Generative AI and Used Car Listings: The Rules That Now Apply

Something changed in the way AI car photography is regulated, and most of the trade has not caught up with it.

Until recently the honest answer to "are AI car photos allowed?" was it depends what the AI changes. That answer is still right. But it is no longer the whole answer, because the question is now partly settled in writing — by a regulator, with a worked example about a car, and with a duty that lands on the dealer rather than the software company.

This article covers what actually applies to a used-car listing photo, what does not, and the questions worth putting to whoever supplies your imaging.

This article is general information, not legal advice. For your own circumstances, take your own advice.

The distinction that now carries legal weight

"AI car photography" is one phrase covering two genuinely different things.

Generative AI synthesises new pixels. A diffusion model or a GAN is asked for an image and it produces one — inventing detail that was never photographed. When it is pointed at a car, it can invent the car: a wheel that is not the wheel, a grille with the wrong number of slats, a badge that does not exist on that trim.

Deterministic computer vision measures and moves pixels that already exist. It finds the edge of the car, separates it from the background, works out where the camera was, and composites the same photographed car onto a different backdrop. Nothing about the vehicle is invented, because nothing about the vehicle is generated.

For years this was an engineering distinction with no commercial consequence — both approaches produced a nicer picture, and the trade quite reasonably called both "AI". That is no longer the case. The distinction is now the hinge on which a transparency obligation turns.

What is actually in force

The EU AI Act's transparency rules are live

The transparency provisions in Article 50 of the EU AI Act (Regulation (EU) 2024/1689) have applied since 2 August 2026. This is not a future deadline to plan around. It is current law.

Two obligations matter to anyone advertising vehicles.

Article 50(2) binds the provider — the software vendor. Where a system generates synthetic image content, the vendor must ensure outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated". You cannot do this for your supplier and your supplier cannot delegate it to you.

Article 50(4) binds the deployer — and the deployer is you. Where the output is a "deep fake", the business publishing it must disclose that the content has been artificially generated or manipulated. Article 50(5) requires that disclosure to be clear, distinguishable, and given no later than the first time someone sees the image.

That second one is the part the trade keeps missing. The compliance burden of generative imagery does not stay with the vendor who sold it to you. A meaningful part of it transfers to the business that publishes the picture.

"Deep fake" is not just faces

The most common objection at this point is that a deep fake means a fabricated video of a person, and a car photo could not possibly qualify.

The Act's own definition in Article 3(60) says otherwise. A deep fake is AI-generated or manipulated content "that resembles existing persons, objects, places, entities or events". The European Commission's implementing guidelines, published 20 July 2026, remove any remaining doubt about what "objects" covers: "realistic, inanimate material items, including buildings, artworks, machinery, consumer goods etc."

A used car is an object. The definition reaches it.

Two further points from the same guidelines are worth knowing. The assessment is objective — no intention to deceive is required for the definition to bite. And a high degree of photorealism makes the classification more likely, which is an uncomfortable finding for a category that sells itself on being indistinguishable from a real photograph.

But the Commission drew the line, using a car

Here the guidelines do something unusually helpful. They give worked examples on both sides, and the example they chose for the compliant side is a car.

Listed under things that do not constitute a deep fake:

"A real product (e.g., a car) shown in an advertisement against an AI-generated background and surrounding environment"

Listed under things that do:

"An AI-generated image of a product in advertisement or packaging that can affect the audience's perception and mislead as to the actual product appearance"

And the guidelines state the trigger for that second example plainly: an image "making the product appear not identical to the real product, more appealing or with improved quality than in real life".

Read those together and the regulator's line is clear, and it is not a line about backgrounds versus no backgrounds. It is a line about whether the product itself is real in the picture. A genuine photograph of a genuine car, placed in a synthetic environment, is expressly fine. A synthesised image of the car is not.

There is one more provision worth knowing, because it is the one the trade press has entirely missed. The guidelines exempt from Article 50(2) content "produced by simple data processing that is not specifically AI-generated" — and the example the Commission gives is a rendered frame. Computer-graphics rendering, of the kind used in games and visual effects for thirty years, is not synthetic content in the Act's sense at all.

The UK route is different, and older

The UK has not adopted the AI Act, and there is no blanket legal requirement in the UK to label an advert as AI-made. The ASA said so directly in May 2025. But this is a much weaker comfort than it sounds, for three reasons.

First, the underlying law already covers it. The Digital Markets, Competition and Consumers Act 2024 has applied to commercial practices since 6 April 2025. A misleading action includes providing false or misleading information about a product — and section 226(3) makes the point that an overall presentation can be deceiving even where every individual piece of information in it is true. The CMA can now impose a penalty itself, by final infringement notice, of up to 10% of global turnover. In its first year of direct enforcement it opened investigations into 14 businesses and imposed £4.7 million in fines.

The government-backed guidance written for car traders specifically is blunter still: misleading information may be given verbally, in writing or visually, and the worked example it gives for the visual route is "the use of pictures of vehicles".

