AI Patent Litigation: Three Defensibility Records
- Pouya Shafabakhsh
- Aug 21
- 6 min read
Updated: Aug 24
Answer first: AI-assisted patent work is best governed as three separate questions: confidentiality, verification, and evidentiary methodology. A proportionate record for each question is more useful than a single generic “AI compliance” log.
AI-assisted patent work is not one risk. It sits at the intersection of confidentiality, filing integrity, and evidentiary reliability. Those are different questions, controlled by different authorities and different facts.
For U.S. patent litigators, the practical issue is not simply whether AI was used. It is whether the team can explain how sensitive information was handled, how material legal and technical propositions were verified, and how an AI-influenced method or item of evidence could be examined if challenged.
This is independent Litigation AI GRC analysis, not legal advice. The governing law, court orders, client requirements, protective orders, professional duties, vendor terms, and facts of a particular matter must be assessed separately.
Why AI-Assisted Patent Work Needs Three Records
Confidentiality record: security is not privilege
A system can be technically secure without answering the legal question of confidentiality. Privilege does not arise from a security label, enterprise contract, or local hosting alone. The analysis depends on who communicated what, for what purpose, under what expectations of confidentiality, and with what third-party access or disclosure conditions.
United States v. Heppner, a 2026 Southern District of New York criminal case, is useful because it should be read narrowly. Judge Jed S. Rakoff held that the particular AI-generated materials at issue were not protected by attorney-client privilege or work product on the facts before the court. The defendant used the consumer version of Claude on his own initiative, and the court focused on confidentiality and third-party disclosure. The decision did not establish a universal rule that every enterprise AI workflow waives privilege.
New York state litigation supplies an equally important caution against blanket conclusions. In Assini v. Hayward, the Nassau County Supreme Court distinguished Heppner, found Morgan v. V2X persuasive, and quashed subpoenas seeking OpenAI materials. Assini likewise did not create a universal shield around AI prompts or outputs. Together, the cases reinforce a basic governance point: context matters.
For patent teams, the operational record can identify the tool and service tier, the categories of data exposed, relevant retention and access settings, material transfers, and client or matter restrictions. That record supports governance; it does not replace a privilege analysis.
Data location also needs precision. 18 U.S.C. § 2713 reaches information within a covered provider’s possession, custody, or control regardless of whether it is stored inside or outside the United States. That is a provider-side disclosure-reach provision, not a privilege-waiver rule.
Verification record: the signer still owns the submission
The second record begins where AI-assisted patent research, claim analysis, prior-art work, or drafting becomes a submission.
The USPTO’s guidance on AI tools does not mandate a special architecture or prohibit AI. It emphasizes that existing obligations continue to apply regardless of how a submission was generated and points practitioners back to duties including 37 C.F.R. § 11.18. The guidance also identifies risks of AI-generated omissions, misstatements, and fabricated content.
Federal court filings carry their own obligations under Rule 11. In New York state courts, Part 161 similarly takes a measured approach: AI use in preparing court papers is not generally prohibited, and disclosure is not automatically required. A court may adopt the model rule, which expects independent review for fabricated or fictitious authorities. Part 161 expressly separates court papers from materials offered as evidence.
For a patent team, the verification record need not capture every lawyer thought. It can show that consequential propositions were checked against authoritative sources: the patent, prosecution history, cited prior art, USPTO records, governing cases, local rules, standing orders, and the evidentiary record.
Where the Controls Become Concrete
Before a USPTO or court filing
Patent filings can compress engineering facts, prosecution history, expert analysis, and legal authority into a few pages. AI may accelerate parts of that work. The governance question is whether it merely organized known material or materially influenced a proposition.
Counsel can calibrate the record accordingly. Formatting assistance does not require the same evidence as AI-assisted prior-art synthesis, claim-chart development, technical summarization, or a proposed argument. The more consequential the influence, the stronger the case for a reproducible source trail and human verification.
This respects professional judgment. It does not require turning every prompt into a litigation exhibit. It asks the firm to know where AI materially entered the workflow and what evidence supports the final human decision.
When AI-derived evidence or methodology is contested
The third record matters when an AI-derived output, classification, comparison, reconstruction, or technical method becomes relevant to proof.
Federal Rule of Evidence 702 addresses reliability when expert testimony is offered. Rules 901 and 902 address authentication and self-authentication in their respective settings. None creates a special “AI evidence” shortcut. The foundation depends on what the proponent claims the item is, how it was created, and how it is used.
