NYC Bar 2025-6 AI Recording: What It Requires Before You Record a Client
New York lawyers spent most of 2025 trying to figure out where the ethical line sits for AI notetakers. In December 2025, the New York City Bar Association's Committee on Professional Ethics answered a large part of the question. Formal Opinion 2025-6 addresses a lawyer's use of generative AI tools that record, transcribe, and summarize client meetings, and it sets out concrete duties around consent, confidentiality, competence, and supervision.
This article walks through what the opinion actually requires before you hit record, how it interacts with ABA Formal Opinion 512, and how to build a repeatable intake step that satisfies both. It is written for practicing New York lawyers, but the analysis applies broadly to any jurisdiction that follows the Model Rules.
What Opinion 2025-6 Actually Says
Opinion 2025-6 treats an AI notetaker as a form of technology the lawyer is deploying in the representation. That framing pulls in several familiar rules at once: confidentiality under Model Rule 1.6 and its New York analog, competence under Rule 1.1 (including technological competence), communication under Rule 1.4, and supervision of nonlawyer assistance under Rules 5.1 and 5.3.
The opinion's core move is to reject the idea that clicking "accept" on a consumer AI tool's terms is sufficient. A lawyer who feeds client information into a system that stores, trains on, or shares that data with third parties is making a disclosure decision, and that decision has to be made competently and, in most cases, with informed client consent.
Three practical requirements emerge:
- Understand where the data goes. The lawyer must have a working understanding of how the AI tool processes, stores, and transmits the recording and transcript, and whether any human or model training touches the content.
- Obtain informed consent where required. When the tool involves disclosure to a third-party provider, the client's informed consent is generally needed. Even where a strong argument exists that a vendor is the lawyer's agent under Rule 1.6(c), the safer course is to tell the client and log the answer.
- Supervise the output. AI-generated summaries are drafts. The lawyer remains responsible for what ends up in the file, the letter, or the pleading.
Why Consumer AI Recording Tools Create a Problem
The opinion is written against a backdrop of real cases showing what happens when lawyers or clients push privileged content into third-party AI platforms. In United States v. Heppner (S.D.N.Y. Feb. 2026), Judge Rakoff held that a litigant's chats with a public AI platform were not privileged, drawing on third-party doctrine reasoning. That is a defendant's problem, not a lawyer's, but the analytical move is identical when a lawyer routes client conversation through a cloud AI provider: the more the content sits on someone else's server, the harder the confidentiality and privilege story becomes.
Civil discovery cases reinforce the point. West Technology Group v. Sundstrom (D. Conn. 2024) treated cloud-stored materials as fair game for subpoena and production disputes when the party had control over them. And the Brewer v. Otter.ai putative class action, filed in 2025, put a spotlight on how bot-based transcription services capture and process meeting audio without every participant realizing what is happening.
Opinion 2025-6 does not name products, but the pattern it is worried about is clear: a lawyer joins a Zoom, an AI bot silently joins too, the audio is streamed to a vendor, a transcript and summary come back, and nobody in the room ever articulated to the client what just happened to their words.
The Consent Conversation, Concretely
The opinion does not prescribe magic words, but it does require that the consent be informed. That means the client needs enough information to make a real decision. In practice, a defensible consent conversation covers:
- That the meeting will be recorded and transcribed.
- What tool is being used and, at a high level, where the audio and transcript are processed and stored.
- Whether any third party (a cloud vendor, a model provider) receives the content.
- How long the recording and transcript are retained, and who inside the firm can access them.
- That the client can decline or ask you to stop at any time.
New York is a one-party consent state for recording under Penal Law § 250.00, so the criminal wiretap analysis is usually straightforward when the lawyer is a participant. Opinion 2025-6 is not about criminal law. It is about the professional-responsibility layer that sits on top: even where recording is legal, the ethics rules can still require you to tell the client and log the answer.
ABA 512 and NYC 2025-6 Side by Side
ABA Formal Opinion 512 (July 2024) was the first major national statement on generative AI and lawyer ethics. Opinion 2025-6 is narrower and more operational. The two overlap heavily, but they are not identical.
| Topic | ABA 512 (Jul 2024) | NYC Bar 2025-6 (Dec 2025) |
|---|---|---|
| Scope | Generative AI use by lawyers generally | AI recording, transcription, and summarization of client meetings |
| Confidentiality | Requires understanding of how the tool handles inputs; consent may be required | Same duty, applied specifically to audio and transcripts of client conversations |
| Client consent | Informed consent when disclosure to third party is involved | Informed consent generally required; document it |
| Competence | Reasonable understanding of benefits and risks of the specific tool | Same, with emphasis on data flow and vendor practices |
| Supervision | Lawyer remains responsible for AI output | Same; summaries are drafts, not final work product |
| Fees | Cannot bill AI time as attorney time | Not the focus, but consistent |
The practical takeaway is that a workflow built to satisfy 2025-6 will generally also satisfy 512. The reverse is not always true, because 512 is more permissive about how you get to informed consent.
A Pre-Recording Checklist for New York Lawyers
Here is a working checklist that maps to the opinion. It is not a substitute for reading 2025-6 in full, but it captures the moves a lawyer should be able to point to if asked.
- Vendor diligence, once. Before adopting any AI notetaker, document where audio and transcripts go, whether the vendor trains on your data, retention defaults, and whether a DPA is available. Keep this on file.
- Matter-level decision. Decide whether AI recording is appropriate for this matter. Some matters (grand jury prep, certain criminal defense conversations, sensitive internal investigations) may warrant not recording at all.
