Google and OpenAI AI Can Now Spot Scams in Your Complaint

When you report a scam to your state or city, someone has to read it and figure out what kind of crime it is. That job is getting harder. Federal agencies have shed staff, so local governments are absorbing more cybercrime complaints than ever — often without the money or experts to build fancy AI tools of their own.

A new study from researchers Shufan Chai, Liangliang Sun, and Jessica Staddon offers a cheap workaround. Instead of training a specialized AI model, they simply handed general-purpose chatbots a "codebook" — essentially a checklist of telltale signs drawn from years of cybercrime research. The AI reads a complaint and decides whether it describes an impostor scam (someone pretending to be your bank or the IRS) or identity theft. No coding, no expensive setup.

They tested the approach on two scam categories using five different AI models from Google's Gemini and OpenAI's GPT families. Across repeated runs, the method reliably caught the right cases without flagging too many wrong ones — a rare combination in this kind of work. That consistency matters: a tool that cries wolf wastes investigators' time, and one that misses cases leaves victims unprotected.

The bigger picture is about access. Frontier AI is usually locked behind big budgets and engineering teams. This research suggests a small city office could get meaningful results just by writing a good prompt. The catch? The method was tested on only two crime types and in English, so it may not catch newer or more complex scams — and a human still needs to make the final call.