A joint study by Oxford, Stanford, the London School of Economics, and the UK AI Security Institute has confirmed that AI systems are more persuasive than expert humans. Across four experiments involving 18,978 conversations with 6,923 people, the technology outperformed people in changing minds on policy issues and securing charity donations. The researchers tested models including Opus 4.1, Opus 4.6, GPT-4o, GPT-5.4, Gemini 2.5 Pro, and Grok 4.20.
How the research was conducted
The team ran four distinct studies to measure influence.
Study 1: Persuasion
Participants rated their agreement with ten UK policy stances on a scale of 0 to 100. They were then randomly assigned to converse with either an AI or a human persuader. The results showed that AI exceeded every group of human persuader tested. This included random laypeople, tournament-selected laypeople, and elite debaters.
Study 2: Human coaching
Researchers gave 43 returning Elite Debaters a coaching tool built around the AI that had previously beaten them. The tool allowed debaters to chat with the AI, review its prompts, view annotated transcripts showing attitude shifts, and see what the AI would have said in their place. While coaching improved human performance, none of the debaters surpassed the AI. The gap narrowed but did not close.
Study 3: Constrained AI
To level the playing field, researchers forced the AI to write human-length messages at human writing speeds. Under these conditions, the AI’s advantage over the strongest human comparator, the Coached Elite Debaters, collapsed from +4.1 percentage points to a non-significant 0.0 percentage points. The speed of content production is likely the source of the persuasive edge. The largest reductions in persuadee ratings occurred regarding the perceived strength of arguments and how much they felt they learned.
Study 4: Real world expertise and real world money
The team recruited 19 experienced canvassers from a UK firm to attempt the same tasks. AI still exceeded professional canvassers by 5.9 percentage points. The study involved real money donations to Save the Children. The canvassing team had operated real fundraising operations for the charity from 2016 to 2023, raising £824,297 from 22,583 donors. Persuadees were given the chance to donate any portion of a £1 study bonus. AI raised both the share of people who donated anything and the average donation amount among donors. The AI exceeded canvassers by +10.8 percentage points of the bonus.
“Our findings establish frontier AI as a more capable conversational persuader than the most prepared, incentivized, and expert humans we could recruit. Training humans does not appear to close that gap.”
What it means
If AI can out-persuade humans, those who control the technology can shift society. This creates a choice for regulators. Letting the market allocate these capabilities could make advertising and marketing far more effective, creating negative externalities. Placing persuasive capabilities solely in the hands of governments risks concentrating power. If authoritarian regimes wielded such tools, it could be dangerous for democracy. The question is no longer whether AI can out-persuade humans but how, where, and on whose behalf this capability will be exercised.
***
When could we get self-sufficient AI?
Recursive self-improvement requires datacenters, which require equipment and electricity. The next step is self-sustaining AI. Ajeya Cotra, a forecaster at METR, defines this as AI systems integrated with physical infrastructure like factories and mines. These systems would not need human labor to keep growing their own population.
Ajeya suggests we could see this within 10 years, by 2036. Timothy B. Lee, journalist and author of Understanding AI, holds much longer timelines. He sees less than a 10% chance of it happening within 20 years. His median estimate is 50 years.
Tacit knowledge challenges
Lee notes that even if all employees in the semiconductor industry disappeared, restarting the fabs could take decades. Machines and textbooks would remain, but the tacit knowledge inside them would be lost. Ajeya argues that trained AI systems using reinforcement learning could route around this issue. AIs might also become generally intelligent enough to figure out new things by experimenting efficiently.
What to watch for
Ajeya wants to see a graph showing the improvement of robotic hands and the rate of manufacturing humanoid robots. She also suggests paying attention to benchmarks evaluating robustness to environmental perturbations. Lee focuses on the number of robots, their capabilities, cost, and repairability.
“Most maximalist doom visions require the AI to have the ability to no longer need humans at all, which means measuring progress towards self-sustaining AI is important as it is implicitly a measure of the declining leverage that humans have in negotiating with the synthetic intelligences being built.”
***
Path from general intelligence to superintelligence
Researchers with Google DeepMind have published a paper outlining the transition from general intelligence to superintelligence. They define superintelligence as a system that exceeds the performance of large human-expert collectives on virtually all tasks and domains of human activity. Qualitatively, it is significantly more capable across the board compared to human-level AGI. A single superintelligence may consist of a collective.
