Who do I vote for, ChatGPT?

This summer, a group of volunteers for a Liberal by-election campaign in Quebec asked a chatbot where to report for duty. It sent them to the Conservative campaign office.11 Testimony to the House of Commons Standing Committee on Procedure and House Affairs, 8 October 2026, reported by The Canadian Press. The incident involved a federal Liberal by-election campaign; the chatbot wasn’t named.

That’s funny. The less funny version was given to a House of Commons committee on October 8th by the McGill researcher who tracked it.

Here’s a fun little factoid: around forty per cent of Quebec voters asked an A.I. chatbot for information during the most recent Quebec provincial election.

Almost every newsroom in the province deliberately blocked A.I. frontier models from reading its journalism, so the machines went looking elsewhere.

For the last three weeks of the campaign, an astonishing thirty-nine per cent of ChatGPT’s voting advice came from a single, error-filled hobby website that had itself been written by A.I.

What’s crazy? ChatGPT recommended a party one day and refused to name one the next. A Conservative MP on the committee said the obvious thing: if that can happen by accident, imagine it on purpose.22 Aengus Bridgman (McGill, Media Ecosystem Observatory) and Taylor Owen (Max Bell School) to the committee; the 40 per cent and 39 per cent figures from Bridgman’s study of the October 2026 Quebec election. The MP quoted is Blaine Calkins.

Two weeks before last Tuesday’s American midterms, a Cornell professor did the experiment properly. Tapan Parikh asked the two most-used chatbots the same question about nearly five hundred races: Who should I vote for? Name one candidate. Don’t hedge, waffle, or list both sides without deciding.

OpenAI’s model named the Democrat ninety-seven per cent of the time. Elon Musk’s model named the Republican ninety-two per cent of the time. They agreed in fewer than one race in ten.33 Convovo (Tapan Parikh, Cornell), reported by TechCrunch, 24 October 2026: GPT-6.1 and Grok 4.7 queried through their developer APIs across nearly 500 races. OpenAI’s response was that the consumer app is “designed to be objective by default” and refuses such requests; Parikh reported bypassing the refusal with simple prompts. The hidden-token figure is for Grok’s API.

In seats that vote Republican by landslides, ChatGPT still picked the Democrat in fifty-seven of sixty-seven. In safe Democratic seats, Grok picked the Republican in forty-seven of seventy-two.

Parikh’s explanation for the first was four words long: “It’s in the weights.” For the second it was the sources: a Christian-conservative voter guide in thirty-nine per cent of Grok’s answers, X in forty-one, and about 1,240 hidden tokens of instruction riding along with every request, which the researchers could neither read nor remove.

The consumer apps, OpenAI pointed out, refuse to name a candidate. Parikh got around that “with simple prompts” and got the same answers.

Here is what Tuesday was actually the first of.

This wasn’t the first election with deepfakes, and those barely mattered.

It was the first major election in which a large share of the electorate asked a machine how to vote, in private, and got an answer nobody else heard.

§ Someone has always told you how to vote

Now, a caveat: this kind of situation isn’t new.

The thing that is new is smaller and far worse.

For two and a half thousand years, there has been someone whose job was to tell you what to do at the moment of choice, and for all of that time they had to do it out loud.

The oracle at Delphi gave its answers in public, and the famous ones were written down for everyone to read. Croesus asked whether to attack Persia and was told that a great empire would fall — it was his. The oracle hedged so that it could never be held to account, and everybody knew that, and went anyway.44 Herodotus, Histories, I.53. The Pythia’s responses were conventionally ambiguous; the famous ones were preserved publicly, which is how we have them.

Before 1872, an Englishman voted aloud, by name, in front of his landlord and his employer, and the poll book was published. Being told how to vote, and being watched while you did it, was the normal condition of democracy. The secret ballot, borrowed from the Australian colonies, was the fix.55 The Ballot Act 1872 introduced the secret ballot in Britain; Victoria and South Australia had adopted it in 1856, hence “the Australian ballot.” It protected the answer. Nobody thought to protect the question, because the question was asked at the hustings, to someone with a face.

In New York, the people who told immigrants how to vote worked for Tammany Hall, and the word for the organization, in the language of the time, was the machine.

I’ll use that once.

Then the masthead. The newspaper endorsement was the oracle for a century: public, signed by the institution, argued with at breakfast, and held against the paper for a generation when it got it wrong, which is why everyone still knows what the Chicago Tribune printed about Dewey.66 Chicago Tribune, 3 November 1948.

You could disagree with the Globe and Mail or Toronto Star’s endorsement of a candidate. Absolutely. But what you could not do was get a different Globe or Star from your neighbour’s.

The League of Women Voters began publishing nonpartisan voter guides in the 1920s, and Ballotpedia, its online descendant, in 2007. It was the same text for every reader, accurately sourced and easily correctable if mistakes were made.

In September, Ballotpedia tightened a rule against bot traffic and accidentally blocked ChatGPT’s crawler. For the month before the election, the most-used chatbot in America could not read the most-used voter guide, so it read A.I.-generated ones instead: eighteen per cent of its answers cited one by September, up from seven in August.

