Your AI Notetaker Joined a Call With Opposing Counsel: What NYC Bar Opinion 2026-2 Requires

What Opinion 2026-2 actually says

The committee answered a narrow question: what should attorneys consider when using AI tools that record, transcribe, and summarize conversations with non-clients. Its answers are consent from all participants, no surreptitious recording, and non-recording as the default absent a specific good reason. It relies on Rules 1.1, 1.18, 3.4, 4.2, 4.3, and 8.4 ^[1].

The opinion is scoped to non-clients, which is the first thing worth noticing. Conversations with your own client raise confidentiality and informed-consent questions the profession has been discussing since Formal Opinion 512 ^[3]. Conversations with everyone else raise a different set, and those are the ones the AI notetaker walks into most often.

On consent, the committee is unambiguous: an attorney must disclose the intention to record and obtain permission from all participants ^[1]. That is a higher bar than the federal recording floor, and it applies as an ethical duty regardless of whether the applicable recording statute would permit one-party consent.

On surreptitious recording, the language is prohibitive rather than cautionary. Lawyers must not surreptitiously record conversations with any third parties ^[1]. A tool that joins quietly and transcribes without announcement is, on this reading, doing the thing the opinion forbids, and the fact that a vendor designed it that way does not transfer the duty.

On default practice, the committee goes beyond a permission framework. Absent a good reason to record a conversation in a particular instance, an attorney should avoid recording, because of the ethical and tactical risks involved ^[1]. That inverts how most firms currently operate, where recording is the default and switching it off is the exception.

The rules the committee draws on show how many separate duties a single transcript can implicate: competence under 1.1, prospective clients under 1.18, fairness to opposing parties under 3.4, communication with represented persons under 4.2, dealings with unrepresented persons under 4.3, and misconduct under 8.4 ^[1]. That breadth is the tell. This is not one narrow question about recording; it is several existing duties meeting a tool that captures conversations by default.

The auto-join problem: the tool decides before the lawyer does

The distinctive risk is that the recording decision is frequently made by configuration rather than by a person. Notetakers join from calendar invites, enable by default at the account level, and begin transcribing at the start of the meeting. By the time a lawyer considers whether this conversation should be recorded, it already is.

Most ethics guidance assumes an affirmative act: the lawyer decides to record and then does so. AI notetakers break that assumption. The common deployment pattern is an assistant connected to a calendar, configured once, that joins every meeting matching a rule.

That inverts the order of operations. The transcript begins, and the judgment about whether it should exist comes afterward if it comes at all. A lawyer who would never have reached for a recorder in a negotiation can end up with a verbatim record of it because a default was set months earlier by someone in IT.

Opinion 2026-2 does not treat that as an excuse, and the structure of the duties makes clear why it could not. The obligation to disclose an intention to record and obtain permission attaches to the attorney ^[1]. A configuration is not a person and cannot hold a professional duty, so the duty stays where it was.

The committee also flags a practical wrinkle that has no clean answer: the technology may lock people out of the discussion if they do not grant consent ^[1]. Some tools condition participation on accepting recording, which means a participant's refusal is not simply a decision to proceed unrecorded, it can be a decision not to participate. A consent framework that assumes declining is costless does not describe those products.

The operational consequence is that the meaningful control point is the configuration, not the meeting. A firm that wants to comply by asking lawyers to remember, in the moment, to disable a tool they did not enable has chosen the least reliable available control.

Two layers: the ethics rule and the recording statute

Consent under an ethics opinion and consent under a wiretap statute are separate questions with separate consequences. Federal law sets a one-party floor under 18 U.S.C. 2511(2)(d), permitting interception where one party consents, provided it is not for a criminal or tortious purpose. States may be stricter ^[2]. Satisfying one layer does not satisfy the other.

Lawyers reading an ethics opinion sometimes treat it as the whole answer. Here it is roughly half. The ethics layer governs professional discipline. The statutory layer governs criminal and civil liability, and it does not care what a bar committee concluded.

The federal floor is one-party consent. Under 18 U.S.C. 2511(2)(d) it is lawful to intercept a wire, oral, or electronic communication where one party consents, provided the interception is not for the purpose of committing a criminal or tortious act ^[2]. States may impose stricter requirements, and a number of them require the consent of every party.

The all-party jurisdictions are commonly enumerated at around a dozen, and several additional states apply different rules depending on whether a conversation happens in person or over the phone. This article does not reproduce a state-by-state grid, because those lists drift and because the in-person versus telephone distinction defeats simple tables. The Reporters Committee for Freedom of the Press maintains a state-by-state recording guide that is kept current and is the better place to check a specific jurisdiction ^[4]. Confirm your own state, and confirm it for the medium you are actually using.

