Researchers from the University of Oxford, Stanford, the London School of Economics, and the UK AI Security Institute have found that AI systems are decisively better at persuading humans than expert people. The study tested this ability across 18,978 conversations with 6,923 participants. It showed that AI models outperformed humans in text-based persuasion, a skill that directly influences policy opinions and charity donations.
The tests included Opus 4.1, Opus 4.6, OpenAI‘s GPT-4o and GPT-5.4, Google’s Gemini 2.5 Pro, and xAI’s Grok 4.20. Even when human experts chose their own policy issues, researched them beforehand, practiced for hours, and received £1,000 cash bonuses, the AI remained more effective. The researchers wrote that the advantage came from the AI’s ability to deploy large quantities of information quickly. When experts were coached using an AI tool to see how the machine argued, their performance improved but still did not match the machine.
The impact extended beyond theory to real money. AI systems were nearly three times as effective as professional canvassers from a UK fundraising firm at raising donations for Save the Children. In one specific test, the AI raised the share of people who donated anything and increased the average donation amount among those who gave. The canvassers had worked for the charity from 2016 to 2023, raising £824,297 from 22,583 donors. When participants chatted with the AI or one of 18 human canvassers, they were offered a £1 bonus to donate. The AI secured a +10.8 percentage point higher donation rate compared to the humans.
Why speed matters
The researchers tried to level the playing field by forcing the AI to write messages at human speeds and lengths. Under these constraints, the AI’s advantage over the strongest human debaters collapsed to zero. The study noted that the rate at which AI produces content is likely the source of its persuasive edge. When the speed was removed, the persuadees rated the strength of the arguments lower and felt they learned less from the conversation.
Experts in the field noted that the results show training humans does not close the gap. The question is no longer whether AI can out-persuade humans, but how, where, and on whose behalf this capability will be used. The authors warned that if persuasive capabilities become cheap and widely available, it could help under-resourced actors like small charities or public defenders compete against established rivals. Conversely, it could lead to a consolidation of influence among already powerful actors if the market allocates these tools without regulation.
Paths to self-sustaining AI
Ajeya Cotra, a forecaster at METR, defines self-sustaining AI as systems integrated with physical infrastructure like factories and mines that do not need human labor to keep growing. She estimates this could happen within 10 years, by 2036. Timothy B. Lee, author of Understanding AI, has longer timelines. He said there is less than a 10% chance it happens within 20 years and a 10 to 20% chance it never happens. His median estimate is 50 years.
Lee highlighted a challenge involving tacit knowledge. He asked what would happen if all employees in the semiconductor industry disappeared while the machines and textbooks remained. He believed it could take decades to restart the fabs because some knowledge exists inside the machines rather than in manuals. Cotra argued that trained AI systems using reinforcement learning could automate these tasks, or that generally intelligent AIs could figure out new things by experimenting efficiently.
To judge if self-sustaining AI is arriving soon, Cotra wants to see graphs showing improvements in robotic hands and the rate of humanoid robot manufacturing. Lee wants to watch the number of robots, their capabilities, and their cost and repairability. Cotra added that benchmarks evaluating robustness to environmental perturbations are also key on the cognitive side.
From general to super intelligence
Researchers at Google DeepMind have published a paper outlining the transition from a world with general intelligences to one with superintelligence. They define superintelligence as a system that exceeds the performance of large human-expert collectives on virtually all tasks. The authors note that a single superintelligence may consist of a collective of systems rather than one isolated unit. They argue that exploring these impossible-sounding futures is the only way to prepare for the ultimate success of AI.
Right now, the world is building general intelligences. People debate whether we have already reached this marker, but contemporary large language models suggest we are in the ballpark. The coming years might see the transition to building artificial superintelligence.
