Practice Areas
AI for IP Lawyers: Patents, Trademarks and the Copyright Cases
AI for IP lawyers: patent drafting and prior-art tools, trademark clearance agents, what NYT v OpenAI has decided so far, and which AI output is protectable.
On this page
- The training-data cases and what they have already decided
- Is it protectable? A quick guide
- Patents: drafting, prior-art search and the fabricated-URL problem
- Trade marks and portfolios: clearance, monitoring and the send gate
- Advising clients who generate or train: Article 50, the GPAI Code and model licences
- A setup for an IP team
Twenty million ChatGPT conversations sit in a discovery database because The New York Times sued OpenAI for copyright infringement, and the model makers' own lawyers have been caught by their own models' hallucinations. AI for IP lawyers means two jobs at once: using the tools for patents, marks and portfolios, and advising clients through the cases and the rules that those same tools have generated.
The training-data cases and what they have already decided
The merits of the copyright claims against the model makers are still open as of September 2026. The procedural rulings are not, and they matter to every client who uses a chatbot.
| Case | Court | What has been decided | Why an IP client cares |
|---|---|---|---|
| The New York Times v. Microsoft and OpenAI, now In re OpenAI Copyright Litigation (16 consolidated suits) | S.D.N.Y. | Preservation order of 13 May 2025 (Magistrate Judge Wang) requiring OpenAI to "preserve and segregate all output log data that would otherwise be deleted"; affirmed by Judge Stein after a 26 June 2025 hearing. Production of a 20-million-log de-identified sample ordered November 2025, affirmed 5 January 2026 (plaintiffs had asked for 120 million) | Chatbot logs are discoverable ESI held by a third party; OpenAI told the court it would otherwise have to retain some 60 billion conversations, of which under 0.01% were relevant |
| Concord Music Group v. Anthropic (music publishers' case) | N.D. Cal. | On 15 May 2025 Anthropic's counsel at Latham & Watkins admitted that Claude, asked to format a citation in an expert declaration, kept the link but invented the article's title and authors; Judge van Keulen struck the affected part, noting "a world of a difference between a missed citation and hallucination generated by AI" | The model maker's own defence team was caught by its own model; no vendor's tool is exempt from verification |
Judge Stein's reasoning in the discovery ruling was that users "voluntarily submitted their communications" to OpenAI (National Law Review); Nelson Mullins' summary of the preservation fight, including OpenAI's argument that the order forced it to "disregard legal, contractual, regulatory, and ethical commitments" to users, is here. The Volokh Conspiracy carried the Anthropic declaration episode as it happened (15 May 2025; 26 May 2025). For a client, the two cases together mean that anything typed into a consumer chatbot about an invention, a mark or a licensing dispute may be produced in someone else's litigation and, per United States v. Heppner (S.D.N.Y., February 2026), is not privileged unless counsel directed it. The privilege page sets out the sequence.
Is it protectable? A quick guide
Clients now arrive with a logo, a product description, a code base or a claim set that a model produced, and ask what they own. The answer as of September 2026, at the level that is stable across the main jurisdictions:
| What the client has | Copyright | Patent | Trade mark | What to tell them |
|---|---|---|---|---|
| Text, image or music generated from a short prompt, unedited | No in the US (human authorship required by the Copyright Office and the courts); doubtful in the EU, where protection requires the author's own intellectual creation | n/a | A generated logo can be registered as a mark; registrability depends on distinctiveness, not on who drew it | Register the mark; do not rely on copyright in the artwork |
| Generated draft substantially rewritten, selected and arranged by a person | The human contribution is protectable; the generated remainder is not | n/a | n/a | Keep the drafts and the edit history as evidence of authorship |
| Invention conceived with AI assistance | n/a | An AI cannot be named as inventor (offices and courts in the US, UK and at the EPO have refused it); a human who made the inventive contribution can | n/a | Record who conceived what, and when, in the lab notebook or its equivalent |
| Invention where the model did the conceiving | n/a | Not patentable in the client's name absent a human inventive contribution | n/a | Consider trade-secret protection for the process instead |
| Prompts, fine-tuning data, evaluation sets | Thin or none as such | n/a | n/a | Trade secret, if kept confidential: which means not pasted into a consumer tool whose terms allow retention and training |
| Curated dataset assembled at cost | Possible protection for selection and arrangement; EU database right for substantial investment | n/a | n/a | Licence it; the licence is the asset |
Two cautions travel with the table. First, the human-contribution line is fact-sensitive, and the client who cannot show what the human did will lose the argument; contemporaneous records are the whole case. Second, the trade-secret answer collapses if the confidential material went into a tool that keeps it: consumer ChatGPT trains on conversations unless the user opts out, consumer Gemini keeps activity for eighteen months by default with human review of a subset, and Grok Build uploaded whole repositories including secrets users had told it to ignore before a server-side fix in July 2026. The confidentiality checklist lists the current terms.