Second, the ASA does not need an AI rule to act. The CAP Code is media-neutral: the same rules apply however an image was made. Rule 3.11 prohibits exaggerating a product's capability, and rule 3.7 requires you to hold evidence for objective claims before you publish. The ASA has been ruling against altered product imagery for years — it upheld a complaint against retouched imagery in 2017, long before generative AI existed, and in June 2024 it ruled against a composited image where a stock photograph had digitally added number plates and was presented as real product output.

Two details dealers routinely get wrong. Your own website listing is inside the ASA's remit — the Code covers marketing on non-paid space you control, not just paid advertising. And responsibility stays with the advertiser even when the ad was produced by an automated platform; it does not transfer to the tool vendor.

Third, if you are accredited, your code may already be stricter than the law. The Motor Ombudsman's Vehicle Sales Code, approved by the Chartered Trading Standards Institute and issued 1 June 2025, requires accredited businesses at clause 2.8 to use images of the actual vehicle when a used vehicle is sold online. No statute required; you signed up to it.

What this looks like when it goes wrong

The clearest publicly documented case is not a dealer, and that matters — but it shows exactly what generative failure looks like on a real listing.

In January 2026 The Autopian examined the photographs on a 1999 Cadillac DeVille consigned to Bring a Trailer. The images showed wheels that were not the car's wheels, a grille with three slats where the model has four, a headlight layout that did not match, tail lights that were not DeVille tail lights, a door frame above the mirror that had simply vanished, and — the detail that gives the game away — cobblestone flooring visible inside the car.

Every one of those is invented detail. No amount of checking the background would have caught them, because the background was not the problem. The car was.

It is worth being precise about why this is the sharp case for used vehicles specifically. A new car is one of many identical units, and a manufacturer's studio image is understood as representative. A used car is a single specific object with a single specific history, and every mark on it is information the buyer is relying on. That is what makes generative imagery structurally riskier here than almost anywhere else it is sold.

The questions worth asking your supplier

None of this makes AI imaging a bad idea. It makes the mechanism worth knowing. Five questions get you a long way:

1. Does your system generate pixels of the vehicle, or only of the environment around it? This is the question the Commission's two examples turn on.
2. If I process the same photo twice, do I get the same picture? A generative model is stochastic by construction — the same input produces different outputs. Determinism is evidence of a pipeline that measures rather than invents.
3. Do you mark your outputs as artificially generated in a machine-readable format? If the system is generative and you sell into the EU, Article 50(2) requires it of the vendor.
4. What exactly is changed in the image, stage by stage? A supplier who cannot answer this precisely is asking you to carry a risk they have not measured.
5. Who did you say carries the disclosure duty? If the answer is anything other than "the business publishing the listing", they have not read Article 50(4).

Where Motuva stands, precisely

We used to run a diffusion model. In September 2026 we replaced it with a pipeline of our own that does not generate images at all, and it is worth setting out exactly what that means rather than asking anyone to take "not generative" on trust.

The car in the picture is the car you photographed. Its pixels are cut from your own photograph and composited. They are not resynthesised, repainted, straightened or improved. The system has no capability to alter the vehicle's bodywork, colour, trim or condition, because nothing in it generates vehicle pixels.

The studio around it is rendered, not generated. It is a 3D scene, path-traced in a rendering engine, at the camera position measured from your photograph. That is computer graphics of the same kind used in film and games — the Commission's "rendered frame".

The same photo produces the same picture. Run it twice and the outputs differ by 0.043 mean levels — effectively identical. There is no seed, no sampling, no variation between runs.

Two things are honestly worth naming, because a vendor who only lists strengths is not being straight with you. The contact shadow beneath the car is produced by a trained neural network, so those shadow pixels are new. It reads only the vehicle's silhouette and never the colour pixels of the car, so it cannot invent anything about the car's appearance — but it is a trained model and we would rather say so than let you discover it. And the number plate is deliberately replaced, which is a change you asked for, control, and can switch off.

That is the whole of it. We would rather publish the mechanism than a slogan, because "we don't alter the car" is a claim any vendor can type, and an architecture that cannot alter the car is a different thing entirely.

The short version

• Article 50 of the EU AI Act has applied since 2 August 2026. It is in force, not coming.
• "Deep fake" includes objects. A car qualifies. Photorealism makes the classification more likely, and no intent to deceive is needed.
• The disclosure duty for deep fakes falls on the dealer publishing the image, not on the software vendor.
• The Commission's own example says a real car against a generated background is not a deep fake. A generated image of the car that misleads about its appearance is.
• The UK has no AI labelling law, but the DMCC Act, the CAP Code and — if you are accredited — the Motor Ombudsman's requirement to use images of the actual vehicle all reach a misleading listing photo already.
• The durable test has not moved: is the car in the picture the car in the yard? What changed is that the way your software answers that question now has legal consequences attached.

Sources: Regulation (EU) 2024/1689, Articles 3(60), 50 and 99; European Commission guidelines on Article 50 transparency obligations, C(2026) 5054 final, 20 July 2026; Digital Markets, Competition and Consumers Act 2024; the CAP Code and ASA rulings; The Motor Ombudsman Vehicle Sales Code, 1 June 2025. This article is general information, not legal advice.

Start free — no card, no demo →

Related reading: Are AI-edited car photos misleading? · The EU AI Act and dealer marketing · The dealer's guide to the DMCC Act 2024