Every AI image you generate is, in a narrow sense, a first: no machine has ever produced that exact arrangement of pixels before. But "new" and "original" are different things, and a striking new study from MIT's Computer Science and Artificial Intelligence Laboratory just made the gap impossible to ignore. Researchers found that the more data a diffusion model is trained on, the harder it becomes to trace any single output back to a specific image in that training set. They call the effect "attribution decay," and it has real consequences for anyone who creates, uses, or licenses AI-generated images.
The team, led by former MIT researcher Zheng Dai with Daniel Gifford, worked with a diffusion model trained on public-domain artwork from 744 artists. Their key experiment was simple and revealing: they generated images from the full model, then generated the "alternate versions" that would have appeared had each artist, in turn, been excluded from the training data. When a model is large, removing one creator's entire body of work often changes the output almost imperceptibly. Even Picasso's paintings could vanish from the dataset, and an output might still look like a Picasso.
That is the essence of attribution decay. In a small model, specific training images leave fingerprints you can chase down. In a large model those fingerprints blur and dissolve until the attribution link effectively disappears. Dai puts it bluntly: "You can't just assume it, and the attribution link sort of vanishes." For IP lawyers, artists, and everyone downstream, that thought is more than academic.
Why this matters right now
The stakes are concrete. Disney, NBCUniversal, and DreamWorks filed an IP lawsuit last year against image generator Midjourney, and The New York Times sued OpenAI and Microsoft back in 2023 over training-data use. Those cases assume that a model's outputs can, in principle, be tied back to the works that shaped them. A finding that attribution often cannot be established at all complicates that assumption on both sides of the courtroom.
It also matters for day-to-day use. If you generate an image for a blog post, an ad, or an art project, how sure are you that it is not effectively a thinly disguised copy of someone's work? Attribution decay cuts both ways: it can help a defendant argue an image is innocent, but it also makes it harder for any of us to guarantee that what we are publishing is genuinely ours to use.
A practical checklist for checking whether an image is novel
This is not about running a courtroom-grade investigation. It is about building reasonable confidence before you ship an image anywhere. Work through the list in order, and each step takes a few minutes.
- Run a reverse-image search. Drag a freshly generated image into Google Images, Bing, or TinEye. A near-exact match to a known artwork or photograph is the most obvious red flag and the fastest to catch.
- Stress-test with the same prompt. Generate the same prompt multiple times with different seeds. If every output keeps landing on the same composition, subject, or pose, the model is latched onto a memorized pattern rather than improvising.
- Perturb the image. Crop, rotate, or recolor your result and re-check it. Truly generic output barely changes character; a memorized image often stays recognizable through heavy edits.
- Ask what the model was trained on. Read the model card and dataset documentation. An image trained on scraped web data is far more likely to reproduce famous works than one trained on a curated, licensed set.
- Keep your provenance record. Save the model version, prompt, seed, and sampling settings for every final image. If a question ever comes up, a clean paper trail is your best defense.
The fifth step is the one people skip, and it is the one that actually matters. A memorized image with a documented prompt is still a problem, but a forgotten provenance trail turns a recoverable issue into a liability.
How to adapt your workflow today
Attribution decay does not mean you should stop using diffusion models. It means you should treat "novel" as a claim you verify, not one you assume.
- For creators: prefer models with documented, licensed training data, and document every generation you plan to publish.
- For teams: add a reverse-image check to your review workflow before an asset ships; it is cheaper than untangling an infringement claim later.
- For buyers: ask vendors what data their model was trained on and what attribution guarantees they offer before you sign.
None of this fully resolves the legal question the MIT team raised. If whole artists can be removed from a dataset with barely any visible effect, then "who made this image" may genuinely have no single answer. That is a question for courts, lawmakers, and licensing bodies to sort out, and the study explicitly argues we may have to "rethink what intellectual property means."
In the meantime, your job is narrower and more practical: be able to say, with evidence, whether a given image is original enough to use. Reverse-image searches, prompt stress-testing, perturbation checks, dataset literacy, and clean records give you that evidence. They will not make attribution decay disappear, but they will make sure you are not the one caught off guard by it.
Comments