A generated image that looks like evidence is a claim nobody sourced

Every image on a page says something, and most of what generated heroes say is harmless. A painted bowl of fruit asserts nothing. A photorealistic office, a dashboard with rising lines or a smiling customer holding your product asserts that something exists, happened or was measured. When a model produced that picture, nobody sourced the claim. The fix is not a better prompt. It is choosing a visual style that cannot be read as evidence in the first place.

A photograph is a claim before it is a picture

Readers have a lifetime of training in what kind of image means what. A drawing is an idea; a photograph, a screenshot or a chart is a record. Nobody consciously decides to trust a screenshot; the format does the persuading before the reader has looked at the content.

That is exactly why realistic generation is tempting and exactly why it is dangerous. An image model will happily produce a convincing analytics panel, a warehouse for a logistics client or a team photo. Each borrows the authority of a format whose whole meaning is that it came from the world. We argued in where the engine gets its facts that fluent prose carries no signal of whether a claim was sourced. A realistic image is the same problem with a stronger disguise, because readers check sentences more readily than they check pictures.

The law now draws the same line

Since 2 August 2026, Article 50 of the EU AI Act requires anyone using an AI system professionally to disclose when generated image, audio or video content is a deep fake. The Act's definition is broader than the word suggests. It covers content that resembles existing persons, objects, places, entities or events and would falsely appear authentic or truthful. A generated photo of a factory floor that a reader might take for your factory is closer to that line than most marketing teams assume.

The Act also carries a lighter regime for content that is part of an evidently artistic, creative or fictional work, explained in recital 134. The Commission's guidelines on Article 50, adopted in July, lean the same way from the other side: whether an image would falsely appear authentic depends on context and on what the audience expects. Read together, the practical lesson is that the regulation is not asking whether you used a model. It is asking whether your reader could mistake the output for a record. We are not lawyers, so treat this as the reason to ask one rather than as the answer.

Google points the same direction with less force. Its guidance on generative AI content suggests adding image metadata to give readers context, and for shopping it goes further: Merchant Center requires AI-generated product images to carry the IPTC TrainedAlgorithmicMedia label. A product photo is the clearest case of an image that is meant to be evidence, so it is the case where Google stopped suggesting and started requiring.

Why our heroes look like prints

Every hero on this blog is generated to the same art direction: a three-ink risograph look in cream, a dark brown and one vermillion accent, with no people and no lettering. That is not only taste. A print with visible grain and three flat colours announces itself as an illustration. Nobody looks at it and asks where it was taken.

The no-people rule follows from the same logic. A generated face is the image that most readily reads as a real customer, a real employee or a real testimonial, and none of them would be. The no-lettering rule exists for a humbler reason. Image models still render text badly, and when we once asked for a scene suggesting several writing systems, the model invented a convincing-looking script that belonged to no language at all. Garbled pseudo-text looks like a document and says nothing.

We should concede what this costs. A stylised print cannot show your product, your interface or your team, and a post about a feature would often be better served by a real screenshot than by a metaphor. The style is a deliberate refusal to make the hero do evidentiary work, which means the evidentiary work has to be done somewhere else.

Real evidence has to be real

So split the job in two. Decoration can be generated, as long as it looks like decoration. Anything that is meant to show the reader what something is actually like has to come from the thing itself: an actual capture of the interface, a photograph someone took on site, a chart drawn from data you can name. That material comes from the brief, like factual claims about the client, and when it is missing the engine should leave the slot empty rather than fill it with something plausible.

This also makes review easier. An approver can check in seconds whether a screenshot is real. Nobody can check whether a generated warehouse resembles the client's closely enough to mislead.

The test is whether anyone could ask where it was taken

Before an image goes live, ask whether a reasonable reader could wonder where it was photographed, whose data it shows or who the person is. If they could, the image is making a claim, and it needs either a source or a different style. Usually the honest choice is also the cheaper one: an illustration that admits it is one, next to an argument that does the proving. Your images and your words face the same test, as we argued in our post on how Google treats AI content. What matters is whether they are true and useful, not who or what made them.