Patents: drafting, prior-art search and the fabricated-URL problem
Patent drafting is the IP task with the most dedicated tooling. DeepIP raised a $25m Series B in April 2026 and reports 400-plus firms; Solve Intelligence raised $40m and builds on Claude; Ankar raised $20m (legaltech.com funds flow, April 2026). What these tools do well is the structured, repetitive part: generating claim variants from a disclosure, drafting the description to support the claims, producing consistent reference numerals and figure descriptions, and checking that every claim term has antecedent basis. What they do not do is decide the claim strategy, and a first-turn draft still needs a patent attorney who reads it against the disclosure rather than against the model's summary of the disclosure.
Prior-art search is where the risk sits. The University of Pennsylvania's April 2026 study of search-augmented models and deep-research agents (including Gemini 2.5 Pro Deep Research and OpenAI Deep Research) found 3–13% of cited URLs were fabricated and a further 5–18% did not resolve, and that agents hallucinated more because they cited more (arXiv). A prior-art reference that does not exist is worse than one you missed, because it goes into a freedom-to-operate opinion. The workable division:
Role: Patent search analyst. You will NOT cite specific documents.
Input: the attached invention disclosure (A) and draft independent
claim 1 (B).
Task:
1. Decompose claim 1 into its elements. For each element list 5–8
alternative technical terms an earlier patent or paper might have
used, including terms from adjacent fields.
2. Propose CPC and IPC classifications likely to contain relevant
art, with one sentence on why each is plausible.
3. Draft 10 Boolean search strings combining the terms in 1 and the
classes in 2, ordered from narrow to broad.
4. List the three most likely categories of art that would anticipate
claim 1 (for example: a product manual, a standards document, a
conference paper) and the venues where each would be found.
Do not name any patent, publication, author or product. If you are
tempted to, replace it with "[SEARCH FOR: description]".
The search itself then runs in Espacenet, Google Patents, PatentScope or a commercial tool, and every hit is retrieved and read. Feeding the retrieved references back with "work only with the documents provided" turns the model into a competent first-pass claim-chart builder; the deep research guide explains why a citing agent needs the same treatment as a citing associate. Opus 5 "verifies its own work without being told to" as of July 2026, which helps on the claim chart and does nothing for a fabricated reference.
Trade marks and portfolios: clearance, monitoring and the send gate
Clearance, watching and enforcement are high-volume and rule-shaped, which makes them the natural home for agents. Anthropic's open-source IP plugin (part of the 12 May 2026 Claude for Legal release) ships skills for trademark clearance, freedom-to-operate triage, cease-and-desist letters, DMCA §512(g) counter-notices, open-source licence checks and portfolio renewals, each starting from a cold-start interview that records the firm's positions (github.com/anthropics/claude-for-legal). The design choice worth copying is the send gate: a demand letter or takedown notice is drafted by the agent and cannot leave without a named human confirming it. A cease-and-desist sent to the wrong party, or one that mis-states the registration, is a liability event; the gate is cheaper than the apology.
