Generated images are cheap, plentiful and arrive already labelled. The research on feeding them back into image models is less comfortable, and it explains why camera-made photographs with a known origin keep their place.
Train a generative model on the output of another generative model, repeat the process, and the results get worse. That finding, published in Nature in July 2024 under the title "AI models collapse when trained on recursively generated data", is the main reason the argument between synthetic data and real photographs has not ended in favour of the cheaper option.
This article is general information, not legal advice. Checked on 18 November 2026.
What counts as synthetic data for image models?
The term covers two different things, and they behave differently.
Rendered or simulated images come from 3D software and game engines. A scene is built, lit and photographed by a virtual camera. Because the scene is constructed, every label is known exactly: the position of each object, its outline, its distance from the lens. This kind of data has long been used for tasks where labelling real photographs by hand is slow or impossible.
Generated images come from a generative model, typically a diffusion model prompted with text. They are fast and almost free to produce, but they are a model's summary of the photographs it was trained on. Their labels are whatever the prompt said, which may not be what the picture shows.
Captions can be synthetic too. OpenAI's DALL-E 3 paper describes training on a mix of 95% machine-written captions and 5% original ones. That concerns the text paired with real images and is a separate matter, discussed in captioning images for machine learning.
Where does synthetic data work?
It has real uses, and the research record is not one-sided.
- As a supplement. A 2023 paper from Google Research, "Synthetic Data from Diffusion Models Improves ImageNet Classification", reported that adding generated samples to the real ImageNet training set significantly improved classification accuracy. The real data stayed in; the synthetic data was added to it.
- Where labels must be exact. Rendered scenes give measurements that cannot be read off a photograph after the fact.
- For uncommon or hazardous cases. A simulator can produce a thousand examples of a situation nobody would want to stage.
- Where real data would intrude on people. Generated faces and figures are sometimes used to avoid processing photographs of identifiable individuals, though that depends on the generator not reproducing real ones.
Where does it fail?
Three problems recur.
Recursion. Shumailov and colleagues found that training on model-generated content can cause "irreversible defects in the resulting models". The less common parts of the original distribution vanish first, and successive generations narrow towards a small range of outputs. A separate 2023 study, "Self-Consuming Generative Models Go MAD", reached a matching conclusion for image models: "without enough fresh real data in each generation" of the loop, either the quality or the diversity of outputs progressively falls.
Inherited limits. A generator cannot show what it never learned. If its training data was thin on a subject, a period or a material, its synthetic versions of that subject will be thin too, or confidently wrong. Generated data redistributes what a model already knows. It does not add observations of the world.
No ground truth. To measure whether a model renders glass, skin or metal correctly, the comparison has to be a photograph of real glass, skin or metal. A generated benchmark measures agreement with another model.
These findings come with a caveat of their own. The collapse studies describe loops in which generated data replaces real data. They do not show that a modest, well-managed share of synthetic material ruins a model. What they establish is that fresh real data is the ingredient that cannot be substituted.
Does synthetic data avoid the copyright question?
It moves it. A generated image is the product of a model with its own training history, so a dataset of generated images carries a question about that earlier training.
Decided. Disclosure rules treat synthetic data as a source to be described, not as an exemption. The European Commission's template for the public summary that general-purpose model providers must publish under Article 53 of the AI Act lists synthetic data alongside public datasets, private datasets, scraped data and user data. Those obligations have applied since 2 August 2025.
Pending. We are not aware of any court decision on whether training on the outputs of a model can attract liability arising from how that model was trained. In the UK, the government's report of 18 March 2026 proposed removing copyright protection for computer-generated works that have no human author, which would affect the status of generated images themselves; that is a proposal, not law.
There is also a practical lesson about contamination. Bloomberg reported in April 2024 that about 5% of the images used to train Adobe's Firefly were AI-generated, having reached the training set through Adobe's own stock library, to which contributors had uploaded them. Adobe said all stock submissions go through moderation. The episode shows how generated material can enter a dataset that is described as licensed, simply because a contributor library accepted it.
