AI Discovery Privilege

State of play.

• The federal court split on AI privilege has deepened and now has a state-court dimension. Heppner (SDNY) treats consumer AI use as a third-party disclosure destroying privilege; Warner v. Gilbarco (E.D. Mich.) protects pro se ChatGPT use as work product; and Assini v. Hayward (N.Y. state court) now holds that attorney-crafted prompts and outputs qualify as opinion work product — three courts, three positions, no circuit resolution .
• The operative variable is attorney direction, not tool identity. The emerging doctrinal axis — attorney-directed use triggers protection, client-initiated use does not — is the closest thing to a working framework, but it has not been tested above the district court level and the Heppner court explicitly reserved the attorney-direction question .
• Agentic AI in digital forensics is creating a new Daubert pressure point. Exterro's ARMOUR for FTK launch — AI-driven remote endpoint investigation with automated reasoning traces — puts AI-generated forensic evidence directly in the path of admissibility challenges, before courts have established what transparency is sufficient .
• Major firms are institutionalizing AI across discovery and litigation functions. Reed Smith's dedicated AI department, with 700-plus lawyers on Harvey daily and firm-wide associate certification, signals that AI integration is now a competitive differentiator rather than an experiment — and that clients will increasingly expect integrated legal-technical guidance (→ Reed Smith Launches AI-Focused Legal Services Team Led by Philadelphia Partner).
• For counsel advising litigation clients or enterprise AI deployers, the practical baseline is that the attorney-direction distinction is the best available shield but remains unvalidated at the circuit level, making platform architecture — closed versus consumer-facing, data retention terms, counsel control of prompts — the load-bearing variable in every active matter.

Where things stand.

• The Heppner framework remains the anchor but is no longer the only data point. Judge Rakoff's three-ground ruling — AI is not a lawyer, platform ToS eliminate confidentiality, materials were not counsel-directed — is the most-cited federal authority, but Warner v. Gilbarco and Assini v. Hayward have created a genuine doctrinal contest rather than a single governing rule .
• Attorney-directed AI use is now the strongest available work-product argument. Assini treats counsel-crafted prompts as opinion work product — the highest protection tier — while Heppner and Assini together draw a line between attorney-controlled and client-initiated AI interactions .
• The open/closed system distinction is judicially operative. Morgan v. V2X (D. Colo.) and Jeffries v. Harcros (D. Kan.) require AI-specific protective order language distinguishing open/public tools from enterprise systems with contractual data-retention prohibitions — a distinction absent from most existing protective order templates .
• AI-generated materials and prompts are treated as standard ESI under FRCP 26. Courts have applied relevance and proportionality analysis without AI-specific exemptions; In re OpenAI Copyright Infringement Litigation compelled production of millions of anonymized user prompts under ordinary discovery rules .
• Generic AI governance policies do not satisfy eDiscovery defensibility requirements. Practitioners and vendors have identified that deploying AI into discovery without tool-specific controls over access, output verification, and decision documentation creates challenges that firm-wide AI policies cannot resolve .
• Agentic AI forensics tools are entering the evidentiary chain. Exterro's ARMOUR for FTK generates automated reasoning traces for remote endpoint investigations — positioning AI-generated forensic evidence for Daubert scrutiny before courts have defined what auditability is sufficient .
• England and Wales courts are moving in parallel. A UK Upper Tribunal ruling has found that uploading documents to public AI systems constitutes privilege waiver, and the judiciary's PD 57AD working group is actively soliciting industry feedback on AI disclosure parameters .
• No federal statute or rule amendment has codified AI privilege or discoverability standards. Proposed Rule of Evidence 707 addresses AI evidence reliability but has not been enacted; the field is developing entirely through case-by-case judicial interpretation .

Latest developments.

• Heppner versus Warner v. Gilbarco — two federal courts reach opposite conclusions on whether consumer AI use destroys privilege, with the disagreement now documented as a formal circuit-level split in waiting .
• Assini v. Hayward (N.Y. state court, June 4, 2026) holds that attorney-crafted AI prompts and outputs are opinion work product, directly contradicting Heppner and establishing the attorney-direction variable as the doctrinal fulcrum .
• Exterro launches ARMOUR for FTK, an agentic AI layer for digital forensics that generates auditable reasoning traces for remote endpoint investigations — raising Daubert admissibility questions courts have not yet addressed .
• Reed Smith establishes a dedicated AI department combining lawyers, eDiscovery specialists, and AI professionals, with 700-plus lawyers on Harvey daily and firm-wide associate certification (→ Reed Smith Launches AI-Focused Legal Services Team Led by Philadelphia Partner).
• Practitioner commentary identifies that generic AI governance frameworks fail in eDiscovery, where defensibility requires tool-specific controls over access, verification, and documentation — not firm-wide policy language .
• ESI management guidance emphasizes email threading, date filtering, and Early Case Assessment databases as foundational pre-AI steps that reduce discovery costs without requiring complex tooling .

Active questions and open splits.

• Attorney direction as the privilege trigger — validated or not? Assini and the Heppner/Warner split converge on attorney direction as the operative variable, but no circuit court has tested it; the question of whether counsel-supervised AI use categorically preserves privilege remains open .
• Third party or neutral tool — which framing governs work product? Heppner treats Claude as a third party destroying confidentiality; Warner v. Gilbarco treats ChatGPT as a neutral tool analogous to a word processor; Assini focuses on attorney mental impressions without resolving the tool-identity question. The split is unresolved at the circuit level .
• What is the scope of waiver when a client independently uses a public AI tool with privileged communications? Heppner ordered disclosure of AI outputs and the underlying attorney communications; whether the waiver extends to all communications on the same subject matter is unresolved .
• What contractual terms satisfy the Morgan v. V2X enterprise AI standard? The D. Colo. model requires contractual prohibitions on data storage, model training, and third-party sharing with on-demand deletion — but no court has yet tested whether specific enterprise agreements (Azure GPT-4, Harvey, and similar) satisfy that standard .
• Will agentic AI forensic evidence survive Daubert? ARMOUR for FTK generates reasoning traces and auditable records of investigative steps — but courts have not defined what level of AI transparency satisfies Daubert's reliability requirements for AI-generated forensic evidence .
• Do existing protective orders require amendment in every active matter? The structural gap in most model orders — no AI tool or data-architecture language — raises whether counsel have an affirmative obligation to seek amendment when AI is in use, and what the consequence of failing to do so is .
• How will courts treat AI use in internal investigations? Companies routinely use AI to process documents before counsel engagement; whether those outputs are discoverable and whether privilege can attach is unaddressed by current case law .

What to watch.

• Any Second Circuit or other appellate court taking up the attorney-direction question or the third-party-versus-neutral-tool split — that ruling will set the framework for all subsequent AI privilege disputes and render the current district-court patchwork obsolete.
• Whether courts begin requiring disclosure of AI tool use and platform terms of service as part of Rule 26(f) conferences, formalizing what is currently an informal practice.
• First Daubert challenges to AI-generated forensic evidence from agentic tools — the outcome will define what auditability courts require and whether reasoning-trace architecture is sufficient.
• The England and Wales PD 57AD working group's publication of findings and proposed rule amendments — the first formal judicial rulemaking on AI disclosure parameters in any major common law jurisdiction, with potential persuasive authority in US courts.
• State bar ethics opinions on attorney obligations when clients have already used public AI tools with privileged information — a gap bar counsel are beginning to address.
• Whether additional major firms replicate Reed Smith's integrated AI department model, accelerating client expectations for combined legal-technical AI governance counsel.