Forum

Attribution Decay: Why Paying Artists Per Output May Never Work

Rippled wet sand on a beach, the ridges catching the light in long parallel lines

One of the most hopeful ideas in the debate over AI and art is that compensation could follow influence. If a generated image owes something to a particular artist’s work, the reasoning goes, that artist should share in the value it creates. Research published by MIT’s Computer Science and Artificial Intelligence Laboratory on 18 August suggests that this idea runs into a hard technical wall.

The researchers describe a phenomenon they call attribution decay. The more data a generative image model is trained on, the less any individual training example matters to a given output. At large scales, they report, you can often remove any single image from the training data, every image by a given artist, or every photograph of a given person, and the model’s output does not change appreciably.

To test this, the team built what it calls a diffusion ensemble, a model made of many smaller components, each trained on a different slice of the data. The design makes it possible to delete large groups of training examples exactly and efficiently, and then measure what changes. Often, the answer was very little.

What the finding means

It is important to read the result carefully. It does not say that artists’ work is irrelevant to AI models. Collectively, the creative work of millions of people is exactly what gives these systems their abilities. What the study suggests is that, at scale, no single contributor is individually decisive for a typical output.

That has uncomfortable implications for several proposed remedies:

  • Per-output royalties based on attribution become difficult to calculate when no individual source can be identified.
  • Removing an artist’s work after training may not remove their apparent influence, because similar qualities are learned from many other examples.
  • Proving that a specific work was used, by looking at outputs alone, may often be impossible.

Where compensation can work instead

If attribution at the level of the output is fragile, attention shifts to the input. Licensing at the point of training, collective agreements that pay creators as a class, and consent before data is collected all avoid the problem of tracing influence after the fact. The settlements and licensing deals of the past year, from music to stock photography, already lean in this direction.

There is also a design lesson for tools. Systems that let users deliberately reference a specific artist’s style create a direct and traceable relationship that diffuse training does not. That is precisely why prompts naming living artists deserve stricter rules than the broad statistical learning the MIT team studied.

What artists can take from this

For individual artists, the study is sobering but useful. It suggests that the most effective protections are collective and preventive rather than individual and after the fact.

  • Support licensing and consent frameworks that operate before training, not only takedown processes after it.
  • Be cautious about services that promise to detect your work inside a model’s outputs.
  • Focus on what remains distinctly traceable: your name, your recognizable characters, your specific works and your direct relationship with clients and audiences.
  • Keep records of where your work has been published and licensed, which matters for collective schemes.

A note for juries and entrants

The research also has a quieter implication for competitions. If influence inside large models is diffuse, originality has to be judged in the work and the process, not by searching for sources. The questions on the International AI Award Rules page, about prompts, models, techniques and ethical considerations, are an attempt to see the human choices that shaped an entry. Those choices are where authorship can actually be found.

The takeaway

The MIT findings do not settle the ethical debate about AI and creative work. They do change its practical terms. If no single image matters, then fairness for artists has to be built in at the scale where their work does matter: collectively, at the input, with consent. Systems that try to trace and pay for individual influence after the fact may be chasing something that, mathematically, is not there.