The disagreement, laid out without a verdict

Why this lesson refuses to conclude

The legal status of training on copyrighted material, and of what comes out, is being decided right now in several countries at once, on different statutes, with rulings that have already gone in different directions on different facts.

A course that told you the answer would be wrong somewhere and out of date everywhere. So this lesson does something more useful and less satisfying: it lays out the structure of the disagreement so that you can read the next headline properly, and it gives no legal advice, because it cannot.

The two positions, stated as their proponents state them

Training is permitted. Learning statistical relationships from lawfully accessible material is not reproducing expression. A model holds patterns, not copies. Copyright protects particular expression, not facts, style or technique. Humans learn from copyrighted works without licensing them, and a training process that produces a system rather than a copy is transformative in the sense the doctrine is designed for. Requiring permission for every work would make the technology buildable only by those who already own large corpora.

Training requires permission. Building the corpus involves making copies, and copying is the act copyright governs, whatever happens afterwards. The analogy to human learning is rhetorical rather than legal: a person reading a book is not a commercial reproduction at scale. Outputs can substitute for the works they were trained on, which goes to market harm. And where the corpus was obtained from unauthorised sources, the acquisition is a separate wrong from the training.

Both of these are argued seriously by serious people. If one of them seems obviously correct to you, you are probably holding an assumption the other side does not grant.

Where jurisdictions actually differ

United States. No text-and-data-mining exception; the question runs through fair use, which is a four-factor analysis decided case by case. That structure means rulings can and do diverge on facts. Decisions since 2024 have distinguished between the act of training and the means by which the corpus was acquired, treating those as separable questions. Several major cases are unresolved and appeals are pending.

European Union. The 2019 Copyright in the Digital Single Market Directive contains TDM exceptions. Article 3 covers research organisations. Article 4 is broader but lets rightsholders reserve their rights in a machine-readable way: an opt-out, whose practical implementation is itself contested. Separately, the AI Act requires providers of general-purpose models to publish a summary of training content.

United Kingdom. A narrow TDM exception limited to non-commercial research. A broader exception was proposed in 2022 and withdrawn in 2023 after opposition from rightsholders; consultations have continued without settled legislation.

Japan. Article 30-4 of its copyright act is comparatively permissive for use in machine analysis, subject to limits where the use unreasonably prejudices the rightsholder.

India. No TDM exception. Fair dealing is an enumerated list of purposes rather than a flexible standard, which makes the analysis different in kind from the US one. The question is open.

The practical consequence: “is this legal” has no global answer, and a company operating in four countries may face four analyses of the same activity.

The separate question about output

Whether your generated code is protectable is a different matter from whether the training was lawful, and people merge them constantly.

The United States Copyright Office has taken the position that material produced by a machine without human authorship is not registrable, while human-authored contributions to a work containing such material can be. Other offices have reached related conclusions by different routes.

Why an engineer should care: if a competitor copies a product whose code is largely machine-generated, the question of what you own is live. This affects businesses whose defensibility rests on code rather than on data, operations or customers, and it is one of several reasons that code was rarely the moat it was assumed to be.

What to do while it is unresolved

The controls are the ones from the previous lesson, and they are all things you would want anyway.

Keep records. Enable the filters and document that you did. Scan releases. Prefer vendors offering indemnities and read the conditions rather than the announcement. Follow your organisation’s existing software licensing process, which exists and has an owner.

And apply one planning rule: do not build a business whose viability depends on a particular outcome in a case that has not been decided. If a ruling in either direction would end your product, that is a risk to name on the risk register, not a question to resolve by picking the answer you prefer.

What this lesson is not

It is not legal advice, and nothing in it is a substitute for a lawyer in your jurisdiction who knows your facts. Two organisations doing superficially identical things can get different answers, because the answer turns on details (what was used, where, under which contract, for what purpose) that a course cannot know.

Anyone who tells you this is settled is describing their preference. The useful skill is holding an unresolved question without either freezing or pretending it resolved in your favour.