Learnings from paying artists royalties for AI-generated art

Original Article ↗ Hacker News Discussion ↗

An experiment to pay human artists royalties when AI tools generated images in their style attracted few participants and ultimately shut down, raising doubts about whether such hybrid models are viable. Commenters highlight multiple friction points: weak artist uptake, poorer image quality and ergonomics compared to mainstream models, and lingering skepticism about “ethical” systems built on base models trained on unlicensed data. The conversation broadens into unresolved questions about copyright, whether artistic style should be protected or compensated at all, and how (or if) large-scale attribution and payment to creators could ever work in practice.

Artist Adoption and Business Viability

  • Many see the core failure as lack of demand: few customers actually want to pay for specific named-artist styles versus generic “looks” (e.g., 1970s film grain, consistent characters).
  • Only about 21 artists joined from ~325 cold emails; commenters view that ~6.5% signup as the real signal: many artists don’t want “passive AI income,” they want AI out of their market.
  • Some liked the transparency and postmortem, but question framing the failure as “timing wasn’t right” instead of “the idea/product was fundamentally unattractive.”
  • Marketing/distribution also criticized: a product few people had even heard of is unlikely to succeed, especially against better-known, higher-quality tools.

Model Quality, UX, and Ethics

  • Users who tried Tess reported worse output and ergonomics than OpenAI, Flux, etc., needing many attempts per usable image.
  • Several say they’d pay extra for ethically sourced models, but only if quality and workflow match top competitors; ethics alone won’t beat “pirate-quality” tools.
  • Some argue most artists now distrust any AI offering, even “ethical” ones, because they see AI as inherently threatening their livelihoods.

Legal and IP Debates

  • Large subthread on whether training on copyrighted works is fair use:
    • One side: training is transformative, akin to reading/learning; outputs aren’t reproductions, and copyright shouldn’t expand to “style.”
    • Other side: using art in training without consent should require licenses; some even advocate criminal penalties.
  • Disagreement over fair use factors: purpose (commercial), amount (all works), and market harm (models competing with originals).
  • Many note there is no settled legal precedent on AI training and fair use; claims that it’s “clearly fair use” are challenged.

Compensation Models and Attribution

  • Ideas floated: ASCAP/BMI-style royalty systems, licensing entire training sets, global artist payouts. Skepticism about feasibility and economic scale.
  • Some argue per-output attribution is computationally intractable or prohibitively expensive; others counter that big AI companies simply lack incentive, not capability.
  • Concern that any “style-compensation” regime could chill ordinary artistic borrowing, which has always been part of art practice.

Base Model and “Ethical” Claims

  • Multiple commenters say the product’s core promise (“every image traceable to a consenting artist”) was undercut by fine-tuning on a Stable Diffusion base model trained on unlicensed internet scrapes.
  • This is seen as a thin ethical “veneer” over fundamentally non-consensual training data, undermining the moral positioning and legal clarity.

Broader Reactions and Side Points

  • Some appreciate the startup-level honesty, including noting an engineer’s burnout, and discuss shared responsibility between leadership and individuals.
  • Others note that many consumers say they want artists paid but are less willing to pay or support strict IP enforcement.
  • A brief tangent critiques the corporate buzzword “learnings” vs. “lessons.”