For the past few years, businesses and individuals using generative AI to draft contracts, marketing copy, articles, and correspondence have operated under a comfortable assumption: no one can really tell what came from a person and what came from a machine; and even if they could, ownership of the result belongs to whoever typed the prompt. Both assumptions are wrong.
The Copyright Office Has Answered the Ownership Question
The U.S. Copyright Office and the courts have taken a consistent position: content generated by artificial intelligence is not protected by copyright simply because a person wrote the prompt that produced it. Copyright protection is reserved for human authorship. If you type your own content, you own it. If an AI model generates the content, that language falls into the public domain—even if you spent considerable time crafting the prompt that led to it.
Many real-world examples fall somewhere in between. A business owner might draft an outline use an AI tool to convert the outline to expanded form, and then manually edit the result. A marketing team might write the concept and let AI handle the phrasing. In these mixed cases, courts and the Copyright Office require a fact-intensive analysis: which portions reflect genuine human creative choices, and which portions are the LLM’s output? Only the human-authored elements are protectable. This distinction may matter if there is value in preventing others from using the content. A company that assumes its AI-assisted materials are fully protected may discover, when a competitor copies them, that much of what it thought it owned was never protectable in the first place.
The Watermark Problem Makes This Harder to Paper Over
It is common for our clients to ask us how anyone will ever know which content is created by AI. Until recently, my response was that registering a copyright requires the claimant to disclose the author to the Copyright Office and certify by signature that the statements are true and enforcement will likely require testifying under penalty of perjury. Those alone should be powerful incentives to tell the truth.
Anthropic, the maker of Claude, has begun a new process that adds another incentive. Claude is now embedding invisible watermarks directly into text generated by its newer models—a mark that is imperceptible to readers, does not alter the wording, and travels with the text when it is copied, pasted, or forwarded. The change is driven by EU transparency regulations. Because the marking happens at the LLM level, it applies globally and across every product built on top of that LLM. Other major AI providers, including Google, Meta, Microsoft, and OpenAI, have made comparable commitments. Generated images carry similar embedded provenance data.
The watermark is broader than most people expect. It does not attach only to text the model invents from scratch — it can also attach to text the model merely processes, such as a paragraph you wrote yourself that the LLM cleaned up, translated, or condensed. Heavy editing and rewriting can weaken or remove the signal, and very short passages typically carry too little text to be reliably detected, but the safe assumption for any substantial passage used largely as generated is that its origin can, in principle, be identified.
Detection tools are not yet publicly available, and a detected mark does not, by itself, prove that a human did not meaningfully contribute to a document—it is a provenance signal, not a verdict. But the technology is coming, and a reasonable assumption going forward is that AI-assisted authorship will be readily verifiable by a client, a court, a business partner, or a competitor.
What This Means for Your Business
The combination of these two developments—an ownership rule that discounts AI-generated content, and a technical mechanism that can reveal how content was produced—has practical consequences worth considering if your business model relies on monetizing content or preventing unauthorized uses of content. Businesses that rely on original content for competitive advantage should understand which of their materials are genuinely protectable and which are not. It is vital to review contracts and vendor agreements for representations about content ownership and originality, adopt clear internal practices for how AI tools are used and disclosed in creating work product intended to be proprietary, and consider consulting with legal counsel well-versed in technology and intellectual property protection.
