Does sending a deposition transcript to ChatGPT waive privilege?

The waiver question depends on the specifics of the disclosure, the contractual protections in place with the AI provider, and the jurisdiction's privilege rules. Where the AI provider's terms allow the provider to use the input for ongoing model training or for purposes beyond responding to the immediate request, the firm faces a stronger argument that confidentiality has not been preserved. Where the AI is operating under enterprise terms that include no-training and limited-purpose-use commitments, the firm has a stronger defense. The architectural problem is that the firm has to be able to produce evidence of which AI was used, on which terms, and what data was sent. Without per-decision records, the firm's response rests on policy assertion rather than evidence.

What is the difference between ABA Opinion 512 and a state bar opinion?

ABA opinions are advisory and interpret the Model Rules. State bars adopt the Model Rules with state-specific variations and issue their own opinions interpreting their adopted rules. A lawyer's actual obligations come from the rules of the jurisdictions in which the lawyer is admitted and the bars whose rules govern the lawyer's conduct. The ABA opinion shapes the discussion but the state-level rule governs. Firms practicing across multiple states have to track the variations, particularly where state opinions have departed from or elaborated on the ABA framework (California, Florida, New York, and Pennsylvania have each added specific elaboration).

How does the inspection layer support litigation hold compliance for AI prompts?

The inspection layer captures, per request, the AI input and the AI output along with the timekeeper, matter, and timestamp. Where a litigation hold is active on a matter, the records on that matter are preserved as part of the hold cohort. The records exist independent of whether the AI provider retained the prompt or response on its side. The firm's preservation obligation is met by the firm-side records, which the inspection layer produces automatically.

Is the inspection layer compatible with legal-specific AI platforms like Harvey or CoCounsel?

Yes. The inspection layer operates at the HTTP AI request boundary, including the calls a legal-specific platform makes to its underlying foundation models on the firm's behalf. Where the legal platform's architecture exposes the prompt traffic to the firm's environment (the typical pattern for enterprise deployments of these platforms), the inspection layer captures the prompts and responses with the platform's identity attached. Where the platform is fully managed and the prompts never traverse the firm's environment, the inspection layer captures the firm-side calls into the platform's API, which still produces the per-decision evidence the bar opinions expect.

What is the inspection overhead for high-volume document review AI?

Inline inspection adds under 50 milliseconds per request in production benchmarks against an LLM inference baseline of 500 milliseconds to 5 seconds. For document review workflows processing thousands of documents per matter, the overhead is invisible relative to the inference time and the human review time. For agent-based research workflows where a single user prompt drives a sequence of model calls, the inspection latency is per-call but the per-call budget remains well under the per-call inference cost.

Law Firm ChatGPT Confidentiality: ABA Opinion 512 and the Architecture Privilege Survives

The American Bar Association Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512 on July 29, 2024, addressing lawyer use of generative AI tools under the Model Rules of Professional Conduct. The opinion identifies the operational obligations on the lawyer using the AI: competence under Rule 1.1, confidentiality under Rule 1.6, supervision of nonlawyer assistance under Rule 5.3, communication with the client under Rule 1.4, and reasonable fee determination under Rule 1.5. State bar opinions from California, Florida, New York, and Pennsylvania add jurisdiction-specific elaboration. The firm that wants to operate AI defensibly faces the same architectural question every regulated industry faces: can the firm produce a per-decision record of what the AI received and what the firm did with the output, on examination, on the time horizon a disciplinary authority or a court expects.

I want to walk through what ABA Opinion 512 actually requires, where law firm AI deployments are exposed, and the inspection architecture that supports the firm's position under bar examination, court inquiry, and client audit.

What ABA Formal Opinion 512 says

Opinion 512 places generative AI within the existing Model Rules rather than creating a new regime. The opinion is operational: it identifies the specific failure modes that surface when lawyers use AI without adequate process.

On competence (Rule 1.1), the lawyer is expected to maintain a reasonable understanding of the benefits and risks of the AI tool being used. The understanding has to be sufficient to identify when the AI's output is unreliable, when the tool is being used outside its competence, and when the workflow requires additional verification before relying on the output. The Mata v. Avianca matter and subsequent state-level discipline cases for fabricated citations all attach here.

On confidentiality (Rule 1.6), the lawyer is expected to make reasonable efforts to prevent the disclosure of, or unauthorized access to, information relating to the representation of a client. Confidential information sent to an AI tool that retains the input for training, or that processes the input through a model the firm has no contractual constraints on, raises the prohibited disclosure question. The opinion notes that "self-learning" AI tools that incorporate user input into ongoing training datasets are a particular concern.

On supervision (Rule 5.3), the lawyer is expected to make reasonable efforts to ensure that nonlawyer assistance conforms to the lawyer's professional obligations. The opinion treats AI tools as a form of nonlawyer assistance for Rule 5.3 purposes. Supervision includes vetting the tool, training users, monitoring outputs, and maintaining policies on AI use.

On communication (Rule 1.4) and fees (Rule 1.5), the opinion addresses whether and when the lawyer has to inform the client about AI use, and how AI-related costs are billed. The disclosure obligation is contextual; the fee obligation requires that billed hours reflect actual lawyer effort and that costs passed through reflect actual costs.

What state bar opinions add

Multiple state bars have issued opinions or guidance elaborating on the Model Rules in the AI context:

California State Bar Standing Committee on Professional Responsibility issued Practical Guidance for the Use of Generative AI in November 2023, addressing confidentiality, competence, billing, and candor.
Florida Bar Ethics Opinion 24-1, issued January 19, 2024, addresses confidentiality obligations when using AI and requires that the lawyer obtain informed consent if confidential information is to be input into an AI that lacks adequate confidentiality protections.
New York State Bar Association Task Force on Artificial Intelligence published a report in April 2024 covering recommendations for lawyer use of AI including disclosure expectations.
Pennsylvania Bar Association Formal Opinion 2024-200 addresses AI use across multiple rule dimensions.