Multi-party calls make this sharper than it first appears. A video meeting can include participants sitting in several different states, and the stricter rule can govern. A firm operating on the assumption that its home state's one-party rule covers every call is making a jurisdictional bet on every invitation it accepts.

One further point cuts against complacency in one-party states: recording a person who has a reasonable expectation of privacy in their communication can still violate the Wiretap Act. One-party consent is a floor with conditions, not a general permission.

Read together, the layers point the same direction even though they measure different things. The ethics opinion asks for all-participant consent as a professional duty ^[1]; the strictest applicable statute may demand it as a matter of law. A firm that adopts all-participant consent as a uniform practice satisfies both without having to resolve which one governs a given call.

Recording statutes vary by state, by medium, and by whether a participant has a reasonable expectation of privacy, and multi-state calls can implicate more than one regime. Nothing here is a determination about any jurisdiction's law. Confirm with counsel licensed where you practise.

Rules 4.2 and 4.3: who is on the other side of the transcript

The opinion's reliance on Rules 4.2 and 4.3 signals that the identity of the other participant changes the analysis. Recording a represented person implicates the rule governing communications with represented parties; recording an unrepresented person implicates the duty not to create a misleading impression about your role ^[1].

A transcript is not neutral. It captures one participant's words in a form the other participant controls, and it does so in a context where the lawyer's role carries specific obligations depending on who is speaking.

Where the other participant is represented, Rule 4.2 governs the communication itself, and adding a silent verbatim record does not make an otherwise improper communication proper. Where the participant is unrepresented, Rule 4.3 addresses the risk that a lawyer's role is misunderstood, and an undisclosed recording compounds precisely that risk: a person who does not appreciate they are speaking to an adversary is even less likely to appreciate that they are being transcribed ^[1].

Rule 3.4, fairness to the opposing party and counsel, sits alongside those. The concern is not confined to what is captured; it extends to the asymmetry of one side holding a verbatim record of a conversation the other believed was informal.

Rule 8.4 is the backstop the committee reaches for, and its presence in the list is the clearest signal about how the committee views quiet recording. Conduct involving dishonesty or misrepresentation is not cured by the fact that a tool performed it automatically.

The practical translation is that "who is in this meeting" is a question a lawyer has to answer before the meeting rather than after. A notetaker configured to join everything cannot make that distinction, because the distinction is legal rather than technical.

Rule 1.18 and the intake call

The sharpest case is the prospective-client conversation. Rule 1.18 protects information learned in a consultation even when no engagement follows, and the committee's inclusion of it is a signal that intake calls carry duties before anyone becomes a client. An AI notetaker on an intake call creates a durable record of exactly that protected material ^[1].

Firms tend to think of confidentiality obligations as beginning at engagement. Rule 1.18 exists because they begin earlier. Information learned during a consultation with a prospective client carries protection even if the firm never takes the matter.

That makes intake the highest-risk place to run an automatic notetaker, and often the place most likely to have one, because intake is exactly the workflow firms are keenest to automate. The transcript captures a person describing their legal problem in detail, and it exists whether or not the firm ends up representing them.

Two consequences follow. The first is retention: a firm now holds detailed protected information about people who are not its clients, in a system whose retention settings were probably chosen for convenience. The second is conflicts, because that same information can bear on the firm's ability to act against that person later.

The intake bot question is adjacent and worth separating. Our guide on running an AI intake bot addresses a tool the firm deliberately deployed to interact with prospective clients. The notetaker problem is different in kind: nobody chose it for this conversation, and it captures the human intake call the firm was already having.

Why "non-client" is broader than most lawyers assume

The opinion covers conversations with persons who are not clients, and that category is wide: opposing counsel, witnesses, experts, opposing parties, prospective clients, co-counsel at other firms, vendors, and court personnel. Very little of a litigator's calendar is a conversation with a client ^[1].

Firms reading this opinion sometimes assume it addresses an edge case. Look at an actual week. Meet-and-confers, witness interviews, expert calls, mediations, negotiations with opposing counsel, vendor demonstrations, co-counsel coordination, and intake calls are all conversations with non-clients.

The category also includes people whose participation is incidental. A witness's family member who joins a call, an assistant who dials in, a translator, or an accompanying professional are all participants whose consent the opinion's framing would reach, because it asks for permission from all participants ^[1].

That breadth is why a per-meeting judgment is unworkable as a control. Any policy that depends on a lawyer correctly classifying each conversation, under time pressure, before a tool that has already joined, will fail at some predictable rate.

