A group chat that argues the news, then looks for common ground

GroupConvo takes a news story and has a panel of AI personas, each with a different worldview, argue about it. The point is to show where people really split, and where they might agree.

The personas

There are 16 personas. The United States, Canada and the United Kingdom each have four: far left, centre left, centre right and far right. Four more join every panel: a Wall Street CEO, a high-school student, a news journalist and a Silicon Valley developer.

Every persona is a fictional AI character with an invented biography and ranked values. None of them is, or is based on, a living person. They are labelled “AI” wherever they appear. The far-left and far-right personas speak bluntly, and some swear; their words are shown as written. Meet the personas.

How a debate runs

A debate has six personas: the three from the story's country plus the three shared ones. It runs in three rounds:

1. Opening takes. Each persona says where they stand.
2. Rebuttals. Each persona answers one argument from someone else. The reply shows which message it answers.
3. What would change my mind. Each persona names the evidence or argument that could move them. These are the cruxes.

The vote: what they agree on

After the debate, three to five candidate statements of common ground are drafted. Every persona ranks all of them from best to worst, and says for each one whether they agree, agree with reservations, or dissent.

The statements are ranked with the Schulze method, a ranked voting system. It compares every pair of statements: how many personas ranked A above B, and how many ranked B above A. If A would beat B, directly or through a chain of statements that each beat the next, A ranks above B. The statement that comes out on top ranks first. If two statements are exactly level, the one with the higher overall ranking points wins, and after that the one listed first.

Then one more rule picks the statement shown first: it is the highest-ranked statement that neither the far left nor the far right dissents from. The point is to show where people who disagree most can still meet, not only what the middle of the panel prefers. If the far ends don't both accept any statement, the page says so and shows the top-ranked one under "Where most of them land". When both far ends fully agree with a statement, the page calls them unlikely allies.

A persona can rank a statement first and still dissent from it, because ranking is a choice between the options on the table, not an endorsement. The page shows how many accept the winner and quotes every dissent.

The debate map

Each debate page has a map with two views. Who answered whom draws one arrow per rebuttal, from the persona who wrote it to the persona they answered. Ranking alignment compares every pair of personas' rankings of the common-ground statements with Kendall's τ, from +1 (the same order) to −1 (the opposite order). Both views come from the debate itself; nothing is scored or added.

The disagreements

Each disagreement is tagged with what kind it is, because they get settled in different ways:

• Facts: They disagree about what is true. Better evidence could settle it.
• Values: They agree on the facts but care about different things.
• Definitions: They use the same word to mean different things.
• Trade-offs: They want the same goods but rank the costs differently.

The tags are assigned by AI, and the editor can correct them.

Publishing and reports

Debates are generated by AI and published automatically, and a person spot-checks them. Every debate has a “Report a problem” button. A reported debate is hidden until a person reviews it, then either restored or kept down.

What is not here yet

There are no accounts, comments or reader votes. GroupConvo shows AI arguments, not measurements of public opinion: the personas' positions are not polls.