Case law, analysis, and governance references

A company executive under criminal investigation used Claude to draft defense strategies without attorney direction. The court held that privilege did not attach because no attorney was involved at the front end, the AI platform is not an attorney, and the platform’s privacy policy permitted data disclosure. Sharing the outputs with counsel afterward did not cure the defect. Cited as 2026 WL 436479; 820 F. Supp. 3d 292. Practitioners read the opinion as a warning for corporate employees who put legal or regulatory questions into consumer or even private LLMs without counsel directing the work.

Takeaway: Enterprise deployment alone does not create privilege. Without counsel direction, AI legal analysis by employees is ordinary evidence.

The leading corporate AI-evidence case. Facing a projected $250 million earnout, Krafton’s CEO asked ChatGPT how to take control of the acquired studio and avoid the payment — after counsel had already told him a for-cause strategy would not work. He pasted the chatbot outputs into Slack with colleagues, so any privilege claim was waived before discovery even started. Vice Chancellor Lori W. Will quoted the exchanges at length, matched the AI playbook to what the company actually did, found pretext and bad faith, reinstated the target CEO, and extended the earnout period by 258 days. The CEO also admitted deleting the original chat logs. Cited as 2026 WL 730977; 354 A.3d 906; C.A. No. 2025-0805-LWW.

Takeaway: This is what happens when a company encourages AI use without routing sensitive strategy through counsel: the prompts become Exhibit A. Copilot and ChatGPT Enterprise do not change the privilege analysis if no attorney directed the work.

Massachusetts first-impression ruling in a commercial note dispute. A represented plaintiff had his romantic partner upload transaction documents to ChatGPT and draft a response without consulting counsel. Judge Debra A. Squires-Lee compelled production: neither the partner nor ChatGPT was a qualifying “representative,” and AI output is not opinion work product because “AI is a tool not a person.” The court distinguished Warner and Morgan as pro se cases and aligned with Heppner. No. 2684CV00799-BLS2.

Takeaway: Companies that tell staff to “just use ChatGPT” on live disputes create the Shealy fact pattern: represented-party AI work, no counsel direction, production compelled.

A pro se plaintiff used AI tools in litigation preparation. The court protected the materials as work product, rejecting the argument that AI platform use constituted waiver. ChatGPT and similar programs “are tools, not persons,” so using them is not disclosure to an adversary. Cited as 2026 WL 373043; 820 F. Supp. 3d 629.

Takeaway: Work product outcomes are fact-specific. Pro se status and treating AI as a tool, not a person, drive the protective civil line.

Magistrate Judge Dominguez Braswell held that a pro se litigant’s AI-assisted preparation is protected work product under Rule 26(b)(3), but that the identity of the AI tool is not protected and must be disclosed. The court also imposed an AI-specific protective order barring confidential information from any AI platform unless the provider is contractually prohibited from training on inputs or disclosing them, and permits deletion on request. Cited as 2026 WL 864223.

Takeaway: Even where work product holds, the choice of AI tool can be discoverable, and confidential material may not enter consumer AI absent contractual safeguards — precisely the controls Saidebar builds into the workflow.

First major New York state decision on GenAI discovery. The court quashed a non-party subpoena to OpenAI seeking a pro se defendant’s prompts, uploads, and outputs used to prepare filings. Citing Morgan as persuasive and distinguishing Heppner, the court treated AI-assisted litigation preparation as conditionally protected under CPLR 3101(d). Cited as 2026 NY Slip Op 26086; 2026 WL 1677232.

Takeaway: State courts are extending the Warner/Morgan line for pro se litigants, while still warning that unverified AI filings can draw sanctions.

First state business-court work-product ruling on GenAI. After in camera review, Judge Dorfman held that a non-lawyer principal’s ChatGPT conversations prepared in anticipation of litigation are protected under Tex. R. Civ. P. 192.5(a)(1), and that using the tool did not waive protection. As in Morgan, the court still ordered disclosure, by Bates number, of every discovery document fed into ChatGPT. Cause No. 25-BC11B-0020.

Takeaway: Courts increasingly protect AI-assisted work product but treat the fact and scope of AI use as discoverable. Feeding confidential discovery into consumer AI can breach a protective order.

In a putative toxic-tort class action against a chemical facility operator, Magistrate Judge Angel D. Mitchell amended the protective order to bar parties from uploading any discovery materials — including non-confidential documents — into public or “open-loop” generative AI tools such as consumer ChatGPT or Copilot. Closed, secure tools with no-training, no-disclosure, and deletion controls remained permitted. The court cited irretrievable training exposure, privacy/GDPR risk, and the incentive for parties to under-produce if open AI were allowed. Cited as 2026 WL 820218; Nos. 25-2352 / 25-2569.

Takeaway: Courts are no longer treating consumer AI as a free discovery workspace. Companies that rolled out enterprise Copilot without closed-loop controls now face protective orders that effectively ban the tools their employees already use.

