a process called preference inference. The AI guesses how you would have voted based on your other choices.

But here's the problem: these guesses aren't neutral. A new study from a team of French researchers shows that AI-filled votes can artificially boost certain viewpoints, bury others, or change which ideas look most popular. The team tested several AI models on a huge dataset of over 90,000 participants, 1 million votes, and 22 languages — the largest of its kind. They found that two models with nearly identical accuracy scores could produce very different collective pictures.

The researchers argue that accuracy alone isn't enough when the stakes are democratic decisions. A model that predicts individual votes correctly 80% of the time might still distort the group's overall preferences. So they created a new way to evaluate these AI systems, focusing on whether the whole landscape of opinions stays intact — which ideas have broad support, which ones split people, and which represent important minorities.

Why does this matter for you? These platforms are already being used in real civic decisions, from city budgeting to public policy. If the underlying AI quietly shifts the apparent will of the people, the debate might look more united or more divided than it really is. This research is a first step toward making that AI more accountable, so that scaled-up democracy doesn't come at the cost of fair representation.