Are AI generative models stealing from artists?
This is an extract from New World Same Humans, a weekly newsletter on trends, technology, and society by David Mattin.
*
This week, a development in the emerging battle between artists and those who own generative AI models. One that offers a way forward on the thorny question of payment for artists whose work is used to train AIs.
Stock photography giant Shutterstock announced a new partnership with OpenAI. The platform will directly integrate OpenAI’s text-to-image generator DALL-E 2, allowing users instantly to create images that they can use themselves or make available to others. Image creators on Shutterstock earn a share of revenue whenever a subscriber to the platform downloads their work.
Crucially, Shutterstock will also launch a Contributor’s Fund, which will pay image creators when their work is sold to OpenAI to train its generative models.
That marks the first attempt by the creators of a popular AI model to address the increasingly acute questions around acknowledgement — and remuneration — of artists whose work informs the model’s output.
Those questions are fast gaining awareness way beyond tech circles. CNN this week ran an an interview with Erin Hanson, an artist whose work was used to train the text-to-image tool Stable Diffusion. That tool can now replicate Hanson’s style so effectively that even the artist herself says of its attempts: ‘oh wow, I would put that on my wall’. Stable Diffusion just raised $101 million at a $1 billion valuation.
And this issue doesn’t apply only to visual arts. The Recording Industry Association of America this week warned that AI music generators, trained on copyrighted work, pose a threat to the livelihood of musicians and songwriters.
⚡ NWSH Take:
OpenAI CEO Sam Altman admits that a huge tranche of Shutterstock data — which the company purchased last year — helped train DALL-E.
Essentially, the creators of those images helped train an AI tool that they’ll now be competing against directly on the platform. It’s right that, in future, Shutterstock contributors are paid when their work is used in this way.
But we’re still a long way from a comprehensive answer. Stable Diffusion can churn out a million and one Erin Hanson works good enough to impress the artist herself. It’s not credible to suggest that this has no impact on Hanson’s ability to sell reproductions of her work. And at the moment, her cut of any potential earnings? It’s zero. We could say the same about thousands of others whose work helped train Stable Diffusion, Midjourney, and DALL-E.
But this doesn’t have to be a bad news story for creators. In fact, quite the opposite.
If some thorny issues — both technical and subjective — can be solved around judging to what extent a particular artist’s work helped inform a particular output of a model, then these models can become a whole new revenue stream for artists.
Imagine a model trained on Picasso’s work, with revenues flowing to the late artist’s estate. Or a music generator trained on the output of Ludovico Einaudi; how many dreamy car ads would be soundtracked by the outputs of such a model, and how many millions of dollars would accrue to Einaudi?
Regular readers already know that I’m obsessed with the generative model revolution and its implications. Artists, meanwhile, deserve to be paid for their work. Stable Diffusion and others should take a lead from OpenAI and Shutterstock’s first draft attempt at an answer. There are deals to be done, here, that work for everyone.
*
You've just read an extract from this week's New World Same Humans, a weekly newsletter on trends, technology, and society by David Mattin.
Also in this week's instalment:
🌎 The UN says that limiting warming to the 1.5C target is now close to impossible.
🗺 A backlash against metaverse hype is gathering steam inside the Tech Industrial Complex.
...and much more!