An AI painting is a digital image produced by a trained model, most often a diffusion model, that starts from random noise and removes it step by step until the picture matches a text prompt. The model learns from very large collections of captioned images, such as the 5.85 billion image-text pairs in the LAION-5B dataset, and a 2025 UK High Court judgment found that Stable Diffusion itself does not store copies of the images it learned from. Whether that training was lawful remains contested, and courts have so far addressed only parts of the question.
Key takeaways
- Diffusion models generate images by iteratively denoising random noise, guided by a text prompt.
- The 2022 latent diffusion method made this affordable by working in a compressed image space instead of on full-size pixels.
- The UK High Court ruled on November 4, 2025 that Stable Diffusion is not an infringing copy, while finding narrow trademark infringement over watermarks.
- That ruling is from the United Kingdom, so it does not settle the question for Canada or the United States.
- Content Credentials can record how an image was made, but they only help when the label stays attached.
From noise to a picture
The method behind most current image generators traces to a June 2020 paper by Jonathan Ho, Ajay Jain, and Pieter Abbeel, "Denoising Diffusion Probabilistic Models." It showed that a network trained to reverse a gradual noising process can synthesize high-quality images. In generation, the process runs backward: the model starts with a field of random noise and removes a little at each step.
The UK judgment in Getty Images v Stability AI gives a plain account of this. Judge Joanna Smith wrote that Stable Diffusion starts with a random noise image and that the trained network, conditioned by the user's prompt, iteratively removes the noise to create an image consistent with the prompt. She also recorded that running the model does not require the training data and that the model itself does not store training data, although the training strongly shapes how it behaves.
Why it became cheap enough to use
Early diffusion worked directly on pixels, which made training and generation slow and costly. A December 2021 paper by Robin Rombach and colleagues, presented at the CVPR conference in 2022, moved the process into the latent space of a pretrained autoencoder. That is a compressed representation that keeps important detail while dropping bulk. The authors added cross-attention layers so text and other inputs can steer the result, and they reported much lower computing needs than pixel-space diffusion. The authors also released their code publicly.
What the model learned from
Training needs images paired with descriptions. The LAION-5B dataset, described at NeurIPS in 2022 by Christoph Schuhmann and colleagues, contains 5.85 billion CLIP-filtered image-text pairs, of which 2.32 billion have English text. It was assembled from material found on the web, and its creators also published detection scores for watermarks and unsafe content. Getty Images argued that its photos were scraped for training, a claim that shaped the lawsuit described next.
What the Getty ruling decided
Getty dropped several claims before closing submissions: it acknowledged it had no evidence that training happened in the United Kingdom, and the prompts behind its examples of copied outputs had been blocked by Stability AI. What remained was a claim that distributing the model weights amounted to dealing in an infringing copy, plus trademark claims. In the judgment of November 4, 2025 ([2025] EWHC 2863 (Ch)), the court dismissed the copyright claim because a model that never stored or reproduced the works is not an infringing copy. Getty succeeded on trademarks only in a narrow, historic way, over some generated iStock and Getty Images watermarks, which the judge called extremely limited in scope.
The ruling does not decide whether training on copyrighted images is lawful in general. In the United States, the Copyright Office released a pre-publication Part 3 of its report on copyright and AI, on generative AI training, on May 9, 2025, and said a final version would follow. Readers who need an answer for a specific use should consult a lawyer, since the rules differ by country and are still developing.
The pipeline in one table
The table lays out each stage and what the cited sources say about it.
| Stage | What happens | Source detail |
|---|---|---|
| Data collection | Captioned images gathered from the web | LAION-5B: 5.85 billion pairs |
| Training | Network learns to remove noise from images | Ho, Jain, and Abbeel, 2020 |
| Compression | Work happens in a latent space, not on raw pixels | Rombach and colleagues, 2021 to 2022 |
| Generation | Random noise cleaned step by step under a text prompt | Getty judgment, paragraph 6 |
| Labeling | Provenance record attached to the file | Content Credentials, C2PA specification |
Telling an AI painting from other images
Visual inspection is unreliable. A more dependable signal is provenance. Content Credentials, a system whose specification is hosted by the Coalition for Content Provenance and Authenticity (C2PA), attach a record of how a file was made and edited, and a pin shown with the image lets viewers open it. The organization notes that bad actors will keep trying to label synthetic content as authentic, and a missing credential proves nothing by itself, because the information has to survive each upload and edit.
- Ask for the tool and the prompt or edit history, not just the title.
- Check whether a credential or similar record is attached.
- Treat "painted by AI" as shorthand: the output is pixels until someone prints or paints it.
- Check the platform's disclosure rules before selling or entering a contest.
FAQ
Do AI image models store the paintings they learned from?
The UK High Court found that Stable Diffusion does not store or reproduce the works in question. That finding concerned one model on the evidence in that case, not every model.
What is a diffusion model in plain terms?
It is a neural network trained to remove noise from images. To make a new picture, it starts from random noise and cleans it up repeatedly, steered by the prompt.
Is the Getty ruling binding in Canada or the United States?
No. It is a decision of a UK court and the UK Copyright, Designs and Patents Act. Courts and agencies in Canada and the United States apply their own laws.
Can a watermark appear in AI images?
Yes, in some cases. The Getty court found that earlier Stable Diffusion versions could generate images carrying iStock or Getty Images watermarks, in limited instances.
Related reading
- AI art and how it is made
- Where AI and art meet
- Using AI to create and enhance art
- Major AI systems compared
More in AI Basics.
Last verified: October 2026
Sources
- Getty Images v Stability AI, [2025] EWHC 2863 (Ch), Mrs Justice Joanna Smith, High Court of England and Wales, November 4, 2025
- High-Resolution Image Synthesis with Latent Diffusion Models, Rombach et al., CVPR 2022 (arXiv, 2021)
- LAION-5B: An open large-scale dataset for training next generation image-text models, Schuhmann et al., NeurIPS 2022
- Denoising Diffusion Probabilistic Models, Ho, Jain, and Abbeel, arXiv, 2020
- Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication), U.S. Copyright Office, May 9, 2025
- About Content Credentials and About C2PA, Coalition for Content Provenance and Authenticity, 2025