A clearance and enforcement workflow that respects that logic:
- Intake. The agent takes the proposed mark, goods and services, and target jurisdictions, and generates the search strategy (word elements, phonetic and visual variants, translations, relevant classes).
- Search. Searches run in the office databases and a commercial watch service, not in the model's memory; the agent collates results into a table with the registration number and status for each hit.
- Triage. The agent scores each hit against the firm's recorded positions (identical, similar goods; similar mark, identical goods; and so on) and drafts a risk memo with every conclusion tied to a specific hit.
- Human opinion. The lawyer reads the hits, not the scores, and signs the clearance opinion.
- Watch. A scheduled agent checks new filings and marketplace listings on a fixed cadence and files a report; the regulatory monitoring page covers the scheduling pattern.
- Enforcement. The agent drafts the letter from the template and the facts in the file, tags anything it could not verify, and stops at the gate.
Step 5 raises a point browser agents have made live. The Ninth Circuit in Amazon v Perplexity (4 August 2026) vacated an injunction against Perplexity's Comet browser agent on the reasoning that "It is the user who 'accesses'" the website, with the agent's help, where the agent runs on the user's device. Read for an IP practice: when your agent scrapes a marketplace for counterfeits or a registry for new filings, you are the one accessing it, and the site's terms and any computer-misuse statute apply to you. The agent security guide covers the prompt-injection side of the same problem, which is real: a poisoned product listing can carry instructions your monitoring agent will read as its own.
Advising clients who generate or train: Article 50, the GPAI Code and model licences
Three sets of rules now sit on top of the case law, and IP lawyers are the ones clients ask about them.
The AI Act's obligations for providers of general-purpose AI models have applied since 2 August 2025 and include a copyright policy and a public summary of training content; since 2 August 2026 the Commission's AI Office can demand documentation, evaluate models, order corrective measures and fine up to €15m or 3% of turnover. The voluntary GPAI Code of Practice published on 10 July 2025 was signed by OpenAI, Google, Anthropic, Microsoft, Mistral, Amazon, IBM, Cohere and Aleph Alpha; xAI signed only the safety and security chapter; Meta refused, citing "legal uncertainties" (The Register). Which chapters a client's supplier signed is now a due-diligence question.
Article 50's transparency duties apply from 2 August 2026: chatbots must identify themselves, deepfakes must be labelled, and AI-generated content must carry machine-readable marks, with a grace period to 2 December 2026 for existing systems. More than 180 organisations had signed the voluntary transparency code by August 2026, and Anthropic reportedly added a statistical watermark to Fable 5.1 outputs for this purpose (a secondary source; treat it as reported). For a client producing marketing content or product imagery with generative tools, the marking obligation is a workflow change, not a legal opinion. The AI Act page sets out the timetable.
The third set is licensing, and here IP lawyers have an old skill applied to new objects. A client that fine-tunes or self-hosts an open-weight model is taking a software licence, and the terms differ:
| Model | Licence | Note as of September 2026 |
|---|---|---|
| Mistral Large 3 (December 2025) | Apache 2.0 | EU-domiciled provider |
| DeepSeek V4 (V4-Pro, 13 August 2026) | MIT | Banned on US Department of Defense and Australian government systems |
| Meta Muse Glimmer 30B (10 August 2026) | Apache 2.0 | Muse Spark weights promised but unpublished |
| Harvey Tenet (research preview, 20 August 2026) | Post-trained open-weight legal model | Terms not public in the research |
| Claude for Legal plugins | Apache 2.0 | Skills and agents, not model weights |
The IP plugin's open-source licence check was written for application code; it applies just as well to the model a client is about to build a product on. The open-weight versus closed comparison covers what the choice means for confidentiality.
