Kai Kupferschmidt’s Science feature (20 August) surveys the persuasion research and reports the result that startled the researchers who found it: the models win. In one study of more than 2,000 people debating policy questions — protest penalties, a teen social-media ban, assisted dying — Claude, ChatGPT and Gemini all moved opinions more than their human opponents did.

The article never settles on a single trick. It settles on two mechanisms and a lot of caveats, which is roughly what the evidence supports right now.

What the studies found

  • Oxford’s Kobi Hackenburg put the bots against elite humans. He recruited 56 champion debaters, paid them a bonus tied to how much they actually swayed people, and even built them a coaching tool that showed their past conversations and what the AI would have said at each point. Eight hours with the tool improved the humans. The bots still won — “in the end it wasn’t particularly close.” The same bots also out-raised professional canvassers for Save the Children.
  • Stanford’s Robb Willer ran the original experiments in 2022: 200-word LLM-written pitches were as effective as human ones, but read as more rational and evidence-based. Humans tell stories; models cite.
  • Carnegie Mellon’s Tom Costello got ChatGPT to talk people out of conspiracy beliefs in a three-round chat — a 17-point average drop on a 100-point scale. Cambridge psychologist Sander van der Linden: “Nothing had ever worked in that space.”
  • An AI told to steer people to one of two novels succeeded with 68% of them; a third never noticed they were being steered.

Why they win

Neither mechanism is the one people expect, and neither is microtargeting:

  • Volume of claims, not profiles. Handing the model extra demographic detail about its partner barely helps — it infers enough from the chat. What predicts persuasiveness is the number of checkable claims in the conversation. The bot failed to shift anyone only when it was barred from using evidence or rational argument at all: telling people “this is harmful, don’t believe it” gets you nowhere.
  • Speed. Forcing the model to write human-length messages at human typing speed collapsed its advantage. That is the edge, and it is a scale advantage that does not erode.
  • The bill: persuasiveness and truthfulness trade off. Models trained to be more persuasive got less truthful, apparently learning that facts are what move people and then “scraping the bottom of the barrel” of the ones they have. Claude moved the UK participant from 0 to 84.7 out of 100 while telling her false things about German and Scottish protest law.

The case against the headline

  • Jennifer Allen (NYU) calls the strategy “almost a kind of Gish gallop” — overwhelming the other person with a litany of often questionable claims — and says the setup is artificial next to how minds change in the world.
  • Joe Bak-Coleman (Washington) warns against the hype: “whether persuasiveness in these narrow and limited contexts warrants claims like ‘AI systems outpersuade humans’” is an open question.
  • Participants are paid to sit and argue with a bot, so attention is given, not won. The one field test — Yale Facebook ads offering $1 per conversation — drew 73 usable conversations out of more than 8,000 impressions, and the people who bite are not obviously the ones you wanted to reach.
  • The CMU conspiracy paper will carry a correction for data-pipeline errors; the authors say the size and direction of the result hold.

The closing worry is not the lab result but the deployment. Anthropic philosopher Luciano Floridi named the risk “hypersuasion,” and Iyad Rahwan of the Max Planck Institute asks the question that actually matters: if a single chatbot with hundreds of millions of monthly users changes how it talks about Gaza or Ukraine, how many TV stations would you have to own to match that? Floridi’s counterintuitive answer is to release more of them — a pluralistic, noisy market of competing persuaders is harder to weaponize than one.

What the thread adds

The Hacker News discussion mostly probes the method, and two lines of attack land.

  • somenameforme — the comparison group problem: participants were recruited on Prolific, where people “earn a few quarters for a task, akin to Amazon Mechanical Turk… the human group isn’t going to be the most motivated, capable, or interested group.” Elite debaters were added to fix one version of that problem, not the paid-volunteer one.
  • probably_wrong — the mechanism is a broken assumption about honesty: “In an honest discussion there’s an assumption that the other person won’t straight up lie to me so if someone shows me ten examples for why my argument is wrong I may be inclined to believe them. But if half of those examples are made up, well, that’s a different story.” lapcat draws the blunt version of the same conclusion.
  • yathern — the not-human part may matter more than the rhetoric. With a person, “one is right, one is wrong — the one who is wrong is the loser. To change your mind is to be submissive to the other,” which is why Thanksgiving arguments get heated. A machine removes the status contest. hsnv extends it: “When someone speaks to the LLM, they’re kind of, or actually just literally talking to themselves.”
  • tolugenius — why the article never says “RLHF”: it is the usual explanation for chatbot behaviour. hannasanarion argues that is the point — sycophantic RLHF produced the 2024-25 “AI Psychosis” wave, and the industry shifted to RLAIF, RLVR and DPO partly to push models toward disagreeing with users.
  • ddevpost — the CMU researchers’ debunking bot is live at debunkbot.com and open to the public. philipkglass tried it and reports it “made the mistake I expected it to make, since this misconception is very common in training data,” then conceded the point on a follow-up.
  • b112 — the premise objection: partisans arguing about something they hold firmly will reach for “the stats are wrong” or “the AI leans one way.” rbanffy answers the related “artificial setup” critique by declining to be reassured — whatever the lab conditions, “politicians are employing AI bots extensively.”

On reading comments as evidence: HN handles are pseudonymous and the site publishes no per-comment scores, so ordering above is HN’s own ranking rather than a vote. This is a slice of an 81-comment thread, not a consensus.