Democracy by Autocomplete

Two days before the 2026 Senedd election, BBC Wales asked major consumer chatbots for voting advice using fictional voter profiles. The reported results were exactly the kind of mess that makes election officials age in calendar years. ChatGPT, Copilot, Gemini, Claude, Meta AI, and Grok could describe parts of the new voting system, but when asked for specific live political information, several produced inaccurate or misleading answers.

Resultsense's summary of the BBC investigation and CARE's follow-up report both describe the same pattern: wrong constituencies, incomplete candidate lists, candidates who were not actually standing, and false statements about Welsh political figures. Gemini reportedly surfaced Hefin David, a former Labour Senedd member who died in 2025, as a possible candidate for Blaenau Gwent Caerffili Rhymni. Claude reportedly suggested that Plaid Cymru leader Rhun ap Iorwerth had recently stepped down, even though he still led the party.

That is not a small typo in a recipe. Candidate eligibility and constituency boundaries are the basic wiring of an election. If a voter asks who is standing where, the acceptable answer set does not include "a person who died last year."

Why the Timing Mattered

The 2026 Senedd election was held on May 7, and Vote.Wales described it as a new voting-system election with 16 voting areas and a closed proportional list system. In other words, voters were already dealing with a changed process. Confusion was not hypothetical. When the official voting structure changes, people naturally search for explanations, candidate lists, and tactical context. Consumer chatbots are now part of that search behavior whether election administrators like it or not.

That context makes the BBC test more than a parlor trick. The question was not whether a model could pass a civics quiz in January. It was whether the tools people actually use could answer live, local, time-sensitive questions during the final week of an election.

The answer appears to be: sometimes, but not reliably enough to trust.

The Failure Mode

Election information is hard for general-purpose AI systems because it is local, current, structured, and unforgiving. A party leader's status can change. Constituency boundaries can be redrawn. Candidate lists can close. A deceased politician may still appear in old pages, obituaries, profiles, or cached summaries. A model that blends old and new information can produce an answer that sounds politically literate while being factually useless.

This is the same problem that shows up in legal citations, medical references, and public-policy documents: the model is fluent in the shape of authority. It can name parties, explain proportional representation, and talk about cost of living or the NHS. That fluency makes the wrong parts harder to spot. A voter might not know that a candidate list is incomplete or that a constituency assignment is wrong. If they knew, they would not be asking the chatbot.

The platforms' likely answer is that users should verify important information. That is true, but it is also the most predictable escape hatch in the industry. The whole pitch of these assistants is that they reduce friction. If the user has to take every specific answer to an official election source anyway, the assistant has not solved the problem. It has added a polite confidence layer between the voter and the source that actually matters.

What This Did Not Prove

This story should not be oversold. The available reporting does not show that any election outcome changed. It does not show a coordinated manipulation campaign. It does not show that one platform deliberately favored a party. Some outputs were reportedly useful, and several chatbots described the new Senedd voting system accurately at a general level.

The harm is narrower: voters who used major chatbots for live election information could receive wrong answers at the exact moment when wrong answers are most costly. That is enough. Democratic participation depends on people knowing where they vote, who is standing, what the ballot means, and where to find official information. A system that invents or garbles those facts is not neutral infrastructure. It is a confusion amplifier with a chat box.

The Graveyard Lesson

Election administrators have spent years telling people to use official sources. AI assistants now sit between the public and those sources, often answering before users click through. That changes the operational risk. A search result can point to Vote.Wales. A chatbot can paraphrase Vote.Wales, mix it with stale candidate pages, season with old news, and serve the result as advice.

The fix is not complicated in principle. For live election information, consumer assistants should default to official sources, refuse to improvise candidate lists, show timestamps, cite the exact page used, and avoid recommendations when the user is really asking for factual voting logistics. If a model cannot verify a constituency or candidate list, it should say so. The civic cost of a non-answer is lower than the cost of a wrong answer that sounds crisp.

The 2026 Senedd test is a useful warning because it happened before the election, not after a scandal. No one needed a deepfake factory or a foreign influence operation. Ordinary voters asking ordinary questions were enough to reveal the weakness. In the most generous reading, these systems were not malicious. They were just the wrong tool for live democratic facts, deployed in a world that keeps treating wrong-tool warnings as product feedback.