Questions to answer before you publish AI product imagery
Three questions decide most of the risk: whether the image must be labelled, who owns it, and whether it matches the product. The sources are clear on the first two and thin on the third.
This primer is for the people who approve marketing: the founder shown a hero image of a product nobody photographed, the marketing director who signs the brief, and the general counsel who will be asked whether anyone checked. Its one point is simple. Before the first generated image goes live, decide three things: whether it must be labelled, who owns it, and whether it is true to the product. The answers come from different places, and the questions below show where each one lives. The primer also marks where its sources run out. That gap is real, and a reader who knows where it is can close it before a regulator or a customer finds it first.
- Do advertising rules apply to AI product images at all?
Probably in some form, but this primer cannot tell you which. The rules on product claims and imagery in advertising are set by advertising regulators and codes in each market, and the sources behind this piece do not cover them. Before you brief an agency, find the regulator in every country where the image will run and read what it says about imagery and product claims. Ask the agency the same question in writing and keep the answer on file.
- Does the EU AI Act cover a product image?
The European Commission calls the AI Act the first comprehensive legal framework on AI worldwide, and it sorts AI systems into four risk levels. The Commission says minimal- or no-risk AI carries no AI Act rules, and gives AI video games and spam filters as examples. Its examples do not mention marketing images, so the Act does not answer the question for you. The narrower question is whether a given image counts as a deep fake under the Act's definition, because that is where the labelling duty begins.
- Do we have to label AI-generated product images?
The Commission says certain AI-generated content should be clearly and visibly labelled, and it names deep fakes and text published to inform the public on matters of public interest. It also says providers of generative AI must ensure that AI-generated content is identifiable, and its page gives August 2026 as the date the transparency rules take effect. Whether a product picture of something that was never photographed is a deep fake depends on the Act's definition, which the Commission's summary does not reproduce. Read the definition in the regulation itself and put the question to counsel. Do not treat the absence of a label as a finding that none is needed.
- What if the generated image sits inside a chatbot or an interactive product tool?
The Commission says chatbot users must be told they face a machine, so they can decide on an informed basis. That rule concerns the conversation rather than the picture, but it matters when a generated image appears inside a chat window. If an assistant shows generated images while it sells, disclose the machine as well as the picture. Ask your product team which screens and flows this touches before launch, not after the first complaint.
- Who owns the copyright in an AI-generated product image?
The US Copyright Office's AI initiative examines the scope of copyright in AI-generated works and the use of copyrighted material in training. Its Part 2 report covers whether outputs made with generative AI can be protected by copyright, and the Office's own summary says so. The sources behind this primer do not state the Office's conclusion, so do not assume one. Read Part 2 and ask a copyright lawyer what it means for your brief. In the meantime, record who made each creative choice in the image and what instructions the tool was given.
- Can the tool's training data expose us to a copyright claim?
The Commission says rules for general-purpose AI providers include transparency and copyright provisions, effective August 2025. Those duties sit with the model's provider, not automatically with the company that uses its output. The sources behind this primer do not show whether any particular tool's training data is licensed, disputed or documented. Ask the vendor in writing what it has published about its training data and what protection it offers users. A vague answer is information in itself, and it belongs in the file.
- What if the picture resembles a real person?
The Copyright Office's Part 1 report covers digital replicas and was published on July 31, 2024. The sources here do not summarise its conclusions, so no rule is stated in this primer. As a working policy, treat any resemblance to a real person as a consent question and obtain written permission before publication. That is our judgement, not a rule quoted from a regulator, and counsel should confirm it for each market where the image will run.
- How accurate must a generated product picture be?
The sources behind this primer set no accuracy standard for marketing visuals, so the standard has to come from inside the company. A simple test is to ask whether the picture shows anything the product does not have, including colour, texture, scale, stitching, screen content or packaging. If a buyer would be surprised by the object in their hands, the image is wrong, whatever the tool produced. NIST's Generative AI Profile is meant to help identify risks unique to generative AI, and an accuracy failure belongs on that risk list.
- Is there a framework we can write our own rules against?
The US National Institute of Standards and Technology released its AI Risk Management Framework 1.0 on January 26, 2023, for voluntary use. A companion playbook was published alongside its roadmap, crosswalk and perspectives, and a Generative AI Profile, NIST-AI-600-1, followed on July 26, 2024. NIST says the framework is being revised as part of the White House AI Action Plan, so cite the version you use and check for updates. Treat it as the skeleton of your internal controls rather than as a legal standard that binds you.
- What applies to generated imagery in the UK?
The Competition and Markets Authority states its goal as promoting competition and protecting consumers, with the end goal of economic growth and household prosperity, according to its GOV.UK page. That page lists consumer-rights and business guidance but, in the text available to us, gives no rule on AI imagery. Read the CMA's consumer guidance directly and check whether any published position covers generated images before you run a UK campaign. Do not infer a rule from the CMA's remit, which is broad by design.
- Are the voluntary EU schemes worth joining?
The AI Pact is a voluntary initiative that invites providers and deployers to comply with key AI Act obligations, according to the Commission. The AI Act Service Desk provides information and implementation support across the EU. Treat joining as a commitment you will have to keep, not as a certificate you can display. For a company placing generated content in the EU, the practical question is whether the Service Desk can answer your labelling question faster than your lawyer can, and whether the answer is written down.
- Why not wait until the rules are settled?
The Commission's view is that existing legislation gives some protection but was judged insufficient for AI-specific challenges. Waiting is still a decision, and it usually means the brief gets written by whoever is most hurried that week. A better course is to write your own rule now, with a named owner, a review date and a log of which images were generated and who approved them. You will need to change that rule as the sources change. That is a reason to write it down, not a reason to leave it blank.
- What is the single most useful thing to set up first?
Set up an image log before anything else. For every generated image in use, record the tool, the vendor, the prompt or brief, the approver, the markets where it runs and the date it was checked against the product. Most of the questions above become answerable from that record, and none of them becomes answerable without it. A log also shows a regulator, a court or a customer that the company knew how its pictures were made.
The mistake behind most of these
Most of these questions go wrong for the same reason. The image looks like a photograph of the product, so the team treats it as one. It inherits the approval steps built for a real shoot, the usual assumptions about who the people in it are, and an unspoken belief that someone checked the product against the picture. None of those things happened. A generated image has no studio session, no model release and no sample on the bench. Each of those absences turns into a question that somebody will eventually ask, usually in the worst week of the campaign. The label, the licence and the accuracy check all depend on knowing how the picture was made, and most companies never write that down. The fix is cheap if it is done before the campaign goes live and expensive if it is done after.
This week, write a one-page rule with three headings: disclosure, rights and accuracy. Under each heading, name the person who decides, the regulator, report or guidance you checked and the date you checked it. Start the image log described above and add every generated picture already in use, including those on partner sites and in paid social. Send the rule to counsel for each market you sell in, and ask the agency and the vendor for their positions in writing. Where the sources in this primer run out, the gap is yours to fill, and it should be filled by a named person rather than skipped. Set a review date three months ahead, because the NIST framework is being revised and the Commission's transparency rules are now in force. Then read the rule again as the buyer would, holding the product in one hand and the picture in the other.