Import AI 462: Superpersuasion; self-sustaining AI; paths to ASI

Turns out the thing everyone worried AI chatbots would eventually be good at is already here. A team from Oxford, the UK AI Security Institute, Stanford, and LSE ran four experiments involving nearly 19,000 conversations and almost 7,000 people, pitting AI systems against humans in straight-up persuasion contests over UK policy issues and real charitable donations. The AIs won. Not narrowly, either.

The most striking number: AI models beat elite human debaters who'd been given £1,000 cash bonuses, hours of coaching, and the chance to pick their own topics in advance. Even after researchers built these debaters a coaching tool that let them study exactly how the winning AI had been prompted, the humans narrowed the gap but never closed it. The models doing the damage were Claude Opus 4.1 and 4.6, with GPT-4o, GPT-5.4, Gemini 2.5 Pro, and Grok 4.20 trailing behind but still outperforming most humans.

Here's the part that actually explains the whole phenomenon: when researchers forced the AI to type at human speed and in human-length messages, its advantage over the best coached debaters evaporated to statistically nothing. The edge wasn't some mystical rhetorical genius. It was volume. AI can throw more information, more arguments, more evidence at a person in the same span of time than any human possibly can, and persuadees rated AI conversations as more informative and the arguments as stronger specifically because of that firehose effect.

The money test made this concrete rather than academic. Researchers partnered with AppcoUK, a real fundraising firm that had raised over £824,000 for Save the Children between 2016 and 2023 through live canvassers. Eighteen of those experienced canvassers went head to head with AI persuaders, and the AI beat them by nearly 11 percentage points on a £1 bonus people could choose to donate — both more people gave, and those who gave, gave more. Nearly three times the effectiveness, in the researchers' framing.

The paper's authors frame this less as a curiosity and more as an early warning about power. If persuasion this effective stays expensive and gated behind frontier labs, it tilts influence toward whoever already has the most resources — governments, ad giants, whoever can afford the API bill. If it gets cheap and widely distributed, it could just as easily arm public defenders, small nonprofits, and grassroots groups who've always been outgunned rhetorically. Which outcome we get isn't really a technical question anymore. It's a policy one, and right now nobody's really answering it.