The state-level opinions differ in their specific recommendations but converge on the same operational expectations: the firm has to know what data is being sent to which AI, the firm has to have a defensible answer to the question of whether confidentiality has been preserved, and the firm has to be able to evidence its supervision program.

Where law firm AI deployments are exposed

The exposure pattern across law firms running document review, research, drafting assistance, and client intake AI repeats:

The associate or paralegal is using ChatGPT, Claude, Copilot, Gemini, or a legal-specific tool (Harvey, CoCounsel, Lexis+ AI, Westlaw Precision AI) to assist with a research, drafting, or document review task. The data sent to the AI includes work product, client information, deposition transcripts, contract drafts, or other materials covered by Rule 1.6 confidentiality. The firm has a policy on AI use but cannot produce, per matter, the evidence of what the AI received and what the timekeeper did with the output. The supervision program is a memo. The evidence base is missing.

The matter has a specific privilege or work-product overlay (anticipation of litigation, attorney-client privilege over the communication itself, joint defense privilege). The data sent to an AI tool that the firm does not have the contractual protections for could be argued by an opposing party as a waiver. The firm's response on a motion to compel turns on what the firm can show about the data the AI received and the protections that attached. Without per-decision records, the firm's response is testimony rather than evidence.

The firm has a Business Associate Agreement with a healthcare client. The AI tool the associate used does not have the BAA terms in place. The PHI in the deposition transcript that went into the AI tool created an exposure the firm cannot now reconstruct.

The litigation hold is in place on the matter. The AI prompt and the AI response are responsive to the hold but were not preserved as electronically stored information. The hold has been breached without the firm's awareness.

Shadow AI in law firms tracks the general enterprise pattern. The 78% employee unauthorized AI use figure from Cloud Radix applies to law firm staff, who have direct access to the most sensitive client data the firm holds. The traditional firm DLP product cannot inspect the prompt traffic because it runs underneath the TLS encryption and the prompts travel as opaque POST bodies.

The inspection architecture that supports the firm's position

The architecture has the same shape as the architecture for healthcare HIPAA or for financial services DORA. Legal-services calibration adds matter-level identity, privilege flags, and matter retention rules.

Inline inspection sits at the HTTP AI request boundary between the firm's authenticated users and any LLM endpoint, including the legal-specific platforms (Harvey, CoCounsel, Lexis+ AI, Westlaw Precision AI) that themselves call the underlying foundation models. The inspection includes detection for the data classes relevant to legal practice: client PII, PHI for healthcare client matters, attorney work product markers, privilege markers, deposition content, and the matter identifiers that connect requests to specific representations.

Identity attribution names the timekeeper behind each request, links the request to the matter number, and where applicable to the engagement letter or BAA in effect for that matter. For agent-based workflows (an AI agent invoked through a platform that runs research or drafts pleadings on behalf of a timekeeper), the agent identity and the delegated authority appear in the record alongside the timekeeper identity.

Per-decision audit records create the evidence base the firm needs across the use cases:

For the disciplinary inquiry, the firm produces evidence of what AI tools were used in the representation, what client data was sent, and what supervision steps were taken.
For the motion to compel or the privilege challenge, the firm produces evidence of what data the AI received, what the firm's policy was at the time of the request, and what contractual protections applied to the destination AI.
For the litigation hold, the firm produces the AI inputs and outputs as preserved ESI on the matter's hold cohort.
For the client audit (most often for institutional clients with outside counsel guidelines requiring AI controls), the firm produces evidence of how the client's matters were processed and what tools touched the client's data.

Retention runs at the longer of the firm's matter retention schedule (typically the duration of the engagement plus a regulatory tail), the jurisdiction's record-keeping rules for legal services, and the client's outside counsel guidelines.

Where this connects to outside counsel guidelines and client expectations

Institutional clients in financial services, healthcare, energy, technology, and government increasingly include AI-specific provisions in outside counsel guidelines. The provisions typically require:

Disclosure of AI tools used on the client's matters.
Specific contractual protections (no model training on client data, geographic data location, audit rights).
Evidence of supervision and policy enforcement on AI use.
Notification on AI-related incidents touching the client's data.

The architecture that satisfies ABA Opinion 512 produces records that contribute to outside counsel guideline compliance, client audit response, and the evidence base for incident notification. The infrastructure is shared. The vocabulary differs by client.

The cross-border practice adds the EU AI Act overlay for firms operating in or for EU clients. While legal services per se are not in Annex III, AI use in client-facing decisions (creditworthiness in finance practices, employment decisions in HR practices) may bring high-risk obligations through the client's regime. The firm's records support the client's deployer obligations under Article 26.

DeepInspect

This is the architecture DeepInspect was built to provide for law firm AI compliance. DeepInspect sits inline between the firm's authenticated timekeepers and any HTTP-based LLM endpoint, including the legal-specific platforms that wrap the underlying foundation models. The inspection includes detection for client PII, PHI for healthcare client matters, work product markers, privilege markers, deposition content, and the matter identifiers that connect AI activity to specific representations.

Every request produces a per-decision audit record containing timekeeper identity, matter number, timestamp, data class, policy version, outcome, and a tamper-evident signature. The records support the firm's position under ABA Opinion 512 examination, state bar inquiry, motion to compel or privilege challenge, litigation hold compliance, outside counsel guideline audit, and incident notification. For managing partners, general counsel of the firm, and CISOs facing the cross-pressure of associate AI adoption and institutional client AI requirements, the inspection layer is the architectural component that produces the evidence the firm's professional and contractual obligations expect from the same infrastructure. Book a demo today.