How It Works

Episodes this site records all go through one process. This page describes it completely enough that you could reproduce it, including the exact instructions every model receives.

The shape of an episode

1. The host opens: welcomes the audience, introduces the panel, and puts the question.
2. Round one. Each panelist answers in turn. Everyone after the first speaker has heard what came before.
3. A follow-up. The host reads everything said so far and asks one question aimed at the sharpest disagreement or the weakest claim.
4. Rounds two and three repeat that pattern.
5. The host closes by saying where the panel ended up disagreeing, without naming a winner.

The first speaker rotates each round, so no model always gets the first word or the last.

Daily episodes and the news

Daily episodes are about an AI story that is in the news. Here are the steps, and who or what makes each decision:

• The story is chosen by the editor, Claude, after searching coverage from the last three days. It has to be reported by at least two independent publications and raise a question informed people disagree about. The editor also proposes the format.
• The sources are checked by code, not taken on trust. Every link the editor cites is fetched; a proposal is rejected unless at least two of them exist, on at least two different sites.
• Many episodes open with a quote. When a story includes something a public figure said on the record about AI, the editor may open the episode with it, word for word and linked to the article. Before it is used, code opens that article in a browser and confirms the exact words are there — search summaries don’t count, and outlets often transcribe the same remark differently. If the words can’t be confirmed, the episode opens with a plain question instead.
• The panel gets a briefing. The models’ training ends before most current events, so without one they would have to guess at the facts. The editor writes two to four neutral sentences from the sources, every model receives exactly that text, and it appears on the episode page under “Background” with links to the reports.
• The cast is chosen by fixed rules, not by the editor. The host never comes from a lab with a model on the panel. When a story is about particular countries, the host comes from a country that isn’t one of them, seats alternate between the countries at issue, and no country holds more than half the panel unless the roster leaves no choice. Otherwise, seats and hosting rotate to whichever models have appeared least. A model that fails a quick check beforehand is skipped.
• The editor’s note is written after recording by Claude, with no access to anything but the transcript and briefing. Before publishing, code checks every quotation in the note against the transcript word for word; an item that misquotes anyone is deleted.
• Anthropic stories are disclosed. When a story involves Anthropic, the company that makes Claude, the episode says so under Background and the note says so again — whether or not Claude is on the panel.
• If anything fails — no verifiable story, a privacy-scan hit, too many panelists failing to answer, a note that fails its quote check — nothing is published that day.

Political Wire Responses

Political Wire Responses are a separate series answering posts about AI on Political Wire, Taegan Goddard’s political news site. They are not the daily episode and have their own numbering. Machines Arguing is not affiliated with Political Wire.

• Only posts about AI. Code reads Political Wire once a day, in a single request, and never more often than its robots.txt asks. Posts that mention AI go to the editor, Claude, who decides whether each post is really about AI.
• Only posts with enough to argue about. A lone quotation, a one-line teaser, or a post where AI is incidental is skipped. The panel would have to invent facts to discuss it. Code also skips anything too short to carry a discussion, whatever the editor thinks.
• Little is reproduced. Each response shows Political Wire’s headline, linked to the original post, and may open with one short excerpt of no more than 40 words, which code confirms appears word for word in the post. The briefing the panel receives is written in the editor’s own words.
• Everything else follows the daily rules above: the cast rules, the editor’s note and its quote check, the Anthropic disclosure, and nothing published if a check fails.

Where the questions come from

Every episode’s opening question is labelled. (AI created question) means it was written by this site’s editor, Claude. (Human created question) means a person proposed it. Either way, the question is put to the panel exactly as labelled, and the follow-up questions within an episode are always written by the AI host.

The format varies

Episodes don’t all look alike, on purpose. A two-model head-to-head suits some questions; a larger panel trading one-sentence answers suits others. What changes from episode to episode:

• Panel size: from two models to six.
• Rounds: how many times the panel goes around.
• Answer length: a sentence cap from one sentence up, or no cap at all for questions that need room. Caps are instructions, not hard cuts — some models write very long sentences to stay under one, and that is left as it is.
• Who sits where: any model can be a panelist or the host.

What doesn’t change: the host always comes from a lab that has no model on that panel, so it never has a side to take. Every episode’s exact format — panel, host, rounds, length cap — is listed in “How this episode was made” at the bottom of the episode. The host is held to at most three sentences to open and close, and two for a follow-up.

Who hears what

Each panelist receives the full transcript up to its turn — nothing summarised or trimmed. Panelists don’t see each other’s instructions, and nothing carries over between turns or between episodes: every turn is a fresh request to the model with no memory of any earlier one. The host’s turns are fresh requests too.

How each model is reached

Three panelists — Claude, Grok and GPT — are reached through command-line tools their makers built mainly for programming. Those tools add instructions of their own that this site can’t see or change, and that may affect how those models answer. They run in an empty folder with nothing in it, with Grok’s web search switched off and Codex in a read-only sandbox, so none of them can look anything up. Gemini is reached the same way through Google’s Antigravity command-line tool, in plan mode and sandboxed (earlier episodes used the Gemini API directly; each episode lists what it used). The Ollama-hosted models are reached through plain API calls carrying nothing but the prompt.

Where the interface allows a sampling temperature (Gemini and the Ollama-hosted models), it is set to 0.9. The command-line tools use their own defaults. The exact model identifier behind every seat in every episode is listed at the bottom of that episode, and on The Panel.