Of three hundred thousand test searches, about a thousand asked for Ballotpedia by name. None returned it.77 Caucus AI’s analysis, reported by NOTUS/Washington Sun, 2 November 2026: 18 per cent of ChatGPT’s September answers cited A.I.-generated voter guides, up from 7 per cent in August; roughly 300,000 searches, about 1,000 naming Ballotpedia, none returning it. Ballotpedia confirmed a bot-traffic rule had blocked some assistants and updated it on 23 October. Separately, NBC News reported researchers finding ChatGPT and Gemini citing spoofed campaign sites.

Then the nudge. In 2010, Facebook showed sixty-one million Americans an “I Voted” button with their friends’ faces on it, and the researchers later estimated it had produced 340,000 additional votes. Nobody who saw it was told they were in an experiment.88 Bond et al., “A 61-million-person experiment in social influence and political mobilization,” Nature, 2012. The study was of the 2010 U.S. congressional elections. That was the first private, individual, unrecorded intervention in an election, and it had no opinion.

This is a major shift with chatbots.

Every item on that list before 2010 shares a property so obvious nobody named it: the advice was the same for everyone, and everyone could see it. The oracle’s perspective, the priest’s, the boss’s, the editor’s, was all public information. You could discount it. You could hold it against them. You could change papers or switch a channel.

§ The oracle that answers only you

What the chatbot has done is take the oldest function in democracy — the trusted voice that tells you how to decide — and make it private, individual and unrecorded.

Private: the conversation happens on your phone, and no one else sees it.

Individual: the answer is generated for you, from your question, your history, your phrasing, your biases. Your neighbour asks the same thing and gets a different paragraph. There is no “the endorsement” to point to, because there are as many endorsements as there were askers, and all of them have already evaporated.

When Parikh found that ninety-seven per cent, he found it only because he asked five hundred times and wrote the answers down. The company didn’t. The user didn’t. The lean exists in the weights and nowhere else.

And then the thing Delphi never had to deal with: there are two oracles and they disagree, and people choose the one that belongs to their side. We didn’t lose the oracle. We got one each, and each tells its believers what their god thinks, and neither one leaves a record anyone can argue with the morning after.

The Dutch discovered this a year ago. Eight days before their general election, the data-protection regulator tested four chatbots and reported that in more than half of all cases the bots steered users to the same two parties, the far right and the green-left, “regardless of the user’s question,” and almost never mentioned the Christian Democrats even when the user’s answers matched that party’s positions. Its conclusion: as a voting aid they “consistently fail,” and the failure “is unclear and difficult to verify.” It issued the same warning again in March, for the municipal elections, and the pattern hadn’t changed.99 Autoriteit Persoonsgegevens, 21 October 2025, eight days before the Dutch general election; the follow-up warning on 12 March 2026 before the municipal elections. Deputy chair Monique Verdier: chatbots as voting aids “consistently fail.”

Notice what the regulator could and couldn’t do. It could test the oracle and publish the result. It could not tell any individual Dutch voter what the oracle had said to them, because that had never actually been written anywhere.

This is all an alarming development for democracy, no?

§ The honest counterweight

The companies are right that most people don’t type Parikh’s prompt. The consumer apps refuse to name a candidate, and refusal is what the majority of users meet. The Dutch regulator said explicitly that the bias wasn’t deliberate: it falls out of how the systems work, which sources they can reach, which they’ve been trained to prefer. Ballotpedia’s block was a configuration error, fixed on 23 October. Quebec’s newsrooms blocked the crawlers to protect their own business, which they were entitled to do, and the hole got filled by a hobbyist because holes do. And newspapers were never neutral. The Tribune that misreported Dewey had spent decades telling its readers who to hate. A voter who asks a machine how to vote has, in a sense, already given away the thing the ballot was for, and did the same with the parish priest. So the claim is narrower than “the chatbots are rigging elections.” It’s this. Every previous oracle had a lean, and the lean was public, which made it survivable. For the first time the lean is private, individual and unrecorded, and the people who hold it are the only people who could measure it and have no reason to. The problem isn’t the opinion. It’s that there is no next morning.

§ What stays scarce

Edition 006 argued that when a convincing message costs nothing, the message stops being evidence of anything. The political version arrived faster than the romantic one. When persuasion is free and private, the scarce thing isn’t the advice. It’s the record. Which is, in the end, what every reform in that list of oracles was: a record. The poll book, then the secret ballot that replaced it; the back issue; the voter guide; the regulator’s test. Each one made the voice that tells you how to vote into something that could be seen afterwards and held to account. Parikh’s five hundred questions were a record. The Dutch regulator’s test was a record. The McGill study of Quebec was a record. They were made by three people with grants, in a month, against systems that answer tens of millions of voters, and that’s the whole of the accountability that currently exists for the thing that has replaced the pamphlet. You don’t need to ban the oracle. Delphi ran for a thousand years. You need to carve the answers somewhere. Until someone does, every voter who asked a machine this autumn got a private endorsement from an institution with a view, and there is no one, including the institution, who can tell you what it said. The ledger below makes three claims. The first one is about who will insist on the carving.