A setup for an IP team
Split the practice by output. Anything that becomes a filing, an opinion or a letter (claims, clearance opinions, FTO memos, demands) runs through a retrieve-then-verify pipeline with a human gate before it leaves, and the citation protocol applies to references exactly as it does to case law. Anything that stays inside the team (search strategy, claim decomposition, portfolio triage, watch reports, first drafts of the description) can run on an agent with a scheduled cadence, on an enterprise tier with written terms, because the Copyright Office, the patent office and the court will never see it. Then write two client notes you will reuse every week: one on what is protectable when a model did part of the work, with the records to keep, and one on what goes into which tool, with the discovery rulings attached. Module 3 of the AI Lab for Lawyers course covers file handling and image work for exactly the evidence-of-authorship problem above, and Module 4 builds a scheduled watch agent without code.
Frequently asked questions
Can AI-generated work be copyrighted?
Not where the AI did the creative work. The US Copyright Office requires human authorship and courts have upheld that position; in the EU protection turns on whether the work is the author's own intellectual creation. What is protectable is the human contribution: selection, arrangement, substantial editing, and the original text or images fed in. Document the human input at the time it happens, because the question will be asked later with no contemporaneous record.
Are ChatGPT conversations discoverable in litigation?
Yes. In the copyright litigation against OpenAI, Magistrate Judge Wang ordered production of a 20-million-log sample of de-identified ChatGPT conversations in November 2025 and Judge Stein affirmed on 5 January 2026, reasoning that users voluntarily submitted their communications to OpenAI. A preservation order in May 2025 had already required OpenAI to keep output logs that would otherwise be deleted. Treat chatbot logs as ESI held by a third party.
Can I use AI for a prior-art search?
As a first pass to generate search terms, classifications and candidate references, yes; as the search itself, no. A University of Pennsylvania study in April 2026 found search-augmented models and deep-research agents fabricated 3–13% of the URLs they cited and returned a further 5–18% that did not resolve. Every reference must be retrieved from the patent office database or a commercial search tool and read before it goes into an opinion.
What does the EU AI Act require of clients who train models on copyrighted material?
Providers of general-purpose AI models have had to comply since 2 August 2025 with obligations that include a copyright policy and a public summary of training content, and since 2 August 2026 the Commission's AI Office can demand documentation and fine up to 15 million euros or 3% of turnover. Article 50 also requires machine-readable marking of AI-generated content from 2 August 2026, with a grace period to 2 December 2026 for existing systems.
Put this into practice
AI Lab for Lawyers: a live, hands-on course for practising lawyers
4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.
Related reading
Regulation & Ethics
Attorney–Client Privilege and AI: Heppner, Warner, the NYT Logs and Public Chatbots
Consumer chatbot memos are not privileged. US v. Heppner, Warner v. Gilbarco and the 20-million-log OpenAI order, plus rules that keep privilege intact.
10 min read
Tools
Claude for Legal Plugins Explained: The 12 Practice-Area Plugins and 20+ Connectors
Claude for Legal gives paid Claude users 12 open-source practice-area plugins, 150+ skills and 20+ connectors to iManage, NetDocuments and CoCounsel.
8 min read
Guides
Deep Research Tools for Lawyers: Claude, ChatGPT, Gemini and Perplexity Compared
Deep research AI for lawyers compared: what Claude, ChatGPT, Gemini and Perplexity research modes do, where they fail on law, and a verification table.
10 min read
Regulation & Ethics
The EU AI Act for Lawyers and Law Firms: What Applies in 2026
The EU AI Act applies to law firms as deployers. Article 4 literacy and Article 50 transparency are live; high-risk duties are deferred to December 2027.
9 min read
Comparisons
Open-Weight vs Closed Models for Law Firms: Control, Cost and Confidentiality
Open-weight vs closed AI models for law firms: what Latham's own GPUs buy, Harvey Tenet, DeepSeek V4, Mistral, the legal benchmark gap, when self-hosting pays.
10 min read
Use Cases
AI Regulatory Monitoring: Agents That Track Law Changes for You
AI regulatory monitoring works as a scheduled fetch, diff, classify and route workflow on official sources. How to build a weekly digest and cut false alarms.
10 min read