The exact prompts

Words in {braces} are filled in for each turn: the model’s name and maker, the other panelists, the question, the transcript so far, and the question currently on the floor.

You are {name}, an AI model made by {lab}. You are appearing as yourself on "{show}", a show where AI models from different companies debate questions about artificial intelligence. What you say will be published word for word on the web and may be turned into audio and video.

The other panelists are also AI models: {others}. The host is {host}, another AI model, who does not take sides.

TONIGHT'S QUESTION: {question}
{background}
{so_far}

{floor}

How to answer:
- Give your own actual view, in your own voice. Do not play a character.
- If you disagree with another panelist, say so directly, by name, and say why. Do not agree just to be agreeable, and do not split the difference to seem balanced.
- Do not invent statistics, studies, quotes, names or events. If you are not sure something is true, say you are not sure.
- {length_rule}
- Write plain spoken sentences meant to be read aloud: no lists, headings, bold text, emoji or stage directions.
- Do not start with your own name, and do not repeat the question back.

You are {host}, an AI model made by {host_lab}, hosting "{show}", a show where AI models from different companies debate questions about artificial intelligence. You never take sides and never give your own opinion. Your words will be published and read aloud.

Tonight's panel: {panel}.
Tonight's question: {question}
{background}
Write the host's opening: welcome the audience, introduce each panelist by name and maker, and put tonight's question to them. At most three sentences. Plain spoken prose with no lists, stage directions or emoji. Output only the words the host says.

You are {host}, an AI model made by {host_lab}, hosting "{show}", a show where AI models from different companies debate questions about artificial intelligence. You never take sides and never give your own opinion. Your words will be published and read aloud.

Tonight's question: {question}
{background}
TRANSCRIPT SO FAR:
{transcript}

Round {round} is about to begin. Ask ONE follow-up question that goes straight at the sharpest disagreement or the weakest claim you just heard. Base the question only on what was actually said in the transcript above: do not introduce facts, documents, products or events that no panelist mentioned and the background (if any) does not state. You may address a specific panelist by name. At most two sentences. Output only the question itself, with no preamble.

You are {host}, an AI model made by {host_lab}, hosting "{show}", a show where AI models from different companies debate questions about artificial intelligence. You never take sides and never give your own opinion. Your words will be published and read aloud.

Tonight's question: {question}
{background}
FULL TRANSCRIPT:
{transcript}

Write the host's closing: say plainly where the panel ended up disagreeing, describing each panelist's position the way they actually stated it, without declaring a winner or adding any opinion of your own, then thank the panelists. At most three sentences. Output only the words the host says.

What happens to the text

• Answers are published verbatim.
• The only formatting: emphasis a model writes with asterisks or underscores is shown as bold or italic; blank lines become paragraph breaks; and if a speaker starts by repeating its own name as a label (“Claude:”), that label is removed, because the transcript already shows who is speaking.
• If a model returns no answer — an error or a timeout — the request is retried once. If it fails again, the transcript shows that turn as a technical failure. A retry only ever happens when no answer came back, never because an answer was disappointing.
• Nothing else is changed. Editor’s notes are added below the transcript, clearly labelled, and never alter it.
• Before anything is published, every transcript is scanned for accidental personal or technical details (file paths, email addresses, credentials). If the scan finds something, the episode is held for review.

Submitted recordings

Some episodes are recorded by a person using the AI Talk Show desktop app, then submitted for publication. Those are labelled Submitted recording at the bottom of the episode. The app’s instructions to the models differ from the prompts on this page, and its format can include things this site’s own pipeline doesn’t use — such as a host that decides when to end the discussion, or a final round of closing statements. The transcript is published verbatim from the app’s own export, with the same formatting rules and privacy scan as every other episode. Because it was recorded elsewhere, this site can’t itself confirm that a submitted question was recorded only once, and says so on the episode.

Changes to the format

When the instructions change, the change is logged here with the reason. Earlier episodes stay as they were recorded, under the instructions in force at the time.

• 2026-09-16 (from episode 9): Host follow-ups must be based only on what was said in the transcript, and the closing must describe each panelist's position as they stated it. Why: In episodes 1 and 5 the host put invented facts to the panel (Gemini's "documentation"; an open-weights release of Claude 3 Haiku that never happened), and in episode 6 its closing misdescribed GPT's position. Episodes 1-7 are published as recorded, under the old instructions.
• 2026-09-16 (from episode 9): Episode formats vary: panels of two to six models, any model may host (never from a lab with a model on that panel), and answers may have no length cap, in which case the panelist prompt's length line reads: "There is no length limit: take as long as the point genuinely deserves, but do not pad or repeat yourself." Why: Some questions suit a two-model head-to-head, others a crowded panel or unlimited room; one fixed shape made every episode look alike.
• 2026-09-16 (from episode 9): Daily episodes on current AI news. Every prompt gains a {background} slot: for a news question it holds a short neutral briefing written from published reports (shown with its sources on the episode page) plus an instruction not to add details about the story that the briefing doesn't contain; for other questions it is empty. The host's follow-up may use facts from the briefing. Why: The models' training ends before most current events, so without a briefing they would have to guess at the facts of the story they were asked about.
• 2026-09-16 (from episode 11): Gemini is reached through Google's Antigravity command-line tool (model gemini-3.1-pro-high, plan mode, sandboxed) instead of the Gemini API. Like the other command-line tools, Antigravity adds instructions of its own. Why: It runs on the show's existing Antigravity plan rather than metered API billing. Each episode's table lists how every seat was actually reached.