What you can actually do with an AI-generated image

You need a picture for a blog post, a slide, a pitch deck, a landing page. About 13 cents an image later you have four of them, and one is good enough [5]. Then the useful part of your brain catches up and asks the question the generate button never answers. Are you allowed to put this on the website, and what exactly is it that you now have.

The honest answer has three separate parts that people keep mashing into one. There is what the vendor’s contract lets you do. There is what copyright law gives you, which is much less. And there is what you have to tell the person looking at the image, which since 2 August 2026 is a legal question in Europe rather than a matter of taste [7]. This guide covers all three for ordinary business use: blog headers, decks, social graphics, mockups, internal documents. It is orientation, not legal advice. If the image is load-bearing, meaning a logo, a package, a character you plan to build a brand on, or anything going into regulated advertising, this guide tells you why to hire someone rather than what to do instead.

The licence and the copyright are two different questions

Ask “can I use this commercially” and OpenAI and Midjourney both answer yes in writing. OpenAI’s terms of use, effective 1 January 2026, say that you “own the Output” and that OpenAI “hereby assign to you all our right, title, and interest, if any, in and to Output” [1]. Midjourney is just as direct: “You own all Assets You create with the Services to the fullest extent possible under applicable law” [4].

Read those two sentences again for the hedges, because the hedges are the whole story. “If any.” “To the fullest extent possible under applicable law.” Both vendors are assigning you whatever rights exist. Neither is promising that any exist. OpenAI adds a second caveat in the same section: “output may not be unique and other users may receive similar output from our Services,” and the assignment “does not extend to other users’ output” [1]. So the picture you generated may already sit on someone else’s landing page, generated from a similar prompt by someone you will never meet, and you have no claim on theirs.

Then there are plan-level traps. Midjourney’s terms state that “If you are a company or any employee of a company with more than $1,000,000 USD a year in revenue, you must be subscribed to a ‘Pro’ or ‘Mega’ plan to own Your Assets” [4]. That clause is written for exactly the reader of this guide: the business that crossed a revenue line while three people kept generating images on the cheaper tier nobody has looked at since signup. Ownership in that case is not a philosophical question. It is a subscription setting.

Now the second question. Do you own it the way you own a photograph you took. In the United States the answer is mostly no. The Copyright Office concluded in January 2025 that “prompts alone do not provide sufficient human control to make users of an AI system the authors of the output,” and that “based on the functioning of current generally available technology, prompts do not alone provide sufficient control” [3]. Writing a better prompt does not fix this. Writing 40 of them does not either.

Thin copyright only bites in specific places

For most of what a small business generates, none of the above costs you anything. Nobody is going to steal the watercolour bakery on your blog header, and if they did, you would not sue. The place thin copyright hurts is narrow and predictable: a logo, a brand mark, a recurring character, an illustration style you intend to be known for, or artwork you want to license to somebody else. Those are all assets whose value is your ability to stop other people using them. A machine-generated image gives you very little of that ability, because the part a stranger would copy is the part the Copyright Office says your prompt did not make you the author of [3].

The Office did leave a real door open, and it is the one worth using. Human authors “are entitled to copyright in their works of authorship that are perceptible in AI-generated outputs, as well as the creative selection, coordination, or arrangement of material in the outputs, or creative modifications of the outputs” [3]. In plain terms: the poster you composed is yours as a composition. Your headline, your layout, your colour choices, your edits are protected. The raw picture sitting in the middle of it is not, on its own. That is a workable arrangement for marketing material and a bad one for a trademark.

Do not wait for this to be fixed. The Office’s own conclusion was that “questions of copyrightability and AI can be resolved pursuant to existing law, without the need for legislative change” [3]. There is no rescue scheduled.

Since 2 August 2026 labelling has a legal answer in Europe

Article 50 of the EU AI Act became applicable on 2 August 2026, and it splits the labelling job in two [7].

The first half is your vendor’s problem. Providers of systems that generate synthetic content must ensure the outputs are “marked in a machine-readable format and detectable as artificially generated or manipulated,” using solutions that are “effective, interoperable, robust and reliable as far as this is technically feasible” [7]. This is already happening whether or not you asked for it. Google’s image generation documentation states flatly that “All generated images include a SynthID watermark,” across the current image models it documents, from gemini-3-pro-image down to gemini-3.1-flash-lite-image [6].

