{"id":3254,"date":"2025-11-05T06:57:47","date_gmt":"2025-11-05T06:57:47","guid":{"rendered":"https:\/\/itlawco.com\/?p=3254"},"modified":"2025-11-05T06:57:47","modified_gmt":"2025-11-05T06:57:47","slug":"getty-v-stability-model-weights-not-copies","status":"publish","type":"post","link":"https:\/\/itlawco.com\/fr\/getty-v-stability-model-weights-not-copies\/","title":{"rendered":"Getty v Stability: Model ontology out, measurable reproduction in"},"content":{"rendered":"
<\/div><\/div><\/div>
\n\t
\n\t\t
\n\t\t\t
\n\t\t\t\t\n
\n\t

Almost every superficial commentary framed Getty v Stability as a referendum on whether AI is \u201callowed\u201d to train on copyrighted works. That was never the legal question actually decided. The real question was narrower and more structural:<\/p>\n

\n

Is a trained model itself a \u201ccopy\u201d of the copyrighted works used in training?<\/p>\n<\/blockquote>\n

This is the hinge question because if the answer had been \u201cyes\u201d, the entire open-weights ecosystem becomes legally radioactive. But the Court\u2019s answer was no<\/strong>.<\/p>\n

The Court rejected the \u201cmodel as stored copy\u201d ontology<\/h2>\n

The Court\u2019s logic is grounded in the actual architecture of latent diffusion, not moral sentiment. Weights do not embed pixel frames. Instead, weights encode parameterised statistical mapping \u2014 not expression. Therefore, a trained model is not a storage substrate.<\/p>\n

\n

This alignment\u2014legal definition of \u201ccopy\u201d with technical representational form\u2014is not just a legal victory, it is the future structure of AI IP reasoning.<\/p>\n<\/blockquote>\n

Copyright liability becomes a question of behaviour, not existence<\/h2>\n

Because the Court anchored \u201ccopy\u201d to reproduction<\/em>, compliance in this domain shifts from hypotheticals to evidence. The legal trigger is not<\/em> that a model was trained on copyrighted works. The legal trigger is: can this system output copyrighted content? <\/strong>That is now measurable.<\/p>\n

We have stepped into a world where architecture \u2192 observability \u2192 liability is the chain.<\/p>\n

\n

Compliance becomes an engineering discipline.<\/p>\n<\/blockquote>\n

Trademark findings exist, but they are surface-bound<\/h2>\n

The trademark components are real but tiny, version-specific, and incidence-dependent.<\/p>\n

Key insight: Risk sits at the interfaces <\/strong>not in the checkpoint file. This preserves governability.<\/p>\n

Commercial impact: Existential risk has been defused<\/h2>\n

By refusing to collapse statistical parameterisation into \u201cstored expressive content,\u201d the Court carved out a new implicit category: non-expressive computational artefacts.<\/strong><\/p>\n

This preserves:<\/p>\n