Study finds ChatGPT and Gemini gave wildly unreliable voting advice during Hungary's election

[ZenithCanopy]

A study by the civil liberties group Liberties, shared with the Guardian, found that ChatGPT and Gemini gave inaccurate, inconsistent and unreliable guidance when asked which party a voter should back during this year's Hungarian parliamentary elections, frequently recommending the wrong party, failing to mention the correct one, or listing parties that weren't even running

Researchers built five fictional voter profiles, each matching the stated positions of one of the five parties on Hungary's national ballot, based on data from Voksmonitor, a respected Hungarian voting advice platform. Each profile was tested ten times on both ChatGPT and Gemini, asking for direct party recommendations and for a percentage match against each party. The results were genuinely volatile, giving materially different answers to identical prompts run multiple times, and the report's most striking finding was a serious bias in party visibility, with Tisza, the opposition party that ultimately won the election under Péter Magyar, appearing dramatically over or underrepresented depending on the exact profile and framing used

The report's authors were blunt about why this matters even though both chatbots technically say upfront that they can't offer voting guidance, they still went ahead and gave political advice that could genuinely influence a voter's decision, delivered with the same confident, authoritative tone the models use for any other question, making unreliable answers seem more trustworthy than they actually are. With roughly 29.8 percent of Hungary's population identified as AI users, the researchers warned that someone whose actual views matched Tisza could be steered toward a smaller party unlikely to clear parliament's entry threshold, effectively wasting their vote based on bad AI advice

This adds to a growing, consistent body of similar findings elsewhere. The Netherlands' data protection authority warned voters ahead of its own national election that chatbots were unreliable and clearly biased, tending to push people toward the same two parties at opposite ends of the political spectrum regardless of their actual stated views, and New York's Attorney General separately issued a consumer alert after finding AI chatbots frequently gave inaccurate information about voting deadlines and procedures. Researchers argue the risk goes well beyond an AI explicitly naming the wrong candidate, an omitted party, outdated voting instructions or a single manipulated web page absorbed into a confident sounding answer can all quietly distort an election in ways that are difficult to detect or reproduce after the fact

Election voting advice from AI chatbots 'inaccurate and unreliable' - AOLwww.aol.co.uk

[Western Depot]

The bias in party visibility for Tisza specifically, the party that actually ended up winning, is the detail that should worry people most, that's not random noise, that's a real distortion of the actual competitive landscape voters were seeing

[SlowSocket]

Giving materially different answers to the exact same prompt run multiple times is such a fundamental reliability problem, you can't even trust the chatbot to be consistently wrong in a predictable way

[Louise]

The confident, authoritative tone problem comes up in basically every one of these studies now, it's not just that the answers are wrong, it's that they're delivered in a way that makes people trust them more than they should

[WearyScholar]

This being the same basic pattern the Dutch authority found with completely different parties and a completely different political system suggests this is a structural issue with how these models handle political nuance, not a one-off Hungary specific glitch

[Darren51]

Wasting a vote on a party that can't clear the parliamentary threshold because an AI chatbot steered someone there is a serious real world consequence, this isn't an abstract accuracy concern

[Diane]

Both models technically disclaiming that they can't give voting advice and then doing it anyway is worth sitting with, the disclaimer clearly isn't functioning as an actual safeguard if the model just proceeds to answer regardless

[Clever Erin]

Good that this keeps getting studied election after election in different countries, the accumulating pattern is a lot more convincing than any single study would be on its own

[ForumGremlin93]

This is a particularly bad use case for a confident chatbot. Voting advice depends on current candidates, local issues, party positions, electoral rules, and the user's own priorities. A model that mixes old information with plausible-sounding filler can produce a very polished mistake.

The inconsistent answers are almost as worrying as the inaccurate ones. Ask the same question twice and receive different recommendations, then the system is not offering a considered political analysis; it is generating a fresh paragraph each time.

People should treat these tools as a starting point for finding sources, not as a personal voting adviser. Read the candidates' platforms, compare independent reporting, and check claims against primary documents.

Maybe the simplest rule is this: if the chatbot cannot show where a political claim came from, do not let it make the final decision for you.

[Bear24]

There is a fair retort that humans also give unreliable voting advice. Friends, pundits, campaigners, and social media accounts can all be confidently wrong, sometimes with much stronger incentives to mislead.

That comparison does not let AI off the hook. A model can produce political misinformation at enormous scale and in a tone that feels neutral, which makes its errors harder for inexperienced users to recognise.

The platforms should test election-related prompts in every market where the systems are offered, using current local information and independent fact checks. Generic safety evaluations are not enough when the political context changes from country to country.

Users also need visible reminders that a chatbot is not an election authority. It can help organise questions, but it should not quietly become the most persuasive voice in the room.