Disclaimer: This article is for educational purposes only. It does not provide legal advice, does not establish an attorney-client relationship, and should not be relied on for legal decisions. Always consult a licensed attorney regarding your specific case.
By mid-2024, the question of whether to use generative AI for legal work had moved from "experimental" to "operational" in most PI firms. Deposition summaries, which used to take a paralegal an hour per twenty-five pages of transcript, can now be produced in under five minutes by any competent off-the-shelf tool. The cost differential is significant; the quality, after some calibration, is comparable for routine deposition transcripts.
The ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512 on July 29, 2024 — its first comprehensive guidance on generative AI in law practice. The opinion establishes the framework against which AI use will be measured for the foreseeable future. For PI attorneys integrating AI into deposition workflow, Opinion 512 is now the operating document.
This article is for PI practitioners whose firms are evaluating or already using AI summary tools and who want a current read on where the practical risk surfaces lie.
What Opinion 512 Actually Requires
Opinion 512 does not impose new substantive duties — it applies existing Model Rules to the AI context. Five rules carry most of the weight.
Model Rule 1.1 (Competence). Lawyers must understand the capabilities and limitations of any GAI tool they use, and must periodically update that understanding. The opinion is explicit that competence is dynamic: a tool a lawyer competently used in 2024 may require renewed evaluation in 2026 as the model, the provider's terms, or the lawyer's use patterns change.
Model Rule 1.6 (Confidentiality). Lawyers must obtain informed client consent before inputting information relating to client representation into a "self-learning" GAI system — one that retains and learns from the inputs. The opinion draws a meaningful distinction between tools that use client data for training versus tools that segregate client inputs from training data and delete inputs after use.
Model Rule 1.4 (Communication). Clients are entitled to be informed of how AI is being used in their representation when that use materially affects the work. The opinion does not require disclosure of every AI use, but practitioner consensus is that significant AI involvement in deposition summarization or brief drafting falls within the disclosure obligation.
Model Rule 3.3 (Candor to Tribunal). Lawyers are responsible for the accuracy of any AI-generated content submitted to a court. Hallucinated citations, fabricated quotes, and invented case law are sanctionable regardless of whether the AI produced them.
Model Rule 5.1 / 5.3 (Supervisory Responsibility). Lawyers must supervise the use of AI tools by associates and non-lawyer staff. A paralegal who uses AI to generate a deposition summary that contains errors is the supervising attorney's responsibility.
Reasonable Fees (Model Rule 1.5). Lawyers may not charge clients for time spent learning generic AI tools. Time saved by AI cannot be billed as if it had been spent manually. This is the most underappreciated portion of the opinion in PI practice, where contingency-fee structures mute its direct application but where statutory fee shifting cases (civil rights, qui tam) make it directly applicable.
The Hallucination Problem
The most prominent failure mode for generative AI in legal work is the hallucination — a confident, well-formed assertion that is factually wrong. Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023), remains the cautionary case: counsel submitted a brief containing fictional case citations generated by ChatGPT, with confident-sounding but entirely fabricated holdings. Judge Castel imposed $5,000 in sanctions and the case became the standard reference for what not to do.
For deposition summaries specifically, hallucination risks are less acute than for brief writing but not absent. Documented failure modes:
Misattribution of testimony. AI summaries can attribute statements to the wrong witness or the wrong speaker in a multi-party deposition. The summary reads correctly; the attribution is wrong.
Invented quotations. AI summaries occasionally produce quoted text that the witness did not actually say but that captures what the AI assesses the witness "meant." The quotation marks make the testimony look verbatim when it isn't.
Fabricated exhibits. AI tools that lack access to the actual exhibit materials sometimes describe exhibit content based on inference from the deposition context. The descriptions can be plausible but inaccurate.
Smoothed contradictions. AI summarization tends to produce internally coherent text. When testimony contains genuine contradictions (which is often the point of a deposition transcript), the AI may smooth them out, losing the impeachment value.
The mitigation framework is verification, not avoidance. Every AI-generated summary that will be relied on for substantive purposes — settlement strategy, witness preparation, motion practice — should be cross-checked against the underlying transcript for the assertions that matter. The verification step is short relative to the time the AI saved on the initial summarization; the protection it provides is substantial.
The Privilege and Confidentiality Problem
The Confidentiality concern in Opinion 512 turns on a binary question about the AI tool: does it use client inputs to train its model, and does it retain those inputs after the work is done?
Consumer-grade tools (ChatGPT, Claude, Gemini consumer versions). These tools typically retain user inputs and may use them for model training. Some providers offer opt-outs through account settings; others do not. The opinion's default treatment of these tools is that they are "self-learning" and require informed client consent before client data is input.
Enterprise tools (ChatGPT Enterprise, Claude for Business, dedicated legal vendors). These tools typically include contractual protections against training on inputs, retention limits, and other confidentiality terms. The opinion is more permissive here, though it does not endorse any specific tool or provider.
On-premises and self-hosted models. Some firms have deployed open-source or licensed models that run on firm-controlled infrastructure. These present the lowest privilege risk but the highest implementation cost.
