Ethics, Copyright & Responsible Use

Image models are trained (imgai1-2) on billions of images scraped from across the internet, a large share of which are under copyright, gathered without explicit licensing from the individual rights holders. The central legal question — does training a model on copyrighted images constitute infringement, or is it transformative fair use? — is genuinely being litigated right now, not settled either direction.

Getty Images v. Stability AI is a real, notable case specifically because Getty's own complaint pointed to outputs that reproduced a recognizable, garbled version of Getty's own watermark — direct, visible evidence that specific training images had been memorized closely enough to leave a trace in generated output, not merely "influenced" the model in some diffuse statistical sense. Andersen v. Stability AI is a separate class action brought by a group of working artists raising the same underlying training-data question from a different angle. Both are genuinely unresolved as of this writing — this chapter states the real question being litigated rather than asserting a confident answer the law itself hasn't reached.

Prompting in the style of [named living artist] (imgai1-3's own Style category) is real, common, documented practice. It raises a genuinely different question from Issue 1 above, and the two shouldn't be collapsed together.

The legal question: under U.S. copyright law, a specific work is protected — a particular painting, a particular photograph — but a general style (a recognizable way of using color, brushwork, composition) generally is not. Mimicking a living artist's style, narrowly, is not the same legal category as reproducing one of their specific copyrighted works.

The ethical question is separate, and real regardless of the legal answer: a working artist's distinctive style is often their own economic livelihood and reputation — the thing clients specifically hire them for. Generating unlimited, uncompensated, unconsented content that competes directly with that artist's own commissioned work, using their own name as a literal prompt term, causes a real, documented economic harm even in cases where no specific copyrighted work was reproduced. Several tools, including Midjourney, have restricted or removed the ability to invoke specific living artists' names by name as a matter of policy — a real, documented response to exactly this concern, independent of how the unsettled legal question in Issue 1 eventually resolves.

This is a different category from both issues above — not about training data or artistic style, but about generating a realistic, identifiable image of a specific real person without their consent. This is the clearest-cut ethical case in this chapter, with the least genuine ambiguity: non-consensual explicit imagery (a well-documented, serious harm, with dedicated legislation emerging specifically to address it in multiple jurisdictions), fabricated images of public figures placed in fabricated situations for political disinformation, and more mundane identity misuse all fall here.

Most major tools now maintain real content policies restricting the generation of photorealistic images of real, named individuals, with genuinely varying enforcement effectiveness across tools and over time. Unlike Issues 1 and 2, there's no real live legal or ethical debate over whether this category of harm is real — the open questions here are almost entirely about detection, enforcement, and legislative response, not about whether the underlying concern is legitimate.

A bare, unspecified prompt like a doctor or a CEO has, across multiple tools, been well-documented to default toward particular demographics far more consistently than real-world demographics for those roles would suggest. This isn't a value the model invented from nothing — per imgai1-2's own mechanism, and directly extending imgai1-3's own explanation of why vague prompts produce generic, averaged results, an unspecified prompt's embedding sits in a broad region shaped by whatever demographic patterns were statistically dominant in the captioned training images associated with that term. If historical stock photography and web imagery skewed a particular way for a given role, the model's own statistical average reflects that skew mechanically, whether or not anyone building the model intended it.

This mechanical explanation doesn't absolve model builders of responsibility — a company choosing what data to train on, and whether to intervene on documented bias afterward, is still making real choices with real consequences. Some companies have made deliberate interventions to diversify default outputs for certain prompts, with mixed and sometimes controversial results when those interventions have been applied inconsistently or without enough care for the actual prompt's own context.