An image of an archeologist adventurer who wears a hat and uses a bullwhip

Original Article ↗ Hacker News Discussion ↗

Where Copyright Liability Should Sit

  • Two main views:

    • Infringement happens when outputs are used commercially, so responsibility should rest with the human user or the company selling infringing outputs, not with “math” or training itself.
    • Others argue the service provider is already exploiting copyrighted works commercially by charging for access, so companies like OpenAI should share liability, not just end‑users.

  • Comparisons to:

    • Torrent sites vs. AI: torrents redistribute exact copies, models store abstract representations, but both can be used to undermine copyright.
    • Commissioned fanart: a human being paid to draw a famous character for private use is at least technically infringing; AI is seen as automating that process at scale.

Copying, Creativity, and Overfitting

  • Many commenters say the behavior clearly shows “regurgitation”: generic prompts reliably yield near‑photographic likenesses of specific characters and actors.
  • Some frame models as lossy storage or compression systems: prompt + seed + model ≈ a kind of low‑bitrate archive of cultural artifacts rather than genuine generation.
  • Others counter that human artists also remix and “copy,” and that novelty typically arises from recombination, not ex nihilo creation. The key difference is scale and automation.

Guardrails, Blocklists, and Corporate Incentives

  • Guardrails are described as crude and inconsistent:

    • Certain franchises or characters are aggressively blocked (e.g., some superheroes, “boy wizard”), while others (Indiana Jones, specific anime styles) pass through.
    • Front‑end keyword filtering is easily bypassed; often the image is generated, then a separate “babysitter” model vetoes it.

  • Hypothesis: major rightsholders may have supplied reference sets or lists to be blocked; smaller studios and individual artists do not get this protection.

Fairness, Scale, and “Theft at Scale”

  • A recurring complaint is asymmetry:

    • Individuals and small shops get DMCA’d or sued for minor uses;
    • AI companies ingest vast datasets, including pirated or unauthorized material, then monetize outputs without compensation or attribution.

  • Several people say this is not just “what fanartists already do” but the industrialization of the same behavior, threatening already‑precarious creative livelihoods.

Homogenization of Culture and Tropes

  • Because models are trained on popularity‑skewed data, vague prompts tend to collapse onto the most prominent pop‑culture instance:
    • “Young wizard” → strongly Harry Potter‑like.
    • “Adventurer archaeologist with hat and bullwhip” → Indiana Jones, often with actor‑level likeness.
  • Concern: positive‑feedback loops where models reinforce a small set of corporate archetypes, making “default” imagery even more uniform over time.

Debates Over the Copyright Regime Itself

  • Large subthread argues copyright and “ownership of content” are legal fictions that now mainly serve big media and censorship, not “progress of science and the useful arts.”
  • Others respond that:
    • Some mechanism to reward creators is necessary;
    • The current system is already heavily biased toward large intermediaries, and AI companies are trying to carve a special exemption for themselves while keeping everyone else bound by strict rules.
  • Proposals span:
    • Shorter terms (e.g., ~20–50 years).
    • Clearer treatment of training as fair use vs. infringement.
    • Focusing enforcement on exploitative commercial use, not small‑scale private copying.

What Users Can and Can’t Expect from These Tools

  • Many note that:

    • If you give an obviously referential prompt, you should now assume you may get a legally risky likeness, even if you didn’t name the IP.
    • Excluding or diversifying prompts (“not X,” specify gender, ethnicity, clothing, style) can push the model away from iconic characters but often leaves strong visual echoes.

  • Overall tone:

    • Enthusiasm for the raw capability (“it’s stealing, but really cool”) is mixed with deep unease about legal exposure, cultural flattening, and the precedent of allowing a handful of corporations to mine all prior work without reciprocal obligations.