AI Patent Risk: The 2026 Firm-by-Firm Consensus

Six months ago this was forward-looking commentary. As of June 2026, the question has a concrete federal data point (the Conservation Law Foundation v. Shell Oil Co. order compelling production of an expert's AI prompts), an updated USPTO procedural framework (the April 30, 2026 SMED memo on Section 101 declarations), and a wave of practitioner alerts from major IP groups. This page synthesizes what the major firms are publicly saying, where the consensus is solid, and where the disagreements matter.

Our short version: at the AmLaw 50 level, the practitioner community has converged on the position that the record of how an AI-assisted patent was drafted is now part of the patent. That conclusion is consistent across firm publications, even when the firms have different client bases and different recommendations.

This article summarizes publicly available law-firm analysis for educational purposes. It is not legal advice. Consult a registered patent attorney for filing or litigation decisions.

The publications we synthesized

These are not the only firms publishing on this set of questions. They are the publications with the clearest practitioner-facing analysis and the most actionable recommendations, spanning from Hogan Lovells's February 2026 framework piece through Baker Donelson's June 3 validity follow-up. Hogan Lovells's piece predates the May 18, 2026 CLF v. Shell order; the rest of the publications respond to it. The 1-day Baker Donelson date gap on Parts 1 and 2 reflects the lag between Baker's own publication and the JDSupra syndication. Citations in the Sources section.

Where the consensus is solid

Five propositions appear across all or nearly all of the firm publications. These are the durable practitioner takeaways.

1. AI prompts are part of expert methodology, not search terms

The firms that responded to CLF v. Shell reach the same conclusion: the court got the Rule 26(b) analysis right. Prompts to a generative model are not search terms; they embed judgment, framing, and instructional choices that materially shape the output. Under Daubert, methodology has to be testable. The prompt is part of the methodology. It is discoverable. (Hogan Lovells's February 2026 framework piece reaches the same conclusion as a general matter, before the CLF order put it on a docket.)

Practitioner implication: testifying experts in patent litigation should expect motions to compel AI prompts, and counsel should not assume search-term doctrine will protect prompts.

2. Attorney-client privilege does not automatically extend to AI interactions

Baker Donelson states this most directly: GAI is not a lawyer, and inputs may be "voluntary disclosure to a third party," with the result that any privilege attaching to the inputs may be waived. Hogan Lovells, Kirkland, and Alston & Bird reach compatible conclusions through slightly different framings. The point of agreement: the privilege analysis is unsettled and counsel should not assume protection.

The mitigation each firm recommends: enterprise GAI platforms with contractual no-training-no-retention terms are easier to defend than consumer-grade tools, but neither is automatically privileged.

3. Work product is at risk in patent prosecution because the "anticipation of litigation" element is weak

Standard work-product doctrine under Hickman v. Taylor and Rule 26(b)(3) requires the material be prepared "in anticipation of litigation." Patent prosecution work, in most cases, is not. That is a structural weakness in any work-product argument over prosecution-phase AI use.

Baker Donelson Part 1 makes this explicit. Hogan Lovells frames it more cautiously. Both agree the doctrine has been doing less work for prosecution-phase materials than practitioners often assume, and that AI-assisted prosecution makes the gap more visible.

4. Litigation holds now have to extend to AI prompts and outputs

Unanimous across publications. Every firm recommends that litigation-hold language explicitly require preservation of prompts, queries, system prompts, model identifiers, and (where available) outputs. A hold notice silent on AI is now a spoliation-risk problem, not a neutral default.

5. Validity exposure runs in multiple directions, not just §101

Baker Donelson's Part 2 is the most thorough on this point. AI-assisted drafting introduces validity exposure across:

• §102 (novelty): If consumer-grade GAI retains prompts or outputs for training, that data could become "otherwise available to the public" and trigger prior-art problems. The one-year grace period may not protect against this.
• §112(a) (written description): GAI-generated technical detail may undermine the "possession" requirement. Baker Donelson notes the counter-argument that written description focuses on the four corners of the spec, not how it was drafted (citing Ariad Pharmaceuticals v. Eli Lilly), but the question is unresolved.
• §101 (eligibility): AI-drafted claims that lean on generic technical language are at heightened risk under the post-Alice line of cases. See our AGI SureTrack v. Farmers Edge analysis for the most recent precedential data point.
• Inventorship (§115): Inventorship still requires human conception. The Federal Circuit's Thaler v. Vidal ruling and the USPTO's November 28, 2025 revised AI inventorship guidance continue to draw the line at AI-as-tool. AI that contributes more than tool-level assistance may invalidate the patent for improper inventorship.

Practitioner implication: AI-assisted drafting is not a single-axis risk. It compounds across multiple invalidity doctrines.

Where the firms disagree, and what that signals

How aggressive to be in restricting expert AI use

Kirkland's commentary is closest to "prohibit consumer-grade GAI for testifying experts entirely, mandate enterprise platforms with no-training/no-retention contracts." Hogan Lovells is more permissive, framing the question as engagement-letter scope rather than prohibition. Alston & Bird falls in between.

Reading the disagreement: the firms with high-stakes commercial-litigation practices (Kirkland) are more conservative; the firms with broader IP advisory portfolios (Hogan Lovells) are more accommodating. Practitioners can read this as a reasonable spectrum, with the conservative end probably right for high-value disputes.

