Tess's Royalty Experiment: Noble Fail or AI Ethics Wake-Up Call?
Tess promised to fix AI art's dirty secret: ripping off artists without a dime. Instead, it exposed a brutal truth—ethics alone won't cut it in the cutthroat world of image gen.
Kapwing launched Tess in 2024 as the first AI image generator paying royalties to artists whose styles powered its models. Each model? Fine-tuned on just 10-20 licensed works from one artist, with a 50/50 revenue split from $20/month subs. They seeded it with 25 artists via $300-$4K advances, hyping passive income and dashboards for usage tracking. By mid-2024, payouts hit $15,500+. CEO Julia Enthoven positioned it as a rebellion against 'large AI companies' hoovering art sans permission, tapping into 70% public sentiment (per 2023 Verge survey) demanding compensation.
<> 'Tess serves the same mission, allowing artists to reclaim their work from large AI companies who copy it without credit, revenue sharing, or permission.' — Julia Enthoven/>
Sounds revolutionary, right? Wrong. After 20 months, Tess generated a measly $12,172 in revenue—a $7K net loss ignoring dev time. They shut it down. Only 1 in 4 artists (5/21) even used their free sub for personal work, citing abysmal quality. Users needed dozens of retries per image, lagging behind OpenAI's DALL-E or Flux. Hacker News roasted it: artists crave style protection over pennies-per-gen, with one quipping a $50K flat fee + per-image would flip 99%.
As developers, this is gold. Tess's opt-in model dodged lawsuits by blocking unlicensed styles, using tight Contributor Agreements for 'clean copyright chains.' But it screamed trade-offs:
- Tiny datasets limit diversity: 10-20 images per artist means niche, inconsistent outputs. Devs fine-tuning? Prioritize massive base models plus ethical tops.
- Quality kills adoption: No one pays for retries when free giants deliver first-try magic. Build for fidelity or bust.
- IP walls everywhere: Kapwing banned output scraping for AI training—smart, but it stifles open-source dreams. Local inference? Viable workaround, but cloud 'copying' debates rage.
Opinion: Tess wasn't a flop; it was a proof-of-concept grenade. It proved royalties work (payouts happened!), but scaling demands big bucks upfront and OpenAI-level polish. Big Tech won't budge—lawsuits from 10K+ authors (Atwood, Patterson) fizzle under 'fair use' shields. Yet, with public backing, devs can win by baking ethics in: revenue-share APIs, verifiable attribution, private fine-tunes. Kapwing's pivot back to video AI tools shows resilience—lesson for us all.
Don't chase 'ethical AI' as a gimmick. Make it perform, pay artists fat stacks, and watch the market explode. Tess lit the fuse; who's building the rocket?
