Do image-generating artificial intelligence (AI) models infringe artists鈥� copyright? With legal action already under way against some of the companies behind these AIs, researchers have now found that the generators can occasionally reproduce some of the specific images used to train them 鈥� a discovery that could play into the upcoming lawsuits.
Popular text-to-image generators like Stability AI鈥檚 Stable Diffusion and OpenAI鈥檚 DALL-E 2 use diffusion models, which work backwards from random noise to produce images similar to those that they have seen before. To work well, these models must be trained on a vast number of images, paired with text captions, which are often taken from the internet without necessarily seeking the owners鈥� permission.
This practice has drawn criticism from some who argue that the models are violating copyright. In the UK, Getty Images because it claims the firm 鈥渋nfringed intellectual property rights鈥� by using Getty鈥檚 images for training. A has been filed in the US by a group of artists for allegedly violating US copyright law.
Advertisement
One possible legal argument for the AI companies to use is that their models aren鈥檛 directly using the copyrighted work, but merely being 鈥渋nspired鈥� in the same way that a human artist would be.
But now, at the Swiss Federal Institute of Technology in Zurich and his colleagues, including researchers at Google, have demonstrated that Stability Diffusion and Google鈥檚 Imagen AI models can almost exactly reproduce images from its training data, suggesting that they are, in some sense, stored within the parameters of the models.
鈥淚t seems to happen extremely rarely, at least in the type of experimental set-up we have, but it鈥檚 definitely the case that, from time to time, you can ask these models to generate an image corresponding to some piece of text and they will generate essentially the exact same image that they were trained on,鈥� says Tram猫r.
To demonstrate this, Tram猫r and his team looked at the training data used in these AI models, identifying images that appeared multiple times and so were more likely to be memorised. They then entered the captions for these images as text prompts for the AIs, using 350,000 prompts for Stable Diffusion and 1000 for Imagen, generating 500 images for each prompt.
When they compared these outputs to the original training data, they found that Stable Diffusion produced 109 images and Imagen 23 that were mathematically similar enough to their training sets to be considered reproductions. Visually, these images were often near identical, with only small artefacts like noise or compression (see above).
, head of artificial intelligence law at legal firm Gowling WLG in London, says the team鈥檚 work is likely to play a role in lawsuits against AI companies. 鈥淚 would expect that the claimants will bring up this research and the defendants will have to deal with it, perhaps trying to show that the research is wrong, or that the extent of memorisation is negligible or that this sort of memorisation is not copyright infringement,鈥� he says.
OpenAI has acknowledged that, , predecessors to DALL-E 2 reproduced images used to train the model, but the company claims to have mitigated this before making the AI public. Stability AI has also removed duplicate images in the newest version of Stability Diffusion. OpenAI and Stability AI didn鈥檛 respond to requests for comment on this new work.
While removing duplicates is a good first line of defence, says Tram猫r, it won鈥檛 eliminate the possibility that these models could still reproduce training data. It is also difficult to achieve, because automated methods may struggle to identify very similar images with only subtle differences.
鈥淚f the model at any time just generated an image from the training set, but slightly rotated it, our current evaluation set-up wouldn鈥檛 be able to find this because we only look for sort of exact pixel level matches,鈥� says Tram猫r.
Part of the problem is that we still don鈥檛 really understand why AI models memorise certain things over others. For example, a model trained on many images of a single cat might then reproduce that specific cat when asked for a cat in general, says Tram猫r, but it is difficult to identify when this would happen.
鈥淭his black box nature is really prevalent in all deep learning models,鈥� says at the University of Surrey, UK. 鈥淣obody understands anything.鈥�
To address this, AI creators must find ways to more closely manage their training data to avoid copyright violations, he says. 鈥淏ut I don鈥檛 think this is something that鈥檚 easily solvable.鈥�
Reference:
