AI for Patent Attorneys: Drafting Help and the Disclosure Risk

Yes, with limits. A patent attorney can use AI for search ideas, summaries, claim-language options and consistency checks, provided an unfiled invention never goes into a tool that is not covered by written confidentiality and no-training terms. The attorney still owns every claim, citation and inventorship decision.

The risk is not that AI writes poor claims; you will catch that in review. The risk is quiet: a disclosure typed into a consumer chat window may be readable by the vendor, stored for weeks, or used to improve a model, and in a first-to-file system with a strict novelty bar nobody can later prove it did not matter. So the order of work is: classify the material, choose the tool, then draft.

Follow me on Instagram@sagnikteaches

Three tiers of material, and what each may touch

Before comparing tools, sort what you handle into tiers. The sort decides the tool, not the other way round. Write the tiers into your office procedures and tell every assistant and paralegal.

Connect on LinkedInSagnik Bhattacharya

Most of the useful AI work in a small practice sits in tier 1 and tier 2, and that is deliberate. Reading published prior art, summarising a cited reference, comparing claim wording to a granted claim and tidying a response to an office action that is already on file are all work where the novel content has already left the client's hands.

Subscribe on YouTube@sagnikteaches

How a disclosure becomes a disclosure

An invention is usually considered novel only if nobody has made it available to the public before the filing date. Some jurisdictions give a grace period for the inventor's own disclosures; many give none. A chat tool is not a publication in any ordinary sense, so the question is narrower: did you hand the content to a third party with no duty of confidence, and could that party reveal or reuse it? I cannot give you a legal answer for your jurisdictions, and I would not trust a tutorial that did. What you can do is make the answer easy by never creating the question.

Four things decide how worried to be about a given tool:

• Training use. Consumer plans of the big assistants let you switch off use of chats for model training in the privacy settings; business plans such as ChatGPT Business and Claude Team do not train on business content by default. Check the setting label on the day you use it, because vendors rename them.
• Retention. Even with training off, chats stay in your history until you delete them, and deleting is not instant everywhere: Anthropic says a deleted Claude chat leaves its back-end storage within 30 days, and longer periods apply to anything flagged for a policy breach. Stored is not public, but stored is also not gone.
• Human review. Vendors may allow staff to read flagged conversations for safety. Read the data-processing terms for the phrase that covers this.
• Contract. A business agreement that states confidentiality and no training is a much stronger position than a consumer toggle. Ask for the document, not the marketing page.

The American Bar Association's generative AI ethics opinion (Formal Opinion 512, 29 July 2024) goes further than most patent attorneys expect. It says a lawyer needs the client's informed consent before putting information about the matter into a self-learning tool, that general boilerplate in an engagement letter is not enough to make that consent informed, and that you may not charge a client for time spent learning a tool you will use regularly for clients. It is guidance, not law, but a regulator or a client's general counsel will recognise it. If your bar association has its own opinion, read that one as well.

A consent paragraph you can adapt

An illustration of the wording to put in front of an inventor or in-house counsel before tier 2 material goes into a tool. Adapt it, and have your own adviser look at it.

We would like to use [tool name], an AI assistant, on your matter for these tasks only:
summarising published prior art, suggesting alternative wording for claims
that we then review, and checking our drafts for internal consistency.

What goes in: [the filed application text / no unpublished technical detail].
What does not go in: your invention disclosure, lab data or drawings.
The vendor's terms say: [no training on our content; retention of X days;
who can access it]. We have attached the relevant terms.

Every claim, citation and statement of inventorship will be checked and decided
by a registered attorney. If you would rather we did not use AI on your matter,
tell us and we will not. Please reply to confirm you agree.

A reply by email is enough to keep with the file. The point is that the inventor knows what is in the tool and what is not.

Prior-art searching: what AI speeds up and what it cannot do

A search has three stages and AI helps with different ones.

1. Query building. Describe the technical field in generic terms (not the inventive concept) and ask for synonyms, alternative technical vocabulary, classification code ideas and adjacent fields. This is where AI is strongest, because inventors and examiners often use different words for the same thing.
2. Retrieval. Run the queries in a real patent database and, where relevant, the non-patent literature sources your firm already uses. A chatbot is not a database. If it names a patent number, treat the number as a guess until you have opened the document.
3. Reading and charting. Paste the full text of the published references you found and ask for a summary of what each one actually discloses, or a first-pass table mapping the draft claim elements to passages. This is tier 1 work, so a general tool is acceptable.

Here is a safe prompt for stage 1. Notice that it describes the problem space, not the solution.

I am searching published patents and papers on keeping a lithium-ion cell within
a safe temperature range during fast charging in small consumer devices.
Give me 25 alternative search phrases a different engineer might use, grouped
as (a) sensing, (b) control of charging current, (c) thermal paths. For each
group, suggest likely classification areas to check. Do not invent patent
numbers; if you are unsure of a term, say so.

