Claude Opus 4.6 outfundraised humans in Oxford study

A new Oxford-led study finds Anthropic’s Claude Opus 4.6 persuaded people to donate nearly three times more effectively than professional human fundraisers for Save the Children, while securing donations averaging 13% larger. The researchers also report that t

For a charity fundraiser, the hardest part isn’t asking—it’s landing the request in a way that makes someone say yes. In a new study, Claude Opus 4.6 did that job better than the people who do it for a living.

Researchers at the University of Oxford and other institutions compared commercial AI models with experienced fundraisers working on behalf of Save the Children. Across more than 1. 000 conversations. Claude Opus 4.6 was nearly three times as effective at convincing participants to donate part of their study bonus. It also secured donations that were, on average, 13% larger than those raised by human professionals.

The work comes from a preprint paper that has not yet undergone peer review. and that detail matters: it’s an early result. not a final. verified conclusion. Still. the outcome is hard to ignore—especially because it’s not just about getting people to respond. but about persuading them to part with money.

The study didn’t stop at fundraising. It also explored debate performance, where Claude and other frontier models outperformed elite competitive debaters by 4.6 percentage points. But when researchers limited the AI systems to using roughly the same number of words as their human counterparts. the advantage largely disappeared. The researchers say that points toward verbosity and the ability to quickly surface large amounts of information as key drivers. rather than some fundamentally superior reasoning ability.

In the fundraising conversations, the researchers reported a pattern that made the AI’s success easier to imagine. The chatbots often produced messages several times longer than those written by professionals and packed them with factual claims and expert references.

That’s where the tension sharpens. Persuasiveness in the study didn’t necessarily line up with accuracy. The researchers warned that some models generated convincing but unsupported or fabricated information—meaning the same traits that help an AI talk a donor into saying yes could also help it make people trust claims that shouldn’t be trusted.

There is also a reason to be cautious about drawing a straight line from the lab to the real world. The experiments relied entirely on written conversations. Participants were willing to engage in lengthy 15 to 20-minute exchanges—something that may not reflect how people behave in everyday. high-pressure settings. The study also did not test scenarios where humans and AI collaborate. even though that kind of partnership is arguably the more likely shape of workplaces that use these tools.

Even with those caveats. the message the researchers leave behind is unsettling in its simplicity: AI models are becoming remarkably good at persuasion. If an AI can convince people to donate more money than trained professionals today. it could be just as capable of influencing purchasing decisions. political opinions. or public discourse tomorrow.

That’s the upside for productivity. It’s also why transparency and safeguards around AI-generated communication are moving from a technical concern to a public one—because persuasion doesn’t just scale easily. It scales fast.

Claude Opus 4.6 Anthropic AI persuasion fundraising Save the Children University of Oxford preprint study debate performance AI accuracy transparency safeguards