The Legal Floor: Human Authorship
AI output ownership in 2026 rests on a rule that has not moved despite intense pressure: in the United States, copyright requires human authorship, and works generated purely by AI are not eligible for registration. The US Copyright Office reaffirmed this across its multi-part guidance, holding that a party who simply enters a prompt — however detailed — cannot claim authorship of the resulting output. What matters is the extent to which a human exercised creative control over the work's expression and actually formed the traditional elements of authorship. This is the legal floor beneath every enterprise AI deliverable, and it sits underneath the commercial terms discussed in the AI contract negotiation deep dive.
The 2026 Supreme Court Position
The position hardened in early 2026. On 2 March 2026 the Supreme Court denied certiorari in the leading autonomous-AI authorship case, declining to consider whether AI alone can create copyrightable works and leaving the Copyright Office and DC Circuit's refusal to register purely AI-generated works in place. For enterprises, the practical effect is certainty: do not assume protectability for anything produced without meaningful human authorship. Separately, bipartisan bills introduced in 2026 would require AI companies to disclose copyrighted works used in training data — a signal that the provenance questions covered in negotiating AI training data licensing are moving toward mandatory disclosure.
AI-Assisted Works That Can Be Protected
The protectable path is AI-assisted creation with documented human control. Where a person selects, arranges, edits and materially shapes the expressive output, the human-authored elements can be registered — but the application must disclose the AI-generated content and explain the human contribution. Vague disclosures or overclaiming human authorship can invalidate the registration. For enterprises producing marketing, software or design at scale, this means building a record of human creative input into the workflow itself, not reconstructing it after the fact.
Treat copyrightability as a process requirement, not a legal afterthought. If protection matters for a deliverable, the human creative control has to be real and documented at the point of creation — the registration disclosure depends on it.
What the Contract Must Say
Copyright law decides protectability against the world; the contract decides ownership as between you and the vendor — and the two are often confused. Many AI vendor agreements reserve broad rights in output or leave ownership ambiguous, while others purport to assign output to the customer without addressing the underlying copyright reality. An enterprise agreement should assign all output IP to the customer, with no vendor-reserved licence to reuse your outputs, and should align with the data-rights and safety terms in AI safety clauses in enterprise contracts. The assignment is necessary but not sufficient — it cannot manufacture copyright the human-authorship rule withholds.
The Indemnity Carve-Out Problem
The most expensive trap is the IP-indemnity carve-out. Only about 33% of AI vendors provide indemnification for third-party IP claims, and a large share of those carve AI-generated output out of the indemnity — meaning if the output infringes someone's rights, the customer defends the claim alone. Negotiate an uncarved IP indemnity that covers output, and place IP infringement outside the general liability cap or under a higher super-cap, consistent with the liability-allocation approach in AI bias liability in vendor contracts. A vendor unwilling to stand behind its output on IP is telling you where it believes the risk sits.
Beyond the US: A Fragmented Picture
The human-authorship rule is a United States position, and multinational enterprises cannot assume it travels. Other jurisdictions take materially different approaches — some recognise a narrower category of computer-generated works with a defined owner, others have yet to rule definitively, and several are legislating on training-data transparency in parallel with the bipartisan US bills introduced in 2026 that would require disclosure of copyrighted works used in training. The result is that the same AI-assisted deliverable may be protectable in one market and not in another, and the infringement risk attached to its inputs varies by where the model was trained and where the output is used.
For an enterprise operating across borders, this fragmentation argues for two things. First, contract for the worst case: assign output IP to the customer and secure an uncarved IP indemnity that is not limited to a single jurisdiction's view of authorship, so the protection holds wherever a claim arises. Second, document human authorship to the highest applicable standard rather than the lowest, because a record sufficient for US registration will generally satisfy the more permissive regimes too. The provenance of training data — increasingly a disclosure obligation rather than a private matter — should be addressed in the same agreement, consistent with the approach in negotiating AI training data licensing.
Practical Steps for Enterprise Buyers
Three steps follow. Map which AI deliverables actually need copyright protection and build documented human authorship into those workflows. Rewrite the contract to assign output IP to you and to remove any vendor reuse licence. And negotiate an uncarved IP indemnity with a super-cap for infringement claims. For the full clause set, download the AI Contract Red Flags brief or request a confidential briefing with our AI practice.
Owning the Risk You Can't Avoid
The 2026 settlement on AI output is clear enough to act on: pure machine output is not protectable, AI-assisted work with documented human authorship can be, and the Supreme Court's 2 March 2026 denial means that floor is stable for the foreseeable future. What remains negotiable — and expensive when ignored — is the contract layer, where only about 33% of vendors indemnify third-party IP claims and many carve output out of even that protection.
The enterprise that treats output ownership as a process plus a contract, rather than an assumption, captures both the copyright it can earn and the indemnity it must negotiate. Build documented human authorship into the workflow, assign output IP in the agreement, and secure an uncarved IP indemnity with a super-cap — the same risk-allocation logic that runs through AI bias liability in vendor contracts and the data terms in negotiating AI training data licensing. Our AI procurement advisory team will review your output-ownership and indemnity terms against the current legal landscape.