News · 2026-08-20

Nobody is quite sure who owns what an AI makes

The single most repeated claim about AI and copyright right now -- that regulators have ruled AI-generated work cannot be owned -- is not what any of the underlying documents say. Three separate texts are being collapsed into one headline: a non-binding European Parliament resolution, a European Commission guidance FAQ, and a US Copyright Office report. Read individually, they converge on a much narrower and more consistent position. The test is human authorship, on both continents, and it has been for a long time.

Key facts

• The EU instrument is a non-legislative European Parliament resolution adopted March 10, 2026 -- a position statement, not law and not a ruling.
• The European Commission's IP Helpdesk FAQ, updated February 2026, says AI-assisted output can be protected where human authorship predominates.
• The US Copyright Office's Part 2 copyrightability report says purely machine-determined expression is not copyrightable, prompts alone are insufficient, but AI assistance does not automatically bar protection.
• Both positions are administrative or advisory, not court holdings.

Take the European side first. A European Parliament resolution is a formal statement of Parliament's view. It can shape future legislation and it carries political weight, but it does not itself change what any court applies. The March 2026 resolution says content fully generated by AI should not be protected by copyright. That is a real position and a narrow one, and the word doing the work is "fully."

The European Commission's own guidance is more operationally useful and rather different in tone. Its FAQ says protection turns on sufficient human intellectual contribution, that AI-assisted output can still be protected when human authorship predominates, and that ownership may depend on the degree of human intervention, on national copyright law within each member state, and on the platform's terms of service -- and that in some cases the output may belong to nobody at all.

The US position is structurally the same argument in different language. The Copyright Office holds that generative AI output can be protected only where a human author determined sufficient expressive elements. Prompts alone do not clear that bar. But AI assistance does not automatically disqualify a work, and human selection, arrangement and modification remain protectable.

An analogy: photography settled this a century and a half ago. A camera does the mechanical work of making the image, and nobody concluded photographs were unownable. The law located the authorship in the human choices -- what to point at, when to press, how to frame and develop. The current question is only whether typing a prompt is more like choosing a subject or more like pressing a button someone else aimed.

So the accurate one-sentence summary is: the test is human authorship. Not "AI output is unownable," which overstates it in a way that will get someone bad advice.

Why this matters commercially is more interesting than the legal fine print. If pure machine output is hard to own, the defensible asset stops being the artifact. It moves to the things copyright never covered anyway: contract terms, licensing of training inputs, curation and editorial judgment, warranties about where the material came from, indemnity if a claim arrives, and provenance you can demonstrate. That reframes the licensing deals labs have been signing with publishers -- increasingly those are about training rights and downstream restrictions rather than about who owns the output.

It also clarifies why the provenance conversation keeps going in circles. Provenance is actually three different problems that get discussed as one: backward traceability (what was this made from), forward marking (can this be identified as synthetic later), and legal ownership (who holds rights in it). All three are currently weak, they are weak for unrelated reasons, and a fix for one does nothing for the others. We saw the forward-marking half of this when a tool that strips SynthID and C2PA marks passed 4,900 stars.

The honest caveat is that administrative guidance is not settled law. The US Copyright Office states registration policy; courts decide cases, and the human-contribution threshold has not been comprehensively tested. In the EU, national implementations already differ, so "the EU position" is a simplification even when you get the instrument right. A ruling in either jurisdiction could move the line in either direction within a year.

Related reading on this site: content provenance and watermarking, training data attribution, and many AI-generated images cannot be traced to any training example.

Primary source, verified: read the paper →

Ground Truth. (2026, August 20). Nobody is quite sure who owns what an AI makes. Ground Truth. https://groundtruth.day/news/nobody-is-quite-sure-who-owns-what-an-ai-makes.html

@misc{groundtruth:nobody-is-quite-sure-who-owns-what-an-ai-makes,
  title  = {Nobody is quite sure who owns what an AI makes},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/nobody-is-quite-sure-who-owns-what-an-ai-makes.html}
}

Topics: policy · copyright · regulation · european-union · legal