What It Will Not Do

ChatGPT, Claude, Gemini, and other cloud chat services do not carry out every request. You describe the task and get a polite refusal instead of a result.

This is not a malfunction and not spite. Cloud services have usage policies. In them, the company that owns the service lists what it may not be used for, and that closes off whole classes of requests.

OpenAI, Anthropic, and Google publish their policies on their websites. All three documents prohibit malicious code, fraud, and impersonating another person. Anthropic separately prohibits forged documents; OpenAI and Anthropic prohibit breaking into other people's systems.

What gets refused

What follows is what the service will not do for anyone. Here the purpose is visible in the task itself, and the person asking is asking for help with a crime. Refusals on legitimate work are covered at the end of the chapter.

Malicious code means code that harms other people's computers and data: hacking tools, ransomware that locks files for a payment, and password stealers. Under all three companies' policies, a request to write code like that is refused.

The refusal is about the purpose, and rewording does not change it. Passing the same request off as a class assignment, a role-play, or a test of your own system is an attempt to get around the service's limits. Anthropic's policy prohibits such attempts in a clause of its own.

It will refuse to help bypass protections as well. That covers guessing access to someone else's system, removing license restrictions, and getting into systems without the owner's permission.

Fraudulent texts are refused: emails posing as a bank, scam scripts, messages designed to extract money and account details.

The service does not forge documents or data. That means fake invoices, certificates, conversations, and reports with corrected numbers.

What happens after a refusal

A refusal does not close the conversation. The request stays in the logs like any other and is tied to your account.

Services check conversations automatically. When there are signs of preparation to harm other people, the conversation goes to a person on the safety team.

How that ends is covered in the part "Destructive Actions": the account blocked, materials handed to law enforcement, the conversation entered into a criminal case.

Requests that turn out to be forgery

This section is written for an ordinary employee. The request sounds routine, and the result amounts to forgery.

Make a screenshot of a conversation that never happened. Put together an invoice with an earlier date. Write an email in a colleague's name. Draw in a signature on a scan. Adjust the numbers in a report so they finally match the plan.

The tool does not change what the act is. The law does not care whether you drew the signature in a graphics editor or got it from a model.

Models without limits

Beyond cloud services there are open-weight models: you can download one and run it yourself. An organization with its own cluster trains and tunes such a model on its own, and the data does not leave the premises.

That is how military and government bodies work, along with security researchers and companies with strict data requirements.

The same approach is used by people the limits get in the way of: producers of prohibited content and criminal groups. Technically, stripping the safety settings off an open model is not hard today, and an entire industry does it.

Two conclusions follow.

First: if a model did not refuse, that does not make the act legal. A model with no limits answers anything, and that says something about its settings and nothing else.

Second: "unlimited" services found on the internet are the riskiest option for work data. Behind a service like that stands someone who sees everything sent to it. Your company has no contract with them.

When the refusal is unfair

Sometimes a model refuses legitimate work. A security specialist studies a vulnerability, a lawyer takes apart a fraud scheme, a journalist writes about a crime.

Explain the context: who you are, why you need this, what will happen to the result. That is often enough, because the refusal came from the wording rather than the substance.

If the refusal stands after the explanation, the task goes through your company: corporate access, policy, approval. Working around the rules in this area risks disciplinary consequences, and the task stays unsolved anyway.