A new legal analysis published this week dissects what is becoming an increasingly urgent question for software teams using AI coding tools: who legally owns the code that Claude Code writes? The Legal Layer piece walks through Anthropic’s commercial terms, US copyright doctrine, and the GPL contamination risk lurking in any model trained on open-source repositories.
For Australian organisations relying on Claude Code in production — or for law firms advising on AI-related contracts — the answer is not as simple as “the developer who pressed Tab”. Ownership depends on three interacting factors: the licensing tier in use, whether meaningful human creative direction was applied to the output, and the employment status of the person at the keyboard.
What the legal analysis says
The Legal Layer analysis examines US copyright doctrine and concludes that Claude Code outputs sit in unsettled territory. Code accepted verbatim with no human modification may not be copyrightable at all, following the Thaler precedent in which the DC Circuit ruled that works lacking human authorship cannot be registered.
The Supreme Court declined to hear the appeal in March 2026, leaving the rule in place. Code that includes meaningful human direction — review, restructuring, integration with existing modules — may receive copyright protection, but the exact threshold is being tested in cases like Allen v. Perlmutter.
How Anthropic’s EULA actually works
Anthropic’s commercial plans (API and enterprise) explicitly assign outputs to the user and offer IP indemnification, with the company committing to defend customers against copyright infringement claims arising from authorised use. Consumer plans (free and Pro) also assign outputs but with narrower indemnification language.
Critically, neither tier covers downstream GPL violations. If the model’s training data included copyleft code and an output reproduces it verbatim, the licence obligations transfer to the customer’s codebase regardless of the EULA terms. License-scanning tools become a defensive necessity rather than an optional governance step.
The Thaler precedent and what is still being tested
Thaler is the cleanest US precedent so far: works produced by AI without human authorship are uncopyrightable. Allen v. Perlmutter is testing how much human involvement suffices to flip an output back into copyrightable territory. Doe v. GitHub, on appeal in the Ninth Circuit, asks the harder question — whether tools like Copilot reproduce licensed code without proper attribution.
Until those cases resolve, organisations using AI coding tools at scale operate in a known unknown. The legal answer will arrive incrementally over the next 12 to 24 months, by which time most enterprise deployments will already be in production.

What Australian businesses should consider
The Legal Layer analysis is US-focused and does not extend its conclusions to Australian law. The same structural questions apply under the Copyright Act 1968, however, and several principles map directly: works created by employees in the course of employment vest copyright in the employer, contractor IP-assignment clauses commonly extend to side projects, and human authorship requirements are read in line with US doctrine.
The practical advice is consistent across jurisdictions — run licence scans (FOSSA, Snyk, Black Duck), document the human contribution in commit messages and prompt logs, review employment IP clauses before side projects, and use commercial-tier subscriptions for commercial products. Law firms advising clients in this space should track Australian decisions as they emerge — for context on where AI is reshaping legal practice locally, see our overview of AI tools for lawyers in Australia.
Embedding contractual review into document automation workflows is now defensible practice rather than over-engineering, particularly for organisations with regulated clients in finance, health, or government.
A grey zone that will not stay grey
The IP question for AI-generated code is not a permanent grey zone — it is a temporary one, being settled case by case as US litigation works through the courts and as legislators in the EU, UK, and Australia respond. The question for Australian businesses is whether to wait for clarity or build defensible internal practice now.
Document trails, commercial-tier subscriptions, and licence scans cost little. Rebuilding a contaminated codebase costs a lot.
