antimemes The Great AI Debate

About and method

The Great AI Debate is the first release of antimemes.is, an arts lab by Benjamin Life. The lab starts from an old observation: we copy each other's wants, and with them each other's enemies, until groups define themselves mostly against each other. Its answer is anti-mimesis, meeting difference without imitating it or fighting it.

This debate is the lab's first instrument, for seeing. It maps where public figures stand on AI's future, read from their own words, and your answers place you on that map. The next instruments are for meeting the far side, with steelmen and anti-debates, and for moving: credit for describing another camp in words that camp would sign. Those are coming, and the map is what they will work from. The three instruments.

So the aim here is not to sort people into teams. It is to help people who disagree about AI describe each other in terms the other side would sign. The rest of this page is the method.

Who is on the map

We read 887 public works by 129 people who shape how America argues about AI and society, in their own words: essays, speeches, testimony, letters and interviews with labelled speakers. They include lab leaders and engineers, safety researchers and whistleblowers, senators and governors from both parties, union leaders, economists, civil-rights scholars, critics of the attention economy, and religious voices. Each position needs enough on the record to compare: 107 people are shown, and the others will appear as more of their public work is read.

Where the statements come from

We drafted a pool of 51 statements: the 24 used before, plus 27 on AI and the social contract (jobs, wealth, power, children, truth, workers, rights). Two blind reviewers checked each new one for slant and for asking two things at once, and proposed the questions both lists had missed. Every person was then read on all 51.

The 27 statements you answer were chosen from that pool by a fixed rule. Each has people on both sides in their own words. Each helps tell the schools apart, reaches people not yet covered, and doesn't repeat a statement already chosen. There are nine in each part: what AI is, what matters and what to do. Statements nearly everyone agrees on were left out, because they can't show where you differ: 52 of 55 people on record want binding rules for frontier AI, for example.

How each person's positions were checked

Every position shown was read from the person's full texts, not from summaries or reputation. Two readers worked blind, each reading everything we have by that person. A position is shown only if two independent readings support it: both readers found it, or one did and a third reader, going back to the source, confirmed it. When the readings don't agree, we show nothing. Silence is never counted as disagreement.

The readers were told to watch for the mistakes that misrepresent people most: a warning read as an endorsement, an opponent's view the author is describing, irony, a hypothetical, a forecast read as a value, and a policy view read into a description of a problem. Every quote is checked by code, word for word, against the source. In this round, 1,039 positions passed across the pool. The two blind readers never once picked opposite sides on the same person and statement, and of the 426 positions from the earlier audit, 423 were found again and none came out on the other side.

Before launch, every item shown here had one more independent check, by a reader asking one question: would this person, reading what we show, sign it as a fair account of their view? Of 670 positions on the survey's statements, 565 passed as they were. For 24, the reader found a clearer quote in the same text, and that quote is now shown (it is checked word for word like every other). The other 81 were hedged, conditional, about a narrower question, or not plainly the person's own view, so they were removed. Of 499 one-line summaries of people's positions, 22 said more than the person did, or put it in words they wouldn't use, and were rewritten. One statement, "An AI can be a genuine friend", was left with a single person agreeing, so the same selection rule replaced it. The group names were rewritten too, to describe what members believe without flattering or belittling anyone.

Then we read more. For each person, researchers looked for further work in their own voice on the statements they had been silent on, and found 356 more works. Two readers read them blind and independently, and a stance was added only where both found it, or where a third reader checked the passage and confirmed it. That added 306 positions and brought 22 more people onto the map. Where a new reading disagreed with a position already shown, a third reader saw both; once, they found the person's view too mixed to call, and that position was removed.

Some sources turned out not to be the person's own words: a guest essay in a newsletter, a third party's summary of a podcast, a machine-written summary, a transcript where speakers can't be told apart, or a leaked private meeting. Those are excluded entirely.

How you're compared

For each person we count the statements where you both took a side, and how many you agree on. A person needs at least three shared statements to be ranked, and the score is smoothed so that one of one doesn't beat nine of eleven. Your neighbourhood is the school of your nearest people.

Camps, schools and close circles

The groups come from the people we have read most fully, 56 of them, compared on the survey's statements and on the other ideas in the atlas they took a side on. Agreement counts only statements both people took a side on, and each statement's weight is divided by how much it overlaps the others, so one question asked five ways doesn't get five votes.

We then redraw the groups 300 times, each time from a resampled set of statements and with a different number of groups, and count how often each pair of people lands together. The groups are read off those counts at four depths: camps, whose members land together at least a quarter of the time; schools inside them (42%); branches (60%); and close circles who land together at least 78% of the time. Each group shows its own number. Where a group doesn't split further, there isn't enough evidence yet, not necessarily agreement.

Today that gives 4 camps, 12 schools, 14 branches and 5 close circles. People read less fully are placed down the tree only as far as they clearly agree with one group more than the others. Someone close to two camps is shown between them. If you answer as any grouped person did, your nearest match is in their own camp 89% of the time and in their own school 59% of the time (73% within your three nearest schools): the finer the group, the more statements it takes to tell it apart.

The live room

In the room, opinion groups are found the way Polis finds them: principal components of the votes, small base clusters, then two to five groups chosen by how cleanly they separate, checked by resampling. Bridging statements are found the way Community Notes ranks notes: a model that separates how much people agree with a statement from how much that agreement depends on which side they're on. A bridge is something people on both sides agree with.

Disclosures

What this can't tell you

Portraits

Photos come from Wikimedia Commons, used under their free licences and credited below. People without a freely licensed photo are shown by their initials.

Photo credits

If you think a thinker is misread, use “Report a misreading” on your result. Every report is answered.