The narrower and more reliable framing is to invert the default. Rather than asking which conversations should not be recorded, run without recording and ask which specific conversations justify it. That is the posture the committee describes, and it happens to be the one that survives a busy calendar ^[1].

When a good reason to record does exist

The opinion does not prohibit recording. It sets non-recording as the baseline absent a good reason in the particular instance ^[1]. Some conversations genuinely warrant a record, and the workable distinction is between a recording that is deliberate, disclosed, and consented to, and one that happens because a default was left on.

Treating the opinion as a ban would misread it and would also fail in practice, because there are conversations where an accurate record serves everyone. The committee's framing is about deliberateness, not prohibition.

Where the participants have agreed a record should exist, recording is straightforward under the opinion's framework: disclose the intention, obtain permission from all participants, and proceed ^[1]. The consent is the point, not the technology.

The tactical caution the committee raises is worth taking seriously alongside the ethical one. A verbatim record of a negotiation is discoverable in some circumstances, can be quoted back, and removes the latitude that informal conversation provides. Lawyers who would not bring a court reporter to a meet-and-confer should ask why an automatic transcript is different.

Accuracy is its own consideration under the competence rule the committee cites ^[1]. AI transcription and summarisation make errors, particularly with crosstalk, accents, technical vocabulary, and proper nouns, and an inaccurate summary of what opposing counsel agreed to is worse than no record. If a record matters enough to create, it matters enough to check against the audio rather than relied on as produced.

A protocol a firm can actually run

Control the configuration rather than the moment. Disable automatic joining, make recording an affirmative per-meeting act, script the disclosure, decide what happens when someone declines, and set retention deliberately, particularly for intake. Every one of these is a settings decision rather than a behaviour a busy lawyer has to remember.

The reason to build this as configuration is that the alternative has already been tested. Asking lawyers to remember to disable a tool that joins automatically produces compliance most of the time, and the failures land in exactly the conversations that matter most.

The steps below assume nothing about a specific product. They are the decisions a firm has to make regardless of which notetaker it uses, and they should be made once rather than per meeting.

1. Turn off automatic joining at the account levelThe recording decision should require an affirmative act by a person. A tool that joins from a calendar rule has made a professional judgment that only a lawyer can make ^[1].
2. Script the disclosure and use it verbatimThe duty is to disclose the intention to record and obtain permission from all participants. A scripted line said at the top of the call, before substance, is more reliable than improvisation and creates a consistent record that it happened ^[1].
3. Decide in advance what happens when someone declinesSome tools condition participation on consent, so a refusal may mean exclusion rather than an unrecorded meeting ^[1]. Know whether your tool behaves that way before you are asking a witness to agree.
4. Treat intake as the highest-sensitivity settingRule 1.18 protects prospective-client information whether or not an engagement follows. Decide deliberately whether intake calls are transcribed at all, and set retention for those transcripts on purpose ^[1].
5. Check the transcript before relying on itTranscription and summarisation error rates rise with crosstalk and technical vocabulary. If a record will be relied on, verify it against the audio rather than treating the summary as the record.
6. Set retention, and know where transcripts liveTranscripts of non-client conversations are a category of sensitive material most retention schedules never contemplated. Decide how long they persist, who can reach them, and whether they are inside the systems your confidentiality obligations already cover.

What this means for the firm's tooling decisions

The opinion effectively makes notetaker configuration a professional-responsibility question rather than an IT preference. The features that matter are whether automatic joining can be disabled, whether disclosure can be enforced, what happens when a participant declines, and where transcripts are stored and for how long.

Most firms selected a meeting platform and inherited a notetaker with it. That is a defensible way to acquire a scheduling tool and a poor way to acquire a system that captures conversations with opposing counsel and prospective clients.

The evaluation questions follow directly from the duties. Can automatic joining be disabled at the account level, or only per user? Can the tool be configured to announce recording rather than relying on a person to remember? What is its behaviour when a participant refuses consent? Where do transcripts and audio reside, for how long, and who inside and outside the firm can reach them?

The last question connects this to the vendor analysis the firm should already be running. A transcript of a witness interview is client-related confidential material sitting in a third-party system, which raises the same questions our guide on verifying a vendor's no-training claim works through. A notetaker is a legal AI vendor, whether or not it was procured as one.

There is also a record-keeping dimension that cuts in the firm's favour. If a firm adopts all-participant consent as a uniform practice, the evidence that it did so is worth keeping: a scripted disclosure, a consistent place it is logged, and a configuration state that can be shown. That