For PI practice, the practical guidance is:
- Do not paste deposition transcripts into consumer AI tools without informed client consent. The transcripts contain client information and are subject to confidentiality.
- For enterprise AI tools, review the data-handling terms. Specifically: does the provider retain inputs? For how long? Are inputs used for training? Is there a business associate or equivalent agreement governing privileged content?
- Inform clients (in engagement letters or by amendment) that AI tools may be used in their representation, with the safeguards in place. Most clients consent without question; the disclosure is what the opinion requires.
What AI Summaries Are Actually Good At
Within the verification framework, AI summarization adds genuine value to PI practice for:
Routine fact depositions. Witnesses on liability facts, treating physicians on routine medical care, corporate representatives on document custodianship. The summarization captures the core testimony reliably, and verification is fast because the substantive content is straightforward.
Initial triage of long transcripts. Transcripts running 200+ pages benefit from AI summarization as a first-pass tool. The lawyer identifies the sections that need detailed review, then reads those sections in the underlying transcript.
Cross-referencing testimony. Some AI tools now offer features that compare testimony across multiple depositions and flag inconsistencies. This is useful even where the AI cannot be trusted to characterize the inconsistencies accurately — it identifies the locations the lawyer needs to read carefully.
Chronological case timelines. AI tools that build timelines from multiple sources (depositions, medical records, business records) save substantial time on case organization. The lawyer verifies the dates and events; the AI does the assembly.
What AI Summaries Are Bad At
The categories where AI summaries should not be relied on without manual review:
Damages testimony. Specific dollar amounts, time periods, percentage limitations, and medical details are exactly the categories where AI errors most affect case value. Damages-related testimony should be reviewed manually.
Impeachment material. Inconsistencies that the AI smooths away are the precise material the lawyer needs at trial. AI summaries of contested testimony are unreliable for impeachment preparation.
Privileged or sensitive content. Beyond the confidentiality concern, AI summaries of testimony involving privilege assertions, settlement negotiations, or client communications should be reviewed manually because the AI may not handle the privilege framework correctly.
Expert testimony. Expert depositions involve technical content where the AI may misstate the expert's methodology, qualifications, or opinions. Manual review is appropriate for any deposition that will inform Daubert or expert challenges.
A Practical Risk Framework
For PI firms adopting AI summarization, a working framework has emerged that meets Opinion 512's requirements without unduly limiting the tool's utility:
- Tool selection. Use enterprise-grade AI tools with documented data-handling terms. Avoid consumer tools for any work involving client information.
- Engagement disclosure. Include AI use in the engagement letter or send a notice amendment. Obtain client consent for the disclosed uses.
- Workflow segmentation. Use AI for triage, organization, and routine summarization; reserve manual review for damages, impeachment, privilege, and expert content.
- Verification protocol. Establish a written verification protocol — what gets verified, by whom, against what source — and follow it consistently.
- Documentation. Note in the case file that AI was used for specific tasks and that verification was performed. This documentation supports any later inquiry about how the work was done.
- Periodic competence review. Periodically re-evaluate the tool's capabilities, limitations, and the provider's data-handling terms. Document the review.
The firms following this framework typically report 40-60% reductions in time spent on deposition summarization without measurable impact on quality, and without ethics exposure.
Frequently Asked Questions
Can I use AI to draft a brief for filing?
Yes, with the verification framework Opinion 512 requires. Every citation must be independently verified. Every factual assertion must be cross-checked against the record. The Mata v. Avianca lesson is that lawyers are responsible for the accuracy of AI-generated content regardless of how the content was produced.
Does ABA Opinion 512 apply in state court?
Opinion 512 interprets the ABA Model Rules of Professional Conduct. Each state's rules of professional conduct are derived from, but not identical to, the Model Rules. The opinion's reasoning will generally apply by analogy, though specific state ethics opinions may diverge. Several state bars have issued their own AI opinions since 2024, generally consistent with Opinion 512.
Should AI use be disclosed to opposing counsel?
The opinion does not require disclosure to opposing counsel. Some courts have adopted standing orders requiring disclosure of AI use in filings; those requirements should be checked at the case level.
What about AI tools that summarize testimony in real time during a deposition?
These tools raise additional concerns beyond Opinion 512's scope. Real-time AI summarization that operates during testimony may implicate the no-coaching rules addressed in ABA Formal Opinion 508 if the tool is being used to suggest answers or strategy to the witness or defending counsel. The defensible practice is to use real-time AI summarization, if at all, on the questioning side only and to disclose the practice if any party requests.
Closing
Generative AI is now part of the deposition workflow for the firms that can use it competently. Opinion 512 does not require firms to adopt these tools, but it does require that firms which use them do so within the ethical framework — competently, with appropriate confidentiality protections, with verification, and with documentation.
For PI practice, the tools' return on investment is real, but the verification step is non-negotiable. The firms gaining the most from AI are not the ones using it most heavily; they are the ones using it for the right tasks, verifying its output appropriately, and documenting their workflow well enough to defend it if questioned. The discipline of verification is the same discipline that has always characterized good deposition practice — and the firms that already had it transfer it to AI work naturally.