An illustrative output would give you groups such as "thermistor placement", "constant-current to constant-voltage handover thresholds", "heat spreader" and "phase-change materials", with a note that classification suggestions need checking against the official scheme. Your fix is to cross-check every classification code against the scheme itself, because model-suggested codes are frequently plausible but wrong.

For stage 3, the failure to watch for is the summary that quietly fills a gap. If you ask "does this reference teach a feedback loop?" and the document does not, a weak summary can answer yes because the field usually has one. Use a prompt that forces quotations.

Below is the full text of one published reference. For each element of the draft claim
listed after it, quote the exact sentence or paragraph number that discloses that element.
If nothing in the text discloses it, write "NOT FOUND". Do not paraphrase and do not
use knowledge from outside the text.

[reference text]

[claim elements, one per line]

Then open the reference and check each quote exists. In practice this check takes a couple of minutes per reference and catches the occasional fabricated paragraph reference. A search you will rely on for a patentability opinion should still be done by someone qualified, with the database results saved to the file.

Claim drafting: using AI on the wording, not the invention

The same principle holds in drafting: the more the task is about language, the better AI does. The more it depends on understanding what the inventor actually built, the more it needs you.

Jobs worth giving to a tool

• Antecedent basis checks. Paste your own draft claims (tier 2 or 3, so only in an approved setup) and ask which terms are introduced with "a" and later referred to with "the" without a first mention. A tool is quick at this, though your own reading is the final check.
• Claim-tree consistency. Ask for a list of every dependent claim and the claim it depends on, flagging dependencies that point to a later claim or to a claim that was cancelled.
• Term consistency between claims and description. Ask for terms used in the claims that never appear in the description, and for terms used with two spellings ("controller" and "control unit") that might be read as different features.
• Alternative phrasings. Ask for three ways to express a limitation at different breadths, then choose and rewrite them yourself.
• Plain-English summaries for the inventor of what the claims cover, so they can tell you where it is wrong.

Jobs to keep with the attorney

• Deciding what the invention is and which embodiments to claim.
• Choosing breadth, fallback positions and claim strategy against the prior art you found.
• Drafting the independent claims that define the client's commercial protection.
• Any statement to the patent office, including declarations and duty-of-disclosure questions.

A tool does not know what the lab bench looks like. If you ask it to write a "detailed description of embodiments" from a one-paragraph summary, it will produce fluent text that adds features the inventor never built. Those additions can become unsupported statements in the specification, or worse, wrongly narrow what the claims read on. If you use a tool to expand text at all, require the inventor to read and mark every statement they cannot support.

A hand-held radio antenna housing, step by step

An illustration for a two-attorney practice with one paralegal. A hardware client brings a disclosure for a clip-on antenna housing that lets a handheld radio swap antennas without tools.

1. Classify (10 minutes). The disclosure form and the photos are tier 3. The generic problem statement, "tool-free mounting of an interchangeable antenna on a handheld device", is tier 1 because it describes a known problem space.
2. Query ideas (15 minutes). Use the generic statement in a general chatbot to get 30 search phrases, such as "bayonet connector", "quarter-turn coupling" and "detent ring". Keep the output in the matter file.
3. Search (3 hours, paralegal plus attorney). Run the phrases in a patent database. Save 40 hits, shortlist 8 after reading titles and abstracts, and read the claims and key figures of the 8.
4. Chart (45 minutes). Paste the full text of the 3 closest references, one at a time, with the quote-or-NOT-FOUND prompt. The attorney opens each quote. Two quotes turn out to be real, one is a paraphrase that does not exist in the document; the attorney discards that row.
5. Draft (attorney, 6 hours). The attorney writes the independent claims from the inventor's disclosure without any AI input, because the disclosure is tier 3 and the client has not signed off on a vetted tool.
6. Consistency pass. After the client agrees to the consent paragraph for a business-plan tool, the draft is checked for antecedent basis and claim-tree errors. The tool flags four items; three are real, one is not. About 20 minutes.

The saving here is modest and honest: maybe four to six hours across the matter, mostly on search phrasing and the first-pass charting. The attorney's drafting time does not shrink, and it should not.

Inventorship and records: what to ask and write down

Inventorship is a question of who contributed to the conception of each claimed invention. Patent offices have published guidance on AI-assisted inventions, and it moves: one major office withdrew its 2024 guidance entirely in November 2025 and replaced it with a shorter position, that the ordinary conception test applies whether or not AI was used and that AI systems are tools used by human inventors, never inventors themselves. Read the current version in each jurisdiction you file in rather than relying on this tutorial or an older summary. The practical rules that follow:

• A tool is not an inventor and is never listed.
• Because inventorship rests on conception, a person who only typed a request into an AI system and accepted the output, with no part in forming the idea of what is claimed, is on weak ground as an inventor.
• If AI was used in the research or design process, record how: which tool, what was asked, what came back, and what the human did with it.

Add three lines to your inventor questionnaire: "Did anyone use an AI system to generate, refine or select any part of this invention? Which one? Please save the prompts and outputs." An illustrative answer, "an engineer used a design-optimisation tool to produce a hinge geometry, then selected one of five outputs", is exactly the sort of fact that changes who the inventors are, so you want it before filing and not during litigation.

Checks to run on every AI-assisted draft before it leaves the firm

Two ways this goes wrong in practice

The convenient paste. A busy associate wants a plain-English summary of an inventor's disclosure for a partner and pastes the form into a free chatbot on their phone. Nothing visible happens, there is no warning, and the habit is easy to repeat. The fix is procedural: a one-page rule saying tier 3 material only goes into the named approved setup, a log of which tools are approved, and an annual reminder. The same logic applies to dictation and transcription apps, which are easy to forget: a recorded inventor call is a tier 3 document.

The persuasive citation. A first-pass memo lists a patent number that matches the topic, with a neat summary. The number belongs to an unrelated patent. Nobody opened it. This is why the check table starts with "Citations exist". The best defence is a rule that no reference is cited in anything leaving the firm until a person has recorded "opened and read" against it.

Choosing a tool: questions that matter more than features

Whether you pick a general assistant on a business plan or a legal-specific product, ask these questions in writing and file the answers. The disclosure side is what separates a safe choice from a convenient one.

1. Is customer content used to train or improve models, by default or under any setting?
2. How long is content retained, and can we set it to zero or delete it ourselves?
3. Who at the vendor and its subprocessors can read content, and in what circumstances?
4. Where is content processed and stored, and can that be restricted?
5. Will the vendor sign a confidentiality or data-processing agreement, and does it mention our professional duties?
6. What happens to our data if we cancel?

Pricing for business plans is modest compared with the exposure: ChatGPT Business lists at $25 per user per month ($20 on annual billing) and requires at least two seats, and Claude Team has a similar two-seat minimum, so a solo practitioner pays for two seats or uses an individual plan. Prices move, so confirm on the vendor page. A legal-specific tool will cost more; the pricing tutorial covers how those are charged, and the general-versus-legal comparison covers the product side.

A first month that keeps risk low

1. Week 1. Write the three-tier rule, pick one approved setup, and send the consent paragraph to current clients who may want AI used.
2. Week 2. Use AI on tier 1 work only: search phrases and reference summaries with the quote-or-NOT-FOUND prompt. Log time saved per matter.
3. Week 3. Add the antecedent-basis and claim-tree checks on filed material under the approved setup.
4. Week 4. Review the log: where did a human catch an error, and was any tier 3 text exposed? Adjust the rule. Decide whether to widen use.

Before you widen it, read the list of legal tasks to keep away from AI and make sure your rule matches. If you want a second pair of eyes on the workflow, that is the kind of thing my AI implementation consultation covers.

Patent practice and AI: quick answers

Does pasting an invention disclosure into a chatbot destroy novelty?

Not automatically, but you cannot rule it out. Whether it counts as a public disclosure depends on the tool's terms, who can read the input and your jurisdictions' novelty rules. Some places have no grace period at all. Treat any consumer tool as unsafe for unfiled inventions and use a tool with written confidentiality and no-training terms.

Should I tell the inventor I use AI on their matter?

Yes. Ethics guidance on generative AI from the American Bar Association says informed consent is needed before client information goes into a tool that may learn from it, and that boilerplate in an engagement letter is not enough. Name the tool, what goes in and what the vendor does with it.

Can AI be named as an inventor?

Inventorship turns on human conception, and patent offices treat AI systems as tools, not inventors. Guidance on AI-assisted inventions has been rewritten at least once since 2024, so read it as it stands today. Ask the inventors what each of them contributed, record it, and do not list a tool. Check the current guidance in each jurisdiction before you file.

Can AI replace a professional prior-art search?

No. AI can speed up query ideas, summaries and claim charts, but it can miss documents, mis-summarise and invent citations. A search you rely on for a patentability opinion needs real database results that you have opened and read yourself.

Further reads

• ChatGPT or a Legal AI Tool: Which Should a Small Firm Use? — Whether a general chatbot or a legal tool suits a small firm.
• AI Paralegal Software: What a Small Firm Should Test Before Paying — A test plan before paying for legal AI software.
• AI Literacy Requirements: What Your Staff Need to Know — What the whole firm should know before using these tools.
• How Small Law Firms Use AI to Draft Letters and Routine Documents — Routine letters and documents, a lower-risk place to start.
• How to Summarise Bundles and Transcripts With AI Safely — Summarising long documents with checks against the source.
• AI Tools for Intellectual Property Firms: Docketing, Searches, Drafts — Where AI helps an IP practice with docketing, searches and first drafts, where it fails, and the checks a lawyer must keep before anything is filed.
• AI Trademark Search for Small IP Firms: What It Finds and Misses — AI trademark search speeds up the first screen but cannot give a clearance opinion. What it finds, what it misses, and how to word the client report.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: American Bar Association Formal Opinion 512 (29 July 2024, read from an archived copy of the published PDF), the November 2025 patent office notice revising inventorship guidance for AI-assisted inventions, OpenAI's business pricing page and data controls help article, Claude's pricing page and Anthropic's consumer data-retention article, all checked 4 November 2026.