Synthetic and licensed photographs compared
| Question | Generated images | Licensed photographs |
|---|---|---|
| Cost per image | Very low | Negotiated |
| Volume | Effectively unlimited | Finite |
| Labels | From the prompt; may not match the picture | From captions and records; quality varies |
| New information about the world | None beyond the generator's training | Yes: each frame is an observation |
| Uncommon subjects and periods | Weak where the generator is weak | As good as the archive |
| Use as evaluation ground truth | Unsuitable | Suitable |
| Origin that can be documented | Depends on the generator's own history | Yes, where chain of title is shown |
| Risk of recursive degradation | Present if used in place of real data | Not applicable |
Neither column wins outright. The sensible reading of the research is that synthetic data extends a foundation of real images and cannot stand in for one.
What do camera-made photographs supply that generators cannot?
Specificity. A studio photograph of a cut diamond records how that stone behaved under that lighting, through a real lens, on a real emulsion or sensor. A street scene records what was actually in the shop windows. A generator asked for either produces a plausible average. For work on materials and products, discussed in why product and still-life photography matters to multimodal models, the difference between observed and averaged is the whole point.
Photographs made before generative image models existed have a further property: they cannot contain generated content. For film originals that is self-evident. As the web fills with generated pictures, a dated, camera-made collection becomes useful as a reference set, on the same reasoning as the collapse papers. This is one of the questions that provenance records answer, described in image provenance for AI. For individual files, the Content Credentials checker shows whether a C2PA manifest is present and whether it declares AI generation, though a missing manifest proves nothing either way.
Where an archive like this one fits
The Robert Harper Photograph Library is the work of one photographer over approximately thirty-five years: 95,487 unique photographs by inventory count, of which 52,684 are analogue, principally 6 x 7 colour transparencies. That film material is camera-made by its nature, and more than half of the archive is still on film and has yet to be scanned. The dataset card sets out what exists as files today and what does not. The collection is not large enough to train a foundation model and is not presented as one; its plausible uses are fine-tuning, domain work on still life and product imagery, and evaluation, which a buyer can test on a sample.
The archive is owned by 825 Ventures Ltd, the owner and rights holder, and licences are negotiated directly with the single rights holder, subject to rights review. Teams weighing real data against synthetic can find the details on the AI training data licensing page.
Frequently asked questions
Is synthetic data as good as real images for training AI?
It depends on the job. Published research shows generated images can improve results when mixed with real ones, and rendered scenes are valuable where exact labels are needed. Studies also show that models trained repeatedly on generated output degrade without fresh real data. Synthetic data is best understood as a supplement to real photographs, not a replacement for them.
What is model collapse?
Model collapse is the name given by Shumailov and colleagues, in a paper published in Nature in July 2024, to the degradation that occurs when models are trained on content produced by earlier models. The less common parts of the original data distribution disappear first, and later generations converge on a narrow range of outputs.
Does using synthetic data avoid copyright problems?
Not automatically. A generated image comes from a model that was itself trained on something, so the question of origin moves back one step. Whether any liability attaches to training on a model's outputs has not, as far as we are aware, been settled by a court. Developers are still asked to document synthetic sources. This is general information, not legal advice.
Why do AI developers still license real photographs?
Real photographs record things generators have not seen or render poorly: specific materials, period objects, uncommon subjects and the way actual lenses and film behave. They also provide a ground truth for evaluating a model, which generated images cannot, and a documented origin that a developer can describe to customers and regulators.
How can I tell whether a dataset contains AI-generated images?
There is no reliable universal test. Documentation is the main defence: who made each image, when, and with what. Material made before generative image models existed is camera-made by definition. Content Credentials can record that a file was generated, where the tool that made it added them, but their absence proves nothing.
Sources
- Shumailov et al., AI models collapse when trained on recursively generated data, Nature 631, 755-759 (2024)
- Alemohammad et al., Self-Consuming Generative Models Go MAD (arXiv:2307.01850)
- Azizi et al., Synthetic Data from Diffusion Models Improves ImageNet Classification (arXiv:2304.08466)
- The Decoder: Adobe trained its AI image generator on Midjourney images (12 April 2024, citing Bloomberg)
- European Commission: Q&A on the training content summary template
- Regulation (EU) 2024/1689 (AI Act), Article 53
- Reed Smith: UK copyright and AI report, the opt-out is dead
- Betker et al., Improving Image Generation with Better Captions (OpenAI)
This article is general information, not legal advice. It reflects the position as we understood it on the date shown and may be out of date. Published by Archive.London. Corrections are welcome under our editorial policy.

