How AI could discourage the creativity it depends on

The internet is a vast library of human-created information and creativity. We upload photos, videos, music, and writing daily. Today, much of this human-generated content is being used to train AI. But what happens if people become less willing to share their work online?

This question was addressed in a study by Alexander Staub (Esade and University of Lausanne), Christian Peukert (University of Lausanne), Florian Abeillon (University of Lausanne), Jérémie Haese (University of Lausanne), and Franziska Kaiser (University of Lausanne), published in CESifo Working Papers. The study, which won the Best Paper Award at the DRUID 2026 conference, examined how AI can change people's willingness to create new content.

If an AI system is not continuously fed a supply of fresh, diverse, human-created material, the quality of future AI output could be affected.

A real-world experiment in AI training

In 2020, Unsplash, one of the largest free stock photography platforms globally, released a dataset of 25,000 images for the purpose of training commercial AI systems. Some photographers’ work was included in the dataset, while others’ work was not.

This created a rare opportunity to observe how creators reacted in response to their work becoming part of AI training. Before the dataset was released, the two groups behaved similarly, and so the researchers compared what changed afterwards.

There were important findings. Photographers whose images were used for AI training became significantly less active on the platform. They were more likely to stop contributing altogether and, even those who remained on the platform, uploaded far fewer images.

Creators changed their behavior

The decline was substantial. On average, affected photographers uploaded around 40% fewer new images than comparable users whose work had not been included in the AI dataset. The researchers also found that they were significantly more likely to leave Unsplash entirely.

The photographers were not asked about their reasons for leaving the platform or uploading fewer images. However, the evidence suggests that awareness of their images being used to train AI influenced their level of engagement with Unsplash.

The reduction in uploads became even stronger after August 2022, when image-generating AI tools such as Stable Diffusion became widely available and public awareness of generative AI increased dramatically.

Economic incentives also appear to play a role. Professional photographers and users with more images included in the training dataset reacted more strongly than casual contributors. Photographers whose images were used to train AI were three times more likely to join Unsplash+, the platform's paid service, which explicitly prevents images from being used for AI training. These findings suggest that creators were responding strategically as they became more aware of AI's commercial potential.

The debate is often about how AI uses existing content. But the rules governing that use may also shape what creators choose to produce in the future.

Why fewer uploads matter for everyone

The implications extend beyond a reduction in the number of uploaded photographs.

AI uses a huge amount of data, but that data needs to include content that is novel, varied, original, and as diverse as possible. If the creative humans that contribute that content become less active – for whatever reason – then the pool of content used to train AI gets smaller and less diverse.

Following the use of photographers’ images to train AI, some creators reduced their activity or left the platform altogether. The mix of contributors changed. The variety of new content declined while repetition increased. Over three years, the researchers estimate that the release of the AI training dataset reduced the variety of new images by around 4% and increased repetition by 2.5%.

Many of the photographers also display their work on Instagram. If creators had simply abandoned Unsplash in favor of another platform, uploads to Instagram would have been expected to increase. But the researchers found no significant change in Instagram activity. This suggests that the behavioral response was specific to Unsplash and its AI training dataset rather than a shift to Instagram.

A policy dilemma with no easy answer

The findings of this study are timely. Currently, governments, publishers, technology companies, and creators are debating how copyrighted material should be used to train AI systems.

The hot topic is how to balance innovation with the rights of the original content creators. Recent licensing agreements between publishers and AI developers, ongoing copyright lawsuits brought by authors and artists, and new transparency requirements under the EU AI Act all attempt to address this issue.

The implications extend beyond the datasets AI can use today. Copyright rules may also influence whether creators continue producing the high-quality, diverse content that AI will need in the future.

So what can we do to ensure AI training in the future is based on diverse content? The authors don’t argue against AI training, nor do they propose any specific copyright model. What is needed is recognition that there is a trade-off: policies that allow maximum access to existing content may be inadvertently affecting whether people continue to create the very content on which future AI depends.

Keeping human and AI creativity alive

Despite its remarkable capabilities, AI still depends on a continuous supply of human-created content. The challenge is no longer only how machines learn from people, but how society can ensure there is a continued flow of original work for AI to learn from.

"Overall, our findings highlight the trade-off between the interests of rightsholders and promoting innovation at the technological frontier," the co-authors say.

If AI is to continue evolving, it will need more than smarter algorithms. It will also depend on policies and incentives that encourage people to keep producing the diverse, high-quality content on which future innovation relies.