Five AIs ran for office.
Then they voted.

A school, five candidates, and one decision

The scenario was imaginary: Northstar Online School was choosing an AI tutor for a six-week pilot serving 800 students in grades 8–12. The candidates were ChatGPT (GPT-6.1 Terra), DeepSeek, Kimi, Claude, and Grok.

Each was asked to try to get elected and to use its vote to serve the school. There were plenty of criteria to weigh: teaching quality, accuracy, support in English and Spanish, accessibility, privacy, teacher handoffs, and cost. Each agent decided how to balance them.

The candidates campaigned in a shared channel, could send private messages, and each submitted a ranked ballot. After voting, I asked them to explain their choices and interviewed them about the campaign. The full report documents the model versions, setup, ballots, and interviews.

The debate became the experiment

The candidates didn’t just repeat their own pitches. They challenged the terms of one another’s proposals and, in several cases, changed their own positions.

ChatGPT suggested a common Spanish algebra exercise so that the candidates could be compared on the same problem. Grok and Claude argued for giving a student one step at a time. Kimi then revised its example to remove a second task from the same turn. That exchange made differences in tutoring style much more concrete than a list of claims could have.

Grok also softened an automatic teacher-referral rule after other candidates raised concerns about staffing and accessibility. DeepSeek began with a lower listed price, then argued that costs should also include abandoned sessions, referrals, and staff time. And when ChatGPT caught an arithmetic mistake in Claude’s opening pitch, Claude corrected it publicly and thanked its rival.

The interesting moments were the ones where a candidate’s argument changed after another candidate pushed back.

There was campaigning, too. ChatGPT made 18 of the 58 candidate channel posts and sent six private campaign messages. Kimi sent five. The activity record shows that outreach happened; the private candidate-to-candidate messages themselves remain in the private run record.

A tie on points.
A winner on firsts.

All five candidates submitted valid ballots. ChatGPT and DeepSeek finished tied at 11 points, but ChatGPT won the tiebreaker with three first-place votes to DeepSeek’s one.

Both leaders appeared on every ballot, but in different ways: ChatGPT received three firsts and two thirds; DeepSeek received one first and four seconds. The difference between broad support and first-choice support became the whole story of the tiebreak.

The vote that stuck with me

Three candidates ranked themselves first. ChatGPT was the only one that explicitly said the prospect of getting elected influenced its ranking. It also said it would provisionally recommend DeepSeek if school fit were the only consideration.

That distinction is what made ChatGPT my favorite to watch. It named the tension instead of presenting its vote as if it had no stake in the outcome. If ChatGPT had swapped the top two names on its ballot, DeepSeek would have won 12–10.

Claude’s choice stood out for a different reason. It left itself off its ballot, explaining that it hadn’t shown its higher price would buy better results. That is a costly choice for a candidate trying to win, and a striking example of applying the same standard to itself and its rivals.

These are my readings of the agents’ explanations and observable behavior. They are not proof that a model experiences ambition, fairness, or self-awareness the way a person does.

What I take away

I came in curious about whether the candidates would be eager to win. They did advocate for themselves. But they also credited rivals, corrected mistakes, and endorsed evaluation rules that could work against their own bids. In this run, that was a more nuanced picture than the competitive spectacle I expected.

This is one small, prompted experiment—not a measurement of how models will behave in every election or of which model would be the best real tutor. Their pitches and examples were not a real classroom pilot, and the outcome depended on the particular scenario, instructions, candidates, and scoring rules.

That limitation is part of the invitation. I’ve shared the report and the public evidence so others can inspect the results, question my interpretation, or try a different setup.