- Client disclosure at engagement. Address AI recording in the engagement letter or in a standalone consent form. Explain the tool, the data flow, and the client's right to decline.
- At-the-meeting confirmation. Confirm on the record, in the first minute, that everyone present consents to the recording. Log it.
- Label the output. Mark the transcript and summary as Privileged and Confidential, and store them in the matter file, not in a general-purpose cloud drive.
- Review before use. Do not forward an AI summary to the client or to opposing counsel without reading it. Correct hallucinations. Delete anything that should not be in a written record.
- Retention. Set a retention policy that matches your document-retention schedule for the matter type.
Where Architecture Matters More Than Paperwork
Opinion 2025-6 is a rules document, but the underlying risk it is trying to manage is architectural. A recording that never leaves the lawyer's device is a fundamentally different disclosure problem from a recording that is streamed to a vendor's servers. The consent conversation is easier when there is less to disclose. The subpoena analysis is easier when there is no third-party custodian to serve. The Heppner third-party-doctrine reasoning has less traction when there is no third party in the loop.
None of this eliminates every risk. Privilege can still be waived by conduct, by disclosure to non-privileged parties in the room, or by careless handling of the transcript. But the risk profile of an on-device workflow is meaningfully different from the risk profile of a cloud-bot workflow, and the opinion's competence duty pushes lawyers to understand that difference before choosing a tool.
What About Virtual Meetings and Bots?
The opinion is technology-neutral, but the practical concern is loudest for virtual meetings. A bot that joins a Zoom as a participant is visible to everyone, which helps with disclosure, but it also streams the entire meeting audio to a third-party service. The Brewer v. Otter.ai litigation illustrates how that model can go wrong when participants do not fully understand what the bot is doing.
An alternative is on-device capture of the meeting audio on the lawyer's own machine, with no bot joining the call. That approach still requires disclosure and consent under 2025-6, but it removes the third-party vendor from the data flow. For a comparison of the two capture models, see our internal write-up on bot-based versus on-device meeting capture.
Common Mistakes to Avoid
- Assuming the engagement letter is enough. A one-line reference to "technology" buried on page four is not informed consent for AI recording. Be specific.
- Recording without confirming on the record. A verbal confirmation at the start of the meeting is cheap insurance and creates a clean audit trail.
- Forwarding raw summaries. AI summaries can misattribute statements and invent details. Review before anything leaves your desk.
- Ignoring third parties in the room. If a non-client sits in on the meeting, think about the privilege implications before you record, not after.
- Using consumer chatbots for follow-up analysis. Pasting a transcript into a public AI platform reintroduces the third-party doctrine problem Heppner flagged.
How Basil Approaches This
Basil is built as an on-device AI notetaker. Audio capture, transcription, and summarization all run locally on the Apple Neural Engine. Nothing is uploaded, there is no server in the loop, and Basil itself never receives user data. For virtual meetings, Basil uses on-device capture on macOS rather than sending a bot into the call.
The Basil for Law edition, launching August 2026, adds features designed around opinions like 2025-6: a Privileged and Confidential label applied to transcripts and summaries, a consent log tied to each meeting, matter organization, and a privilege attestation flow. Pricing for solo attorneys is $19.99 per month or $199.99 per year, with a 3-day trial on the monthly plan and a 7-day trial on the annual plan. The general Basil app is available today with a 60-minute-per-month free tier. Basil signs DPAs and NDAs on request.
None of that removes the lawyer's duty to have the consent conversation, document it, and review the output. What it does is reduce the number of third parties in the data flow, which is the part of the problem that paperwork alone cannot solve.
This article is for information only and is not legal advice.
Frequently asked questions
Does NYC Bar Opinion 2025-6 ban AI notetakers?
No. The opinion permits AI recording and transcription of client meetings, but it requires the lawyer to understand how the tool handles the data, obtain informed client consent where a third party is involved, and supervise the output. It rejects the idea that clicking accept on a consumer tool's terms is enough.
Is a general reference to technology in my engagement letter enough consent?
Probably not. Opinion 2025-6 requires informed consent, which means the client needs enough specific information about the AI tool, where the data goes, and how long it is retained to make a real decision. A generic technology clause buried in the engagement letter is unlikely to satisfy that standard on its own.
Do I need consent if the AI runs entirely on my own device?
You should still disclose and obtain consent. On-device processing changes the confidentiality risk profile because there is no third-party vendor receiving the content, but the client's right to know they are being recorded and transcribed does not go away.
How does 2025-6 interact with ABA Formal Opinion 512?
They are consistent. ABA 512 (July 2024) sets out general duties for lawyers using generative AI, including competence, confidentiality, and supervision. NYC 2025-6 applies those same duties to the specific context of AI recording and summarization of client meetings, with more operational detail on consent and vendor diligence.
What happens if I use a consumer chatbot to analyze a client transcript?
That reintroduces the third-party doctrine concern highlighted in US v. Heppner (S.D.N.Y. Feb. 2026), where a litigant's chats with a public AI platform were held not privileged. Pasting a client transcript into a public AI tool creates disclosure and privilege risk that the on-device recording did not have.
Is New York a one-party consent state for recording?
Yes, under Penal Law § 250.00, one-party consent generally satisfies the criminal wiretap statute when the lawyer is a participant. Opinion 2025-6 is a separate ethical layer: even where recording is legal, the professional-responsibility rules can require disclosure to and consent from the client.
Keep client conversations on your device
Basil transcribes and summarizes entirely on-device — no cloud, no bot, no server to subpoena. See Basil for Law → · Legal-tool reviews →
This article is for information only and is not legal advice.