OpenAI just pulled the curtain back on a series of sophisticated online fraud networks it banned, originating primarily in Cambodia, Myanmar, and Nigeria. These aren’t lone-wolf scammers but organized outfits using ChatGPT to industrialize the age-old “pig butchering” playbook. In a new case study, the company details how these groups deployed its models for translation, generating biographies for fake investment gurus, and even managing the mundane office logistics of a scam compound, like desk assignments and dorm allocations. One audacious operator in Myanmar apparently used ChatGPT to research the criminal penalties for getting caught.
The scam playbook, which OpenAI calls the “ping, zing, sting” method, starts with a cold outreach—often a social media ad or a wrong-number text. The “zing” is the emotional hook, promising wild returns on a crypto platform or faking panic over a financial loss. The “sting” is the money grab. AI fits into this as an efficiency scaler, not a strategy shift. The teams used ChatGPT to churn out content for fake trading websites, translate messages to target victims across multiple countries, and even craft detailed personas for non-existent “investment experts.”
But the most striking finding isn’t the criminals’ sophistication. It’s the counterpunch. OpenAI estimates that people are using ChatGPT to identify and avoid scams up to three times more often than the model is being used to commit them. The report cites millions of help-seeking interactions per month, often where a user pastes a suspicious message into the chat and gets a definitive warning in return. An OpenAI investigator even used it to immediately flag a scam SMS impersonating a TikTok recruiter.
This cat-and-mouse game reveals a clear asymmetry. While the scammers’ use of AI remains largely a blunt instrument for scaling text generation, the defensive use is proving to be more immediately practical. You can automate thousands of fake scripts, but a single person with a screenshot and the right prompt can dismantle the lie in seconds. It’s a rare instance where the defensive utility of a large language model is not just theoretical but measurably outpacing the offensive one, right now.
