Tax Training – 29/08/2025 – Adrian Cartland: Stability AI, Compression v Vectorisation, and Copyright

This week we move from the Anthropic case to Andersen et al v Stability AI Ltd in the United States, a case that directly raises whether AI training involves simple “vectorisation” (statistical abstraction) or “compression” (storing compact but recognisable copies). This distinction matters for copyright law, and the pleadings and court order make the tension clear. In Anthropic it was assumed by both parties that vectorisation involves compression, and we noted that this was a subject that we would come back to.

In the Second Amended Complaint, the artists allege that Stable Diffusion does not merely store abstract vectors. Instead, they say:

“Stable Diffusion contains compressed copies of the training images, which are stored at and as digital information within the model. These compressed copies can, with sufficient decompression steps, be turned back into recognisable versions of the original training images.” (FAC, §X, 2024)

Judge Orrick, in his earlier order, highlighted the inconsistency between describing Stable Diffusion as statistical processing (vectorisation) and also as containing compressed copies:

“The Complaint repeatedly describes Stable Diffusion as an algorithmic process that uses training images to develop a statistical model, and not as storing or retaining those images. At the same time, plaintiffs allege that Stable Diffusion ‘contains compressed copies of the Training Images.’ Those allegations appear inconsistent, and plaintiffs must clarify their theory.” (Order on Motion to Dismiss, 2023, pp. 11–12)

Turning to Australia, the Copyright Act defines “reproduce” in terms of “material form,” which “includes any form (whether visible or not) of storage of the work or a substantial part of the work” (s 10(1), s 31(1)(a)(i)). If compression stores a form that could be decompressed back into the original, that arguably is a reproduction in material form. Vectorisation, if it truly strips expression down to abstract statistics, might not fall within that definition. The High Court in Autodesk v Dyason (1992) treated the copying of functional elements of a program as infringement, showing how the line between expression and function can be blurred.

Discussion Points for the Session:

How the distinction between vectorisation and compression maps onto Australian copyright’s definition of “reproduce in material form.”
The importance of “material form” and “substantial part” (ss 10, 14, 31) to analysing AI training.
The relevance of Autodesk v Dyason as an Australian precedent where functional copying still counted as reproduction.
Whether Australian law, lacking a broad “fair use,” would treat training as infringing if compression = reproduction.

Please see below link to case materials which is assumed reading in order to participate in the discussion:

Discussion led by Adrian Cartland.

Redirect to tax training webpage