Useful methodology records may include the model or software version, material configuration, defined inputs, source corpus, preprocessing steps, human interventions, validation method, known limitations, timestamps, and preserved outputs. Not every field is necessary in every case. Preserve what allows a qualified examiner to understand and, where appropriate, reproduce the consequential process.
A Practical Litigation AI GRC Control Model
Matter-level checklist for patent teams
Ask five questions before the dispute hardens:
What confidential or proprietary information entered an AI-enabled system, and under what access, retention, and transfer conditions?
Which AI-assisted outputs materially influenced claim analysis, prior-art review, technical characterization, expert work, or a filing?
Which primary sources independently support those material propositions?
If a method becomes disputed, what records would allow a qualified examiner to understand its inputs, versions, controls, limitations, and human interventions?
Which records should not be preserved or disclosed without a matter-specific privilege, work-product, protective-order, and discovery analysis?
That fifth question matters. Good governance is not indiscriminate logging. Unnecessary copies of privileged strategy or sensitive matter data can create their own risk.
When independent testing may be appropriate
Independent testing can be useful when counsel needs technical assurance without asking the production environment to validate itself. An Air-Gapped Shadow AI Audit can examine defined workflow evidence inside an isolated environment designed to reduce additional external exposure. A jointly retained Judicial Forensic AI Audit can address agreed technical questions where parties seek a common examination baseline.
Neither approach is required by the authorities discussed here, and neither guarantees privilege, admissibility, compliance, or litigation outcomes. A jointly retained examiner is also not the same as a court-appointed expert; Federal Rule of Evidence 706 provides a separate court-controlled mechanism.
Conclusion: Build the Right Record Before the Challenge
The practical lesson is not “never use AI,” and it is not “log everything.” It is to maintain the right record for the right question.
Confidentiality needs a record of custody, access, retention, and material transfers. Filing integrity needs authoritative verification for consequential propositions. Evidentiary reliability needs enough methodology evidence to examine a disputed process without confusing vendor assurance with admissibility.
Once opposing counsel, a client, the USPTO, or a court asks how the process worked, retrospective reconstruction is slower and less reliable than disciplined matter-level governance.
Practical checklist
Classify sensitive patent, technical, client, and expert data before AI use.
Record the relevant AI service, access boundary, retention conditions, and material transfers.
Identify consequential AI-assisted work rather than treating every use identically.
Verify material patent and legal propositions against primary sources.
Preserve proportionate methodology records for AI-influenced evidence or expert work.
Keep privilege analysis separate from technical-security claims.
Reassess controls when the model, provider, workflow, protective order, or litigation posture changes.
FAQ
Does using AI automatically waive attorney-client privilege?
No. Privilege is fact-specific. Heppner illustrates a confidentiality problem on its facts; it should not be converted into a universal rule for every AI service or workflow.
Does keeping patent data in the United States eliminate disclosure risk?
No. Data location is only one factor. Section 2713 addresses certain provider obligations for information within possession, custody, or control regardless of location.
Does the USPTO prohibit AI-assisted patent drafting?
The USPTO’s published guidance does not impose a blanket prohibition. It emphasizes that existing duties and certifications continue to apply.
Must New York lawyers disclose AI use in every court paper?
Part 161 says disclosure should not automatically be required. Courts may adopt the model rule, so the actual court and part must be checked.
Does a technically reliable AI output automatically satisfy the Federal Rules of Evidence?
No. Reliability, authentication, and admissibility depend on the purpose for which the evidence is offered and the applicable rules.
Is a jointly retained forensic AI examiner the same as a court-appointed expert?
No. Parties may jointly retain an examiner by agreement. FRE 706 is a separate court-appointment mechanism.
Before AI use becomes the issue opposing counsel, a client, the USPTO, or the court asks you to reconstruct under pressure, it is worth knowing where your current matter-level controls actually stand. If your team is using AI in patent research, prior-art review, claim analysis, drafting, expert work, or evidentiary preparation, complete our Litigation AI GRC Assessment to identify potential gaps across confidentiality, verification, methodology, and defensibility. The assessment is designed to help Managing Partners and patent litigation teams determine where stronger governance records may be warranted before the challenge begins.
We appreciate the completion of the Assessment Form at:
Author: Pouya Shafabakhsh Co-Founder, CAIO & Principal Forensic AI Auditor, Radsam Academy of AI Sovereign Governance. The Architect of North America's: Judicial Forensic AI Audit Standards, AI Governance, Risks & Compliance Standards, Air-Gapped Sovereign Sanctuary AI Audit System.
Comments