The second half is yours. Deployers of an AI system that generates or manipulates deepfake content “shall disclose that the content has been artificially generated or manipulated” [7]. You are the deployer. If the content forms part of “an evidently artistic, creative, satirical, fictional or analogous work or programme,” the obligation shrinks to disclosing the existence of generated content “in an appropriate manner that does not hamper the display or enjoyment of the work” [7]. A stylised illustration on a blog is a long way from a deepfake. A photorealistic image of a person or a place is not.

The practical version, which travels well outside Europe too: a machine-readable mark is not a disclosure to a human being. That is why Article 50 asks for two different things, one from the provider and one from you [7]. If a viewer could reasonably take an image for a photograph of something real, say so in the caption or the alt text. If the image is obviously drawn, you do not need ceremony.

Generated images carry other people’s marks

The most useful recent illustration of what can come out of an image model is Getty Images v Stability AI, decided in the High Court of England and Wales on 4 November 2025 [8]. Getty’s case narrowed sharply before judgment, with the training and output-copyright claims abandoned, and the secondary copyright infringement claim failed. As the judgment records, Getty “do not say that Stable Diffusion is itself a copy of, or that it stores within it any copies of, the Copyright Works” [8]. The model itself was not held to be an infringing article.

What survived was trade mark, and it survived because the model could produce synthetic images bearing the Getty Images and iStock watermarks [8]. The findings were limited, cut down further by which versions of the model Stability had itself released [8], so the case is not a warning about your generator. It is a warning about your file. A generated image can contain a watermark shape, a recognisable logo on a shirt or a shopfront, or text that reads as a real brand at a glance. You are the one publishing it.

So look before you ship. Open the image at 100%, not at thumbnail size in the tool that made it. Check corners and edges for watermark-shaped artefacts. Read every piece of text in the picture, including signage, packaging and screens. Check hands, teeth and repeated objects. That pass catches the failures that make a business look careless.

The generation cost is trivial and the review cost is not

Prices are now low enough that the line item stops mattering. Gemini 3 Pro Image costs $0.134 per image at 1K or 2K resolution and $0.24 at 4K; Gemini 3.1 Flash Image is $0.067 per 1K image; Gemini 3.1 Flash Lite Image is $0.0336 per 1K image. Batch processing halves all of those, taking the Pro model to $0.067 per 1K or 2K image [5]. So 40 images a month at the top tier is about $5.40.

Which means the money is somewhere else. It is in the minutes you spend generating six versions to keep one, opening the keeper at full size, reading the text inside it, deciding whether it implies a fact, and writing the caption. Budget that time explicitly, because the failure mode of cheap images is not overspending. It is shipping 40 of them that nobody checked.

An illustration is not evidence

The last line is the one that matters most and has the least to do with law. Some images persuade because they are pretty. Others persuade because the viewer believes a camera was there. The second kind is a factual claim, and generating it is not a style decision.

That covers your team page, your premises, your product on a table, a before-and-after, a customer’s face next to a testimonial, a result you are promising. It also covers anyone real. OpenAI’s usage policies prohibit “use of someone’s likeness, including their photorealistic image or voice, without their consent in ways that could confuse authenticity” [2]. Do not generate identifiable real people for business use, including the ones you employ.

This is the rule that survives every model release, every terms-of-service revision and every jurisdiction. If the persuasive power of an image comes from “this is real,” then photograph it, licence a real photograph, or drop the image.

What still goes wrong

The terms move under you. Ownership language is set per plan, and plans get renamed, repriced and re-tiered without anyone emailing you about the clause that matters. The Midjourney revenue threshold is the clearest example [4]. Put a calendar reminder to reread the terms of the one tool you actually use, once a year, and check that the account generating your images is the account you think it is.

The law is unsettled and it is local. The Copyright Office position is guidance rather than statute, and it describes United States law [3]. Getty v Stability is a single English judgment on claims that had been cut down before the end of trial [8]. Article 50 is European law, applicable since 2 August 2026, and its deepfake disclosure duty leaves a real amount of judgement to you about what counts [7]. If your images travel across borders, and on the internet they do, design for the strictest rule you are subject to.

Provenance marking is not the protection it sounds like. The mark in the file is machine-readable, which is to say it is aimed at detection tools rather than at the person reading your page, and the disclosure Article 50 asks of a deployer is a duty separate from the marking it asks of the provider [7]. Treat the caption as the disclosure and the watermark as a bonus. The technical half of that job sits with providers of generative systems rather than with deployers like you [7], so if you are the one shipping an image feature inside your own product, work out which role you are in before assuming this guide covers you.