Invisible AI Patent Risks Your Legal Team Is Probably Missing

Artificial intelligence doesn’t alter the law regarding method infringement. If a company performs every element of a patented method, the origin of the instruction isn’t a defense.

What AI does change is the visibility of risk. The critical decision often occurs before patent counsel is involved: a channel that existing intellectual property processes weren’t designed to catch.

A recent compelling argument said the most consequential AI and IP risks begin before any output is generated, and urged companies building large language models on scientific and technical literature to conduct freedom-to-operate analysis on what those models are trained to replicate.

They are right—from inside the IP function of a company whose scientists aren’t building models but are using them constantly, I want to add the part of the picture that sits at the other end of the prompt.

Where the risk enters: Consider this scenario: The R&D team prepares a product user manual and needs a method for a customer-facing protocol. An employee uses an large language model to obtain a method, which is then included in the manual. Standard legal review occurs, but R&D deems the method routine, so no invention disclosure is filed, and IP legal isn’t involved.

The product ships with the user manual. Six months later, a patentee sends a notice letter. From that point on, the exposure decomposes cleanly. Under Section 271(a), the company’s pre-notice practice of the method is strict liability.Knowledge of the patent is irrelevant.

Under Section 271(b), the post-notice period is where inducement crystallizes. Every customer following the manual is a direct infringer, and the company’s continued distribution of the manual, or its failure to issue prompt corrective communication, is the conduct that grounds the claim.

Why the patent trigger gets missed: The problem isn’t that R&D is refusing to involve IP legal. It’s that an LLM removes the conditions under which that involvement would normally be requested.

Traditional research workflows carry provenance. A journal article has authors, citations, and a trail that elicits questions. A patent search returns assignee names and numbered claims. R&D doesn’t need to be IP-trained to encounter these signals, because the medium delivers them automatically.

An LLM response carries none of this. It arrives without authorship, without citations, and without patent context, reading like a knowledgeable colleague answering a question rather than a document with a trail worth scrutinizing. The researcher’s instinct to flag isn’t overridden—it’s never activated. The signal was never there to be found.

Downstream AI Use

With those signals stripped away, four controls can close the gap:

Define the term. AI-use policies should distinguish executable technical instructions from summaries and background. “LLM-derived method” needs to be a recognizable category. A synthesis route, assay protocol, formulation, manufacturing process, diagnostic workflow, or customer-facing method isn’t a literature summary.

Build a fast escalation path. A brief form or designated IP contact is enough. A process that feels like a full legal stop will be routed around.

Preserve provenance. When an LLM-derived method enters an experimental program or deliverable, the project record should reflect it. “When did the company know?” Shouldn’t depend on email reconstruction. Provenance also matters for inventorship, trade secret hygiene, and later defensive narratives.

Review the vendor agreement. Confidentiality, data-use restrictions, and copyright indemnity don’t protect a company that practices a patented process suggested by a tool. Ask about patent indemnity directly, and let its absence inform procurement, policy, and training.

None of these controls require treating LLMs as inherently risky. Research organizations should use them because they’re valuable for summarization, ideation, drafting, and technical orientation. However, the legal significance of LLM output depends on how the company uses it. Searching isn’t the same as practicing, and brainstorming isn’t the same as instructing customers. Once a method is included in a protocol, it becomes company conduct.

The legally significant act isn’t the model generating text, but the company’s decision to practice the method, scale it, or instruct customers to use it.

Model builders should consider what their systems are trained to do. Model users should ask employees how they use the answers they receive. Both ends of the prompt create risk, but the user side may be the one that faces litigation first.

This article does not necessarily reflect the opinion of Bloomberg Industry Group, Inc., the publisher of Bloomberg Law, Bloomberg Tax, and Bloomberg Government, or its owners.

Author Information

Gregory Kline is senior director, transactions, and IP counsel at Thermo Fisher Scientific.

Interested in writing? Review our author guidelines, and submit pitches to Insights@bloombergindustry.com.