When a machine writes the scam

What the model is doing

The systems people call AI here are language models. They are trained on enormous amounts of text: websites, articles, manuals, code, and the kind of letters and emails that already exist in public. Their job is narrow. Given the words so far, guess the next one, then keep going.

Do that fast enough and you get paragraphs that look planned. The model is not thinking about you. It has no file on your tax office or your bank unless a person pastes those details in. It is completing a pattern, which is why the writing can sound official or urgent without anyone meaning it.

Why that is useful for crime

A convincing scam used to need a competent writer, or a cheap script that was easy to spot. A language model is a competent writer on demand. It can produce French that reads like a ministry, English that reads like a bank, and many slightly different versions of the same story so that filters see less repetition.

It does not get tired or need a translator. A short instruction such as “write as customer service, mention a delivery, ask them to sign in” becomes a finished message in seconds. The person aiming it no longer has to be good at language. They have to be willing to send the result.

The useful part for crime is not a machine that breaks into systems by itself. It is a machine that makes the email, the SMS and the call script cheap enough to run all day.

Leaked facts do the real work

A fluent generic message is still easy to ignore. What makes it dangerous is a true detail sitting next to that fluency. Once a stolen file has your name, address, IBAN, school, operator or a contract number, those facts can be dropped into the prompt. The model then writes a letter that recites them as if a real agent had opened your file.

This dossier does not claim that any of the incidents here were carried out with a language model. The same leaked fields that already make ordinary phishing work also make generated phishing work. The machine did not invent your IBAN; a person supplied it. The machine made the paragraph around it sound like the tax office.

Voice clones and fake login pages follow the same logic. A model that has heard or seen enough examples can imitate a tone or a layout. You still decide whether to stay on the line or type a password.

What the model cannot do

A language model is not a master hacker and it is not a person who has logged into your account. It cannot approve a payment. It cannot read the code on your banking app unless you read that code out. Most of what the public will meet is still social engineering: a better-written ask.

Using these tools to break into systems, steal data, or run fraud is a crime. This page is not a guide to doing that. It is a guide to recognising the output. The test is the same as for any other unexpected contact. Did you open the official site yourself? Did anyone ask you to hurry, to pay, or to share a code? Knowing a real detail about you is still not proof of identity, even when the French is perfect.