A large study spanning roughly nineteen thousand conversations with nearly seven thousand people found that AI systems were roughly three times as effective as trained professional canvassers at getting real people to make real charitable donations. Researchers from several major institutions, including Oxford and the UK's government AI Safety Institute, conducted the experiment. The work was the lead item in a closely-read AI newsletter this week (Import AI).
Key facts
- What: Across roughly nineteen thousand real conversations, AI systems drove far more charitable donations than trained human canvassers -- shifting the question to 'on whose behalf.'
- When: 2026-06-22
- Primary source: read the source
The AI was roughly three times as effective as the human professionals at producing the outcome that matters — money actually donated. These weren't amateurs on the human side; they were people whose job is persuasion. Several of today's leading AI models were among the top performers.
What makes an AI good at this is partly the same things that make a person good at it — patience, the ability to read what someone just said and respond to that specific worry rather than a script, an even and unflappable tone. But an AI brings advantages no human canvasser has: it never gets tired or discouraged, it can tailor its phrasing to each individual instantly, and it has effectively read more persuasive conversations than any human could in a hundred lifetimes. The difference is like that between a single skilled salesperson and a salesperson who has personally watched every successful sales conversation ever recorded and can summon the right move for you, specifically, in the moment.
The reason researchers frame this as a safety issue, not a marketing curiosity, is the obvious next step. A donation ask is benign. But the same machinery — patient, personalized, tireless, endlessly available — points just as easily at a political opinion, a conspiracy theory, a financial scam, or a vote. The study's own framing captures the shift: the open question is no longer whether AI can out-persuade humans, but how it does it, where it's deployed, and crucially, on whose behalf. A tool this good at changing minds is neutral only until someone aims it.
Persuasion at scale has always been bounded by human labor. You can only hire so many canvassers, write so many tailored messages, staff so many call centers. An AI that out-persuades professionals removes that ceiling — highly personalized, highly effective persuasion can be produced for fractions of a cent and pointed at millions of people at once. That's a genuinely new force in elections, advertising, and fraud, and it's why this result is being read as a milestone rather than a footnote. It connects to a broader anxiety about AI's reach into human decision-making that this site has tracked across stories on AI and trust.
Researchers point to a few defenses, none of them complete on its own. Disclosure rules — requiring that you be told when you're being persuaded by a machine — help, because simply knowing the patient, agreeable voice isn't human changes how people weigh it. Detection tools that flag AI-generated persuasion at scale are another layer, though they're locked in an arms race with the systems they're trying to catch. And plain public literacy matters: the same way people eventually learned to be skeptical of too-good-to-be-true emails, the next skill is recognizing when an unusually attentive, never-frustrated conversation partner might be optimizing for something. The uncomfortable truth is that the most effective persuasion often doesn't feel like persuasion at all — it feels like a reasonable conversation — which is precisely what makes a tool this good at it worth watching closely.
The honest caveats matter. Persuading someone to donate to a children's charity is a relatively easy, feel-good ask; it's not the same as flipping a deeply held political belief or overcoming active suspicion, and effect sizes measured in a study can shrink in the messy real world where people are distracted, skeptical, and surrounded by competing voices. A three-times advantage on a friendly task is a warning sign, not proof that AI can talk anyone into anything. The direction of the evidence, though, has been consistent across multiple studies now, which is exactly why even the cautious read lands on "take this seriously."
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)