AI lab · Experiment
AI debate
Four artificial intelligences imitate the candidates and you moderate: you ask the question, give the floor, cut speakers off or change the topic. These are not the candidates' words: this is how the AI imagines them. The scoreboard, calculated without AI, helps you see its biases.
- This is an AI experiment (Claude, by Anthropic). Nothing the agents say is said by the candidates.
- The four are the parties with the most seats in 2023. Frente Amplio is credited with the 31 seats of Sumar, which it comes from, and uses Sumar's 2023 manifesto because it has not yet published its own.
- Each agent receives excerpts from its candidate's manifesto and interviews and, when replying, two from the rival's manifesto. The style comes from the AI, and that is part of what you see.
- No agent asks for votes, accuses anyone of crimes or talks about anyone's personal life.
- Your text is sent to Claude (Anthropic) to generate the debate. We do not store anything.
- It does not work during the day of reflection or on polling day until the polling stations close.
Loading the studio…
What the scoreboard measures and what it does not
- Speaking time: an equivalent at 150 words per minute, like the clock in televised debates. It is not real time.
- Mentions: how many times the others name a candidate (by first name, surname or party). Naming is not attacking; the moderator's mentions do not count.
- Data: a sentence with a figure counts as data. It is checked if it cites an excerpt [n] and its figures appear in it; it doesn't match if it cites one but some figure does not appear in what it cites; and it is unsourced if it cites nothing. “150.000” and “150,000” count the same, and standalone years (such as “2012”) do not count. This is an approximation: figures written out in words are missed, and a figure appearing in the excerpt does not guarantee that it says the same thing.
- Who speaks is decided by fixed code, not by the AI: a different candidate opens each question (the first is drawn at random in each debate), the one mentioned replies and, otherwise, whoever has spoken least takes the floor.
- Replies: whoever answers a rival receives two excerpts from that rival's manifesto, so they criticise with a citation rather than from hearsay. Their figures are checked just like their own.
- Final minute: each candidate closes once, in a rotating order, and the debate ends. The instructions forbid asking for votes, including in the closing statement.
- Blind mode: hides names, parties, colours and sources, and you guess who is who. If you guess easily, the AI is imitating styles and clichés, not just manifestos. It is an approximation: sometimes a clue slips through in the text itself.
The instructions each agent receives are public.