Whether a Rule 29 stipulation can cure the discovery exposure

Arnold & Porter and Kirkland both emphasize the role of the Rule 29 stipulation in CLF v. Shell (which was silent on AI and therefore did not protect plaintiff). The implicit framing: better-drafted stipulations can shape the discovery exposure.

Baker Donelson and Hogan Lovells are more skeptical that stipulations will reliably contain the issue, particularly when one party has strategic incentive to seek production. Negotiated stipulations help; they are not a complete answer.

Whether validity exposure under §112 should change drafting practice

Baker Donelson Part 2 stakes out a clear position: yes, drafting practice should change. Specifically: substantively review AI outputs, document the review/revision process, and prefer enterprise GAI with data isolation.

The firms publishing primarily on discovery (Alston & Bird, Arnold & Porter) do not engage as deeply with §112 written-description exposure. The asymmetry suggests the discovery angle is more developed; the validity angle is still building.

The consensus recommendations, distilled

Across the publications, the recommendations cluster into a short list of practical steps. None is novel in isolation; the collective weight is the signal.

1. Discuss AI use with clients up front. Engagement letters should specify whether GAI use is permitted, on what platforms, with what retention controls, and with what disclosure obligations. Default-silent engagement letters carry hidden risk.
2. Prefer enterprise GAI with contractual data isolation. No-training, no-retention, single-tenant, auditable. Consumer-grade GAI is harder to defend in discovery and creates §102 exposure through retention.
3. Substantively review every AI output. The Baker Donelson framing: the attorney's review, revision, and integration of AI output is what creates defensible authorship of the application. Outputs treated as production-ready dictation are more exposed.
4. Document the review and revision process. Time-stamped records of what changed between AI draft and filed application, and why. This is the practical answer to "did the attorney exercise independent judgment over the AI's contribution."
5. Update litigation-hold templates to cover AI prompts and outputs. Both for in-house counsel issuing holds to employees, and for outside counsel issuing holds to expert witnesses.
6. Negotiate Rule 29 stipulations with AI explicitly in mind. Silence is not protection.
7. Build the spec around technical evidence the inventor controls. The more the patent record rests on artifacts the inventor's team produced (code commits, design docs, measured behavior), the less the AI-assisted prosecution workflow becomes the story in litigation.

What this means for the structured-disclosure approach

The firm-by-firm consensus matters for ObviouslyNot in a specific way. We have been arguing for the past 18 months that there is a structural difference between two workflows:

• Generic generative AI for patent drafting: prompts are the methodology, outputs are non-deterministic, the audit trail lives in chat history at best, and the practitioner has to manually reconstruct what the AI contributed.
• Structured disclosure tooling: deterministic scans of an inventor's actual codebase, with citations to specific files and commits, producing an audit trail by design rather than by retrofit.

The AmLaw firm consensus, read carefully, validates the structural difference. Every one of the seven consensus recommendations above maps to a property the structured-disclosure approach has by construction: deterministic file-and-commit identification, controlled retention, reviewable artifacts, and an audit trail rooted in the technical record itself. The recommendations are designed to retrofit generic GAI workflows toward that property set.

This is the framing of the question that lets ObviouslyNot's product positioning intersect cleanly with what major-firm IP practitioners are now publicly recommending. The recommendation set was not designed for our product. It happens to be a description of what our product does.

See AI Patent Discovery & Privilege for the underlying analysis of why this matters in litigation, the CLF v. Shell discovery order for the case the firm coverage is responding to, and Section 101 Declaration Strategy for how the prosecution-side record is built.

Nuance: AI prompts are not uniformly discoverable

A reasonable reading of the synthesis above is "AI prompts are discoverable in litigation, period." That reading is too broad and the firms know it. Arnold & Porter's June 2026 framework piece ("The Emerging Framework on AI Prompts, Privilege, and Discovery") makes the distinctions explicit. Hogan Lovells and Kirkland reinforce them. The right reading of the synthesis is conditional, not categorical.

Categories drawn from Arnold & Porter, "The Emerging Framework on AI Prompts, Privilege, and Discovery" (June 2026), cross-referenced against Hogan Lovells and Kirkland coverage of the same framework, and corroborated by Akin Gump's May 28, 2026 analysis of diverging Q1 2026 federal rulings on GenAI privilege, work product, and protective orders. The category list is descriptive of the current federal-court landscape, not exhaustive.

For patent practice the implication is concrete: blanket statements about "AI prompts" obscure the doctrinal sorting that actually determines exposure. The right operational question is not "is my AI use discoverable" but "in which of these four categories does my AI use sit, and what discipline does that category require." See our emerging-framework page for the longer treatment.

What we are not claiming

• Firm consensus is not law. Practitioner alerts are good guidance and useful predictive signal, but the binding authority is judicial decisions, USPTO guidance, and statutes. We have cited those above; the firm analyses interpret them.
• Not all AmLaw firms agree. We synthesized the six publications above. Other firms are publishing in this space and some may take different positions. The convergence we describe is real but not unanimous across the entire bar.
• Generic GAI is not banned. The consensus is that it carries risk in patent prosecution and litigation. Risk can be managed. The recommendations above are management strategies, not prohibitions.
• This is not legal advice. Consult counsel about your specific matter. The publications cited are for educational reference.

Sources

Related from the resources