What's the most effective way to validate art as human-made?
Organic or Diffused?
Late last year, Ben Zhao’s research group at the University of Chicago published a paper assessing the ability of human and machine to determine whether art was human-made or AI-generated. Here’s the paper:
Ha et al.: “Organic or Diffused: Can We Distinguish Human Art from AI-generated Images?”.
(You may already be familiar with their work: Glaze, a tool which can be applied to artwork posted on the internet to hinder the training of generative AI models, was created by Ben Zhao and his team).
Key findings
If you have the stomach for academic papers and are interested in art verification, I recommend giving it a read.
If not, here are some key points:
- the authors looked to assess performance of humans and machine at determining the provenance of art
- they looked at different styles of art
- they assessed the performance of different AI-art detection tools
- similarly, they assessed the performance of humans of different art expertise (non-artists, professional artists, and artists expert in AI art detection)
- the most effective way to determine if a piece of art is AI-generated or not is to combine the efforts of human validators and machine classifiers!
The most effective way to detect AI art
That last point is somewhat unexpected. The best performance in detecting whether a piece of art is AI or human-made results from combining both human and AI input. This collaboration appears to lead to more accurate assessments than either group working alone, suggesting that the strengths of one may be compensating for the weakness of the other.
A big thank you to the authors of this study for conducting important research that can help in addressing the challenges posed by AI. Research like this is crucial for protecting the integrity of human art!