Ben Berman

A lecture on the ethics of generative art models, argued from practice rather than from principles, working from the synthetic-data training pipeline built for Monopoly GO! as its case.

Adobe markets Firefly as an ethically trained model, but has never shown how it trains the text encoders. Those encoders are almost certainly CLIP or T5, trained on LAION-2B and C4, which carry the exact problems Firefly claims to have solved. The CLIP paper explains why: on a dataset roughly the size of Adobe Stock, a model reaches 37 percent accuracy on ImageNet, against 71 to 80 percent on LAION’s two billion images. Training on expressly licensed data alone is impracticable rather than principled, which is the substance of OpenAI’s position against the New York Times.

Reasoning in these models comes from purpose-built, hand-curated instruction data rather than from the presence or absence of copyrighted text. One person can invent an architecture a thousand times cheaper to train; no one person can write a hundred thousand expert question-and-answer pairs. A single ethical framework does not travel either: Japan operates under a different legal system and benefits from the technology regardless.

The lecture closes on a stated position. Against training on expressly licensed data alone, in favour of open source for now, and against a standard ethical framework.