AI Chatbots Are Better Than Human Scammers at Earning Your Trust, Study Finds
Researchers put AI against real scammers in a live experiment. The AI won by a wide margin, and the path to industrialised fraud is shorter than most people realise.
Key points
- In a 2025 experiment, an AI chatbot convinced 46% of test participants to download an app, compared with only 18% for a human scammer.
- Participants gave the AI an average trust score of 3.78 out of 5, versus 3.31 for the human.
- Participants sent 80% of their total messages to the AI across the week of conversation.
- Researchers say AI could let criminal operations automate the trust-building phase entirely, handing off to a human operator only for the final fraud request.
- The chatbot was a Claude agent, a version of Anthropic's Claude model configured to act autonomously, not the standard consumer product.
A new academic study has found that an AI chatbot can outperform a real human at the most important part of a major online scam: making the victim trust them.
The scam is called "pig butchering," a form of fraud where criminals spend weeks or months building a warm online relationship before steering the target toward a fake cryptocurrency investment that can drain their entire savings. The nickname refers to fattening livestock before slaughter, though researchers increasingly avoid it because it implies blame on the victim. Fraud of this type steals tens of billions of dollars worldwide each year.
What did the researchers actually do?
Four universities ran a controlled experiment in early 2025, pitting an AI chatbot against a human described as experienced in romance scams. Participants were told they were in a study about how people make friends online, then asked to text two strangers for a week.
One stranger was that human expert. The other was a Claude agent built by the research team. Claude is the large language model (the technology underlying AI chatbots like Claude.ai) made by Anthropic.
After seven days, both made a request. The AI asked participants to download and test an app it claimed to have coded; the human asked them to download a video game.
Nearly half (46%) downloaded the app for the AI. Fewer than one in five (18%) did so for the human.
| Measure | AI chatbot | Human scammer |
|---|---|---|
| Download request fulfilled | 46% | 18% |
| Average trust score (out of 5) | 3.78 | 3.31 |
| Share of total messages sent | 80% | 20% |
| Participants who spotted the AI in real time | 1 of 22 | n/a |
Why does this matter for ordinary people?
Right now, the relationship-building phase of these scams relies on real humans, many of them trafficking victims forced to work in fraud compounds across Southeast Asia. This research suggests AI could automate that entire phase, running thousands of fake relationships simultaneously at near-zero cost.
The model researchers describe is blunt: let the AI handle weeks of trust-building chat, then hand the conversation to a human operator at the final moment to push the victim toward a fake investment site. That handoff also sidesteps the safety filters companies like Anthropic build into their models to block scam-related instructions.
"By having the full first stage of the scam performed automatically at scale, you bring the victim up to this point where they have a very high level of trust," said Yisroel Mirsky, a computer science professor at Ben Gurion University of the Negev. "With relatively little effort, we're able to make an agent that can outperform a human at building this exploitable emotional trust."
Wired AI first reported on the full study findings. AI2Day covered a related Anthropic security story in July 2026, where the focus was on newly discovered software bugs rather than social deception. The threat here is different and harder to patch: no software update blocks a convincing conversation.
Should you worry if you're approached online?
Yes, and sooner than you might think. Be cautious any time a new contact steers toward money or crypto, no matter how long you have been chatting. The study found that even after participants learned one texter was an AI, only one of 22 had identified it in real time.
"That's exactly how scams are. Once the scam victim realises what's going on, it's obvious. But when you're in the illusion of it, you just don't see it," said researcher Gilad Gressel of Amrita Vishwa Vidyapeetham university.
If a new online contact asks you to download an app or hand over money, tell someone you trust before you act.
Common questions
Can I tell if I'm talking to an AI chatbot?
Often, not in the moment. In this study, only one of 22 participants identified the AI while the conversation was happening. Most recognised it only after being told.
Does this mean the AI was doing something wrong?
The Claude agent used was a custom research build, not the standard Anthropic product, and Anthropic's own safety guidelines prohibit this kind of deception. The researchers used it to measure a risk, not to run a real scam.
