How to Make AI Art Safely

“Safely” covers two questions that get searched as one. There’s protecting yourself — your data, your payment, your account. And there’s not creating something that harms someone else, which is where the actual legal exposure lives. The second is the one people ask about least.

The line that isn’t negotiable

Generating sexual imagery of a real, identifiable person without their consent is prohibited on every legitimate platform and unlawful in a growing number of jurisdictions. Enforcement has moved quickly, and it now reaches individuals who made the images rather than only the services hosting them.

Two things people underestimate. First, “identifiable” is broader than a name or a face — a distinctive combination of context, setting and description can identify someone. Second, it applies to people you know, not just public figures; in practice, most enforcement involves someone the creator knew personally.

Anything involving minors is categorically illegal everywhere, including entirely synthetic material with no real subject. There is no jurisdiction where this is a grey area, no platform where it’s permitted, and no framing that changes it. A service that allows it is not permissive — it’s collecting evidence about its own users.

Uploading a photo: what actually happens to it

Many platforms let you upload a reference image. It’s the feature that most often turns a legitimate use into a problem, because the safeguards vary enormously and none of them are visible from the outside.

Before uploading anything, the questions worth answering are whether the platform states what it does with reference images, whether they’re retained after generation, whether they’re used for training, and whether your own face and the faces of people you know are things you’re prepared to have stored indefinitely by a company in a jurisdiction you haven’t checked.

A platform that answers those clearly is telling you something. A platform whose terms are silent is also telling you something.

Protecting yourself

The practical side, in the order it tends to matter:

• A separate email address. Costs nothing, takes a minute, and keeps adult accounts out of the inbox you use for everything else — including out of any future breach disclosure that arrives there.
• Know the billing descriptor before you pay. The merchant name on your statement is rarely published in advance and matters more to most people than a few dollars of price difference. It’s also how you identify who to contact when a charge won’t stop — the cancellation guide covers that.
• Assume everything is stored. Not as a worst case — as the default. Your conversations are stored because storage is the memory feature, and generated images generally sit in an account gallery indefinitely. What varies, and what to check.
• Check whether deletion is self-service. Before you create anything, not after. A platform where deleting your account requires emailing support has told you how it plans to handle the request.
• Look for a training opt-out. They’re uncommon. Their absence means your prompts and images may improve the platform’s models, which is a different thing from being stored.

Where you generate changes the risk

Running an open model on your own machine keeps everything local — nothing uploaded, nothing retained by anyone else, no account to breach. The trade is that no filter protects you from your own mistakes, and the legal limits above apply exactly as they do everywhere else. Local generation is private, not permitted.

Hosted platforms give you moderation that catches some errors before they become problems, and a company that holds your data. Both statements are true simultaneously, and which one weighs more depends entirely on what you’re worried about.

Account suspensions, and why they arrive without explanation

Platforms rarely explain suspensions, because explaining precisely which rule was triggered teaches people how to work around it. That’s understandable and it makes them impossible to appeal usefully.

What reliably triggers them: uploading photographs of real people, attempts to work around content filters, and prompts that read as involving minors even when nothing of the sort was intended — automated systems are tuned to be far more sensitive than a human reader, and descriptions of youthful appearance are the most common accidental trigger. If you lose an account and can’t think why, that’s usually where to look.

How this shows up in our evaluations

Safety findings sit outside the score, alongside billing discretion, the cancellation path and account deletion. We report what a platform’s terms say about reference images, retention and training, and whether the deletion route works when we use it.

They’re kept unscored on purpose. Folding them into a rating would let a platform earn points for meeting the legal minimum, and it would let a good score hide a bad answer on the one question that matters most to you.

Read the full protocol · How image generation works · Adult generation specifically