Plaintiff’s counsel uploaded Walmart’s discovery responses into an AI tool, asked it to identify deficiencies, and copied the unreviewed output to opposing counsel and the court. Magistrate Judge Tim A. Baker held that exclusive reliance on AI does not satisfy Rule 37(a)(1)’s good-faith meet-and-confer duty: AI “can be a useful discovery tool” but “is not a substitute for good lawyering.” No. 1:25-cv-01120-RLY-TAB.

Takeaway: Attorney judgment remains non-delegable even in routine discovery. Unreviewed AI output sent to the court or opposing counsel is a process failure, not a productivity gain.

The Ninth Circuit sanctioned and suspended two attorneys for filing briefs with nonexistent cases, misattributed quotations, and gross misrepresentations, then failing to candidly disclose that generative AI hallucinations were the source. The court stressed it was not punishing AI use itself — the violation is signing and filing unverified material. For two years, every attorney at their firm must disclose AI use in filings and certify personal citation review under penalty of perjury. No. 24-4790.

Takeaway: Appellate courts now treat unverified AI filings as professional misconduct with firm-wide operating consequences. Supervised verification before sign-off is the compliance baseline.

Magistrate Judge Farrish ordered production of generative AI prompts an expert witness used to cull and analyze document productions. Expert methodology remains fair ground for discovery; protections that may shield a party’s litigation strategy do not extend to a testifying expert’s AI process. Case No. 3:21-cv-00933.

Takeaway: Wall off testifying experts from privileged attorney-directed AI workflows. Expert prompts and queries are methodology, not strategy.

In a discovery dispute, the court compelled production of AI prompts, outputs, account settings, and testing process. The court held that placing a subset of test results in the complaint waived work product protection over related negative results.

Takeaway: AI prompts, outputs, and testing process can become central evidence. Account settings and interaction history are discoverable.

March 2026 corporate advisory: unless working at the express direction of counsel, non-lawyers should not use even private AI tools for legal or regulatory analysis. C-suite and junior employees alike create discoverable records when they query AI about legal exposure — enterprise firewall alone does not fix it.

Detailed reconstruction of Krafton: ChatGPT outputs pasted into Slack waived any privilege argument; deletion of original logs became additional motive evidence. Explicitly flags Microsoft 365 Copilot, Google Workspace AI, Slack AI, and Notion AI as hold-scope data.

March 2026 employment guide: HR and manager prompts like “how to document a problem employee before firing her” are discoverable, unprivileged, and quotable in the discrimination suit that follows. Train staff that privilege does not attach to AI use.

Information-governance alert: Copilot Chat is enabled by default and quietly captures prompts, responses, referenced files, and linked documents in hidden Exchange locations searchable via Purview eDiscovery. Rolling out enterprise Copilot without retention and hold workflows creates a discoverable archive by design.

August 2026 employment-side brief translating Heppner for employers: supervisor and HR consumer-AI chats about terminations and investigations are ordinary ESI. Litigation holds must reach prompts, outputs, and activity logs.

July 2026 synthesis of eight opinions in four months mapping the Heppner outlier against the Warner/Morgan/Assini/Tate civil line, plus expert-prompt, protective-order, and deposition-misuse cases. Practical checklist: direct the use, choose closed tools, assume tool identity is discoverable, wall off experts.

April 2026 analysis of Morgan and Jeffries: courts are writing AI restrictions into protective orders, distinguishing open consumer tools from closed enterprise systems with contractual no-training terms.

July 2026 alert explaining why represented-party AI use routed through a non-attorney sank work product, and why courts favorably cite Heppner when counsel did not direct the work.

August 2026 analysis of Shealy as Massachusetts first impression: AI output is not opinion work product absent attorney direction or input.

Structural reading of Judge Rakoff’s memorandum: GenAI is not incompatible with privilege; unsupervised consumer use is. Points to Kovel-style protection where counsel directs a confidential AI workflow.

Comprehensive bar association report on how model training, retention, and privacy policies interact with privilege and work product, surveying Heppner, Warner, and Morgan.

July 2026 Fried Frank analysis synthesizing Heppner, Krafton, Warner, Morgan, and Tate. Identifies the central divide as whether AI is a litigation tool or the equivalent of a person giving legal advice, and recommends that companies require counsel to direct GenAI use on legal matters.

April 2026 analysis of Fortis Advisors v. Krafton detailing how the Court of Chancery quoted the buyer’s AI prompts and how the absence of counsel direction removed any privilege claim.

Side-by-side of Warner and Heppner decided the same day, framing the tool-versus-person split and the counsel-direction requirement.

Practitioner summary of Lnu v. Blanche: the misconduct is signing and filing unverified material and lack of candor, not AI use itself; firm-wide disclosure and verification orders follow.

Comparative analysis of Heppner, Warner, and Morgan, noting the open question of how courts will treat a represented party who uses AI without counsel — later answered in part by Shealy.

Coverage of State of Florida v. OpenAI (June 1, 2026), the first state attorney general action against OpenAI and Sam Altman, signaling escalating enforcement pressure on unsupervised consumer AI.