Due to a combination of illness, travel commitments, and general availability, we’ve decided to take a short mid-semester break from our weekly tax training sessions. We’ve had strong interest in our upcoming session on Anthropic, Fair Use, and the Nature of Intellectual Property, and we’re looking forward to continuing the discussion when sessions resume on Friday, 1st August.
In this session we examine a recent US copyright decision that raises fundamental questions about the nature of property in the digital age. The case is Bartz v Anthropic, involving the use of millions of books, some purchased and others pirated, in training a large language model (LLM) for Anthropic’s AI system called Claude.
As professionals increasingly rely on generative AI tools in legal, accounting, and business contexts, the question of whether LLMs can lawfully be trained on copyrighted material has become a live legal issue. This case provides one of the clearest judicial explanations to date of how LLMs are trained and how that training process intersects with traditional property rights.
At the heart of the dispute was whether copying books for AI training was a “fair use” under US copyright law. The Court drew a clear distinction between two types of source material:
-
Lawfully purchased books, which Anthropic destructively scanned and converted into digital format for internal use.
-
Millions of pirated books, which Anthropic downloaded and retained in a permanent digital library for potential future use.
The Court allowed the use of purchased books, finding that format-shifting for internal storage and search purposes was transformative. It also ruled that the use of both sets of books for LLM training was a fair use, likening the process to human learning and noting that the final AI outputs did not reproduce the original works.
However, the Court was highly critical of Anthropic’s decision to retain the pirated copies permanently. It emphasised that the company had no property right in those copies, regardless of their eventual use.
Drawing analogies to Australian law, the decision invites comparison with Autodesk v Dyason, where the High Court held that copying a program’s functional behaviour was not an infringement of copyright in the original code. Both cases explore the boundary between copying expression, which engages property rights, and using material as functional input, which does not.
Discussion Points for the Session:
-
How copyright operates as a form of property right, with control over copying, distribution and derivative use.
-
The distinction between lawful acquisition (title) and mere possession of digital content.
-
Whether internal uses like format-shifting or machine learning training amount to a misappropriation of property.
-
The relevance of Autodesk v Dyason in Australian law as a comparator.
-
Broader implications for how tax, research and development incentives and digital asset rules might engage with intellectual property issues arising from AI training.
Please see below link to case materials which is assumed reading in order to participate in the discussion:
Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson v Anthropic PBC
Discussion led by Adrian Cartland.