Using AI Images Commercially Without a Legal Review

Nobody gets fired for using a stock photo. People do get awkward emails about AI-generated images, and the email usually arrives after the campaign is live.

The risk is not that AI image generation is illegal — it is not. The risk is a specific, narrow set of clauses that differ between tools and that almost nobody reads, because they are in the terms rather than on the pricing page. Three of them matter.

1. Indemnification: who pays if it goes wrong

Indemnification means the vendor covers your legal costs if the output turns out to infringe something. It is the single most valuable clause in this category, and it is rare.

Most AI image tools do not offer it. One major vendor does, and offers it specifically because its training data is restricted to licensed and public-domain material — which is a genuine engineering and procurement decision, not marketing.

What indemnification covers varies:

• Some cover copyright claims arising from the output.
• Some cover only claims arising from the training data, which is the more likely source of a problem.
• Some exclude claims arising from your prompts, which is reasonable and worth knowing.

If your images go into advertising, packaging, or anything a competitor's lawyer might see, indemnification is worth a large premium. If they go into an internal deck, it is close to irrelevant.

The practical question to ask a vendor: do you indemnify output, and does the indemnity cover training-data claims? If the answer is no or unclear, that is your answer.

2. Attribution: the clause that turns free into not-free

"Free with attribution" is not free. It is a licence with a condition, and the condition has a cost in every context where you control the layout:

• A website footer credit for a hero image is a design problem and a trust problem.
• Attribution on a printed product or a social ad is usually impossible.
• Attribution requirements in client work are a contract problem, because the client's brand standards will not accommodate it.

Check whether the tier you are on requires attribution, and whether it requires it per image or per site. Some stock libraries with AI features require attribution on the free download tier and drop it entirely on the paid tier, which makes the paid tier much cheaper than it looks.

Also check the reverse: some tools prohibit attribution-removal even when you paid, because their attribution requirement is part of how they licence the underlying model.

3. Output format: vector or nothing

This is the clause that is not a clause, and it costs more than the other two combined.

Most image generators produce raster output. Raster is fine for a screen and wrong for almost everything else:

• A logo at one size only. Scaling it produces a blur, and a blurry logo in a deck reads as careless.
• Print, where resolution requirements are absolute.
• Anything a printer, a sign-maker, or a merchandise supplier will accept.

The workaround is redrawing in a vector tool, which is a skilled task that takes longer than generating the image did. That step is what turns a $12 subscription into a $400 invoice from a designer.

If your images need to scale, you need vector output, and very few tools provide it. One or two do, and they are worth their price for that single capability. Everything else in the category is solving a problem you do not have.

Test this before subscribing: ask for an SVG. If the tool cannot produce one, decide now whether you will need one later, because retrofitting is expensive.

The model-provenance question

Some tools are trained on licensed and public-domain material. Some are trained on scraped data of contested provenance. Some are trained on user-generated content with an opt-out.

For a commercial user, the practical difference is:

• Licensed/public-domain training supports indemnification and is defensible in a procurement review.
• Scraped training is contested, and while the legal position is unresolved in most jurisdictions, it is a live question rather than a settled one.
• User-content training with opt-out means your prompts and outputs may improve the model. Check whether that is acceptable for your inputs.

If you work for an organisation with a legal or procurement function, they will ask about provenance, and the answer needs to be specific. "It's an AI tool" is not an answer that survives a review.

A short decision procedure

For any AI image going into customer-facing material:

1. Does it need to scale? If yes, shortlist only the tools with true vector export. This eliminates most of the category immediately.
2. Does it need indemnification? If it is advertising, packaging, or anything visible to a competitor, yes — and that means one vendor.
3. Does the tier require attribution? If yes and the placement cannot carry a credit, upgrade or change tool.
4. Where was the model trained, and does your organisation care? Ask the vendor directly and keep the answer.

Four questions, about twenty minutes, and they eliminate almost every wrong choice.

The honest summary

For internal decks, social experiments, and mockups, the cheap tools are fine and the licence questions are academic. For anything a customer sees, the licence questions are the product, and the tool that answers them well costs more for a reason.

The mistake is applying the internal-deck standard to customer-facing material because the tool is the same subscription. It is the same subscription and a different risk profile, and the difference is worth twenty minutes before launch rather than after.

For the per-tool position on each of these, the AI image tools reviews record export format, attribution requirements, and training-data provenance where the vendor states it. Adobe Firefly is the indemnification case; Recraft is the vector-output case; they solve different problems and are not substitutes.