AI copyright: separate human authorship from training rights

On 29 January 2025, the US Copyright Office released Part 2 of its AI report, addressing copyrightability of generative-AI outputs. This was a US agency report interpreting existing principles and discussing policy, not a new statute or a judgment resolving every dispute. The Office concluded that sufficient human determination of expressive elements remains central. Its announcement said AI assistance does not itself disqualify a larger human-created work from protection. Swedish readers must not mistake that US administrative analysis for a statement of Swedish or EU law. 1

The Office distinguished human creative selection, arrangement and modification from the mere provision of prompts. That does not mean every AI-assisted project is unprotectable. It means the protected contribution must be identified rather than inferred from time spent interacting with a tool. An editor who writes original passages and arranges material raises a different question from someone who selects a generated image without further creative control. Assess the actual expressive contribution, not the prestige of the tool or the complexity of its prompt interface.

A work’s eligibility for copyright protection does not determine whether creating or distributing it infringes someone else’s rights. Nor does it settle whether training on a dataset was lawful. Part 2 addresses output copyrightability; the report series treats training as a separate subject. Avoid headlines suggesting that the Office either legalized all AI training or prohibited all generated content. A business needs separate reviews for input permissions, generated similarities, authorship and distribution contracts. Those questions may involve different facts, legal doctrines and jurisdictions even within the same project.

A useful editorial record preserves original human drafts, meaningful revisions, asset provenance and the final selection rationale. This is a process recommendation, not a guarantee of registration or litigation success. A prompt log alone may show instructions without showing control over the resulting expression. Retain enough evidence to explain what a person actually created while avoiding unnecessary storage of confidential source material. Connect this to company knowledge, because internal reference collections also need provenance, access rules and a distinction between permission to read and permission to reuse.

A vendor may contractually permit commercial use without promising that every generated element receives copyright protection. Equally, a customer contract may demand exclusivity that a team cannot substantiate for unedited machine-generated material. Identify which assurances concern tool access, third-party claims, confidentiality or ownership of human contributions. Do not replace a legal review with a marketing phrase such as commercially safe. Provider obligations offer a separate EU governance context, while AI security addresses confidential inputs; neither resolves US output copyrightability by itself.

Before publishing an asset, ask who created its expressive elements, which source rights were checked and which country’s law governs the intended transaction. Escalate uncertain high-value uses rather than pretending a general article can clear them. This analysis is published on 25 October 2026 and revisits the report released on 29 January 2025. It does not verify later court rulings, current registration practice or a final outcome for training-data litigation. It is not legal advice, and it does not import US conclusions into Swedish law.