AI in Patent Drafting: Discovery, Privilege, and the Evidence Defense

On May 27, 2026, Nicole Berkowitz Riccio and Dominic Rota (Baker Donelson) published the first installment of a JD Supra series titled Privilege, Discovery, and Litigation Risks in Enforcing AI-Drafted Patents. The headline finding is uncomfortable for any practice that has quietly integrated generative AI into patent prosecution: "Communications with GAI platforms may not enjoy the protections of the attorney-client privilege or the work-product doctrine."

The IPWatchdog companion piece (Patent Law Firms Face an AI Reckoning, May 26, 2026) frames the operational scale of the problem: lawyers are now receiving roughly 50 pages of AI-generated material per matter that is "bloated junk, inaccurate, redundant or untethered to what the inventor actually invented." The combination of those two realities is a discovery problem waiting to happen.

This page walks the Baker Donelson risks, draws the distinction between generic GAI use (which creates the problem) and structured-disclosure tools with evidence chains (which do not), and offers practical guidance for prosecution practices integrating AI in 2026.

This article presents law-firm analysis and published commentary for educational purposes. It is not legal advice. Consult a registered patent attorney and your jurisdiction's professional responsibility authority for filing and litigation decisions.

The Baker Donelson risks, summarized

Riccio and Rota identify three privilege-and-doctrine risks that practitioners using generative AI in patent drafting should anticipate:

1. GAI platforms are not lawyers. Attorney-client privilege protects confidential communications between an attorney and her client made for the purpose of obtaining legal advice. A communication with a third-party GAI platform is not that. The attorney's prompt and the platform's response sit outside the privilege.
2. Feeding privileged information into GAI may waive privilege. If an attorney pastes a client's confidential disclosure into a hosted GAI tool, the act may constitute voluntary disclosure to a third party — the platform operator — and may waive privilege over the underlying client communication. The waiver risk depends on platform terms, jurisdiction, and how the material was used downstream.
3. Work-product doctrine probably does not protect GAI outputs in prosecution. Work-product protection generally requires that the material be prepared "in anticipation of litigation." Patent prosecution is not generally that. So even if the AI's output reflects attorney mental impressions, the prosecution context may exclude it from work-product protection.

Riccio and Rota also enumerate six categories of discovery requests that future patent litigants should anticipate. Each is a question discovery counsel will ask, and each is a question that becomes more or less defensible depending on the prosecution practice's AI workflow:

• What AI tools were used in drafting the application?
• What prompts and inputs were provided to those tools?
• What outputs were generated, and how were they incorporated?
• What §101 arguments were drafted with AI assistance?
• What prior-art searches were performed by or with AI tools?
• What internal policies governed AI use during prosecution?

The practical advice in the article is straightforward: discuss GAI implications with clients in R&D, and ensure litigation hold notices cover prompts, inputs, and outputs. Practitioners who do not anticipate these discovery questions will be ambushed by them later.

The IPWatchdog frame: AI inverts the disclosure problem

IPWatchdog's analysis adds operational context. The traditional patent-disclosure problem was scarcity: inventors did not write enough down. The 2026 disclosure problem, in practices that have adopted generative AI, is the opposite — abundance without discipline.

From the IPWatchdog piece: lawyers now receive "50 pages of bloated junk... inaccurate, redundant or untethered to what the inventor actually invented." The author's argument is that the winners in this transition will be firms that know "where AI belongs in the workflow, and where it absolutely does not."

The two observations stack. If a practice produces 50 pages of low-discipline AI output per matter, and those 50 pages may not enjoy privilege protection in discovery, the result is litigation exposure proportional to AI volume. The more aggressively a practice has integrated unreviewed AI, the larger its discovery footprint.

The distinction that matters: generic GAI use versus structured disclosure tools

Not all AI use in patent prosecution carries the same exposure profile. The Baker Donelson risks attach to generic use of generative AI platforms — pasting client confidences into a hosted LLM, prompting it for draft claims or arguments, accepting the output more or less as written. That workflow is undisciplined by construction and creates exactly the discovery profile the article warns about.

A structurally different category of AI tools exists and behaves differently in discovery. Code-scanning disclosure pipelines that produce structured technical disclosures anchored to specific source-code citations have a fundamentally different evidence profile than generic GAI drafting. Four reasons:

1. The output is a technical observation, not legal work

A scanner that reports "this function implements an unconventional approach to X, distinct from the conventional Y approach used in [comparator]" is not making a legal argument about patentability. It is making a technical observation about a codebase. The attorney's downstream legal work — drafting claims, arguing eligibility, framing prior-art distinctions — happens after that technical observation lands. The technical observation itself is the kind of factual content that would be discoverable regardless of how it was produced; nothing in the privilege analysis changes because a computer produced it instead of a junior associate.

2. The evidence chain is reproducible

Generic GAI output is famously non-reproducible: prompt the same model with the same input twice and get different results. That non-reproducibility is itself a discovery liability — opposing counsel will argue that the prosecution attorney relied on outputs that cannot be verified. A structured-disclosure tool that points to specific files, specific lines, specific commits, and specific scoring rubrics produces an evidence chain that is far more reviewable. When the repository snapshot, scanner version, rubric, model configuration, prompts, and outputs are preserved together, the analysis can be reviewed and reliably reproduced. Bit-identical reproduction depends on which of those inputs are locked; the discipline of locking them is itself the defensible practice.

