Memorisation is not a glitch

Foto: Monique Shaw

In para. 84-92 of his recent opinion in the Like Company case Advocate General Szpunar goes along with the narrative of the AI-industry that ‘memorisation’ is ‘a glitch’, ‘incidental’ and the result of ‘overfitting’. ‘Memorisation’ is the phenomenon that generative AI systems ‘store’ or ‘copy’ works, as they are able to ‘regurgitate’ large chunks of copyrighted works or almost identical copies of such works. According to the Advocate General this phenomenon is something ‘incidental’ that can be mitigated by ‘guardrails’ , and should therefore not be considered a copyright infringement for which the AI-system is liable, as long as the operator takes ‘appropriate measures that can be expected from a reasonably diligent operator in its situation’.

‘Memorisation’ is however not a glitch. There can be little doubt that the super intelligent Generative AI of the near future will be both able to recite, ‘redraw’ or ‘recompose’ all existing copyright protected works verbatim (or almost verbatim) and also produce an unlimited number of variations of such works.

The only way to prevent such ‘regurgitation’ of protected works is to have ‘guardrails’ on a massive scale to filter out all the copies of protected works that will inevitably be produced without such guardrails. Any AI-system is already able to draw a perfect picture of Mickey Mouse, because to an AI Mickey Mouse is just a phenomenon that carries a lot of weight in its training material and is therefore regurgitated perfectly. Unless there are guardrails to prevent it from doing so. And in the near future this will apply not only to Mickey Mouse, but to any copyright protected work.

Dirk Visser will discuss the recent opinion in the Like Company case on 9 November, Monday afternoon during Actualiteiten Auteursrecht, on 11 November, Wednesday during Leiden Law Lunch IE and on 13 November, Friday during the course Artificial Intelligence, Recht en Ethiek.

Dirk Visser

DOSSIERS

Memorisation is not a glitch