3. The "scanner identifies, attorney determines" framing matches Thaler

Thaler v. Vidal (Fed. Cir. 2022) requires that inventors be natural persons. The USPTO's November 2025 guidance treats AI strictly as a tool under human conception standards. A prosecution workflow where AI assists drafting, and the attorney determines patentability, conception, and inventorship, fits cleanly inside that doctrinal posture. A prosecution workflow where AI produces drafts the attorney does not meaningfully review is, in the IPWatchdog framing, "bloated junk" and is also, doctrinally, in tension with the human-conception requirement.

4. Audit trails are litigation-grade by design

The six discovery categories Baker Donelson identifies all benefit from contemporaneous logging. A workflow that records which scans were run, when, against which codebase commit, with which scoring rubric, and which output the attorney actually relied on, produces a record that survives discovery. A workflow that does not log AI use, or that logs it informally, will struggle to defend against discovery requests about what was used and how.

Practical adoption guidance for prosecution practices in 2026

For practices that are using or about to use AI tooling in patent prosecution, six practical steps reduce exposure:

1. Stop pasting client confidences into hosted GAI platforms. The privilege risk is concrete and the waiver question is uncertain. Use local-first tools where the technical material does not leave the client's infrastructure or counsel's infrastructure.
2. Distinguish technical-observation outputs from legal-draft outputs in your workflow. Technical observations (this code uses an unconventional approach) are likely discoverable regardless of how produced; treat them accordingly. Legal drafts are different — keep AI use for legal-draft work disciplined and reviewed.
3. Maintain contemporaneous logs of AI tool use. Which tool, which version, which inputs, which outputs, which the attorney relied on. Litigation hold notices should cover these logs.
4. Document attorney judgment exercised on AI output. Every patent drafted with AI assistance should have a clear record of the human attorney's decisions: what was accepted, what was modified, what was rejected. This is the inventorship-and-conception story for the file history.
5. Prefer tools that produce reproducible, evidence-anchored output. A scanner that points to specific code is harder to attack in discovery than a chatbot that generated unreproducible text.
6. Have the GAI conversation with clients early. R&D teams are using AI too. The privilege analysis is no easier on the client side than on the attorney side. Discuss tool selection and information handling before a dispute arises, not after.

What to log: a discovery-defensible AI workflow checklist

The Baker Donelson piece's six discovery categories all become easier to defend when the underlying AI workflow has produced a contemporaneous record. The following items, locked together per matter, turn a "what did you use AI for?" discovery question into a documented answer:

This checklist is derived from Baker Donelson's discovery-category enumeration (JD Supra, May 27, 2026) and applied to the specific properties of structured-disclosure tools. It is not legal advice; consult counsel and your jurisdiction's professional-responsibility authority before adopting any specific logging practice.

How structured-disclosure tools fit prosecution practice

ObviouslyNot's code scanner is built to the structured-disclosure-tool profile described above. The output is:

• Anchored to specific source-code citations, with file paths, line numbers, and commit references.
• Structured as technical observations ("this approach is distinct from X conventional approach"), not as legal conclusions about patentability.
• Scored with explicit criteria that opposing counsel could re-run if challenged.
• Framed as input to attorney judgment: the scanner identifies, the patent attorney determines whether what was found may be patentable.

That output profile is specifically defensible against the Baker Donelson risks. The technical observations are the kind of factual content a layperson could produce by code review; nothing about computer-assisted production changes the discovery analysis. The evidence chain is reproducible. The human-conception requirement is preserved. The audit trail is litigation-grade.

See our attorney landing page for how this fits prosecution practice, or download the local-first scanner for the workflow that keeps client code on the client's infrastructure.

The doctrinal and policy landscape this sits inside

The Baker Donelson risks do not exist in a vacuum. Three 2026 developments shape the surrounding terrain:

• USPTO Director Squires's April 30, 2026 SMED memo (see our Section 101 Declaration Strategy page) encourages structured evidentiary support for §101 arguments. SMEDs ask for exactly the kind of anchored, comparator-specific technical evidence that scanner-based workflows produce.
• USPTO November 2025 inventorship guidance treats AI strictly as a tool under human conception standards. The structured-disclosure-tool workflow is doctrinally aligned with this; the unreviewed-GAI workflow is in tension with it.
• State-level legislation on AI ownership (Arkansas and Iowa in particular, per Mayer Brown analysis) is beginning to fill the federal gap on AI-generated works. The disclosure question — who contributed what, when — becomes the central question of inventorship determination as that landscape develops.

The 2026 USPTO posture is reasonably coherent: it wants quality patents with documented evidentiary support, preserved human-conception responsibility, and disciplined AI integration. Prosecution practices that align their AI workflow to that posture will be better positioned for the next decade of patent litigation than practices that integrated AI without considering its discovery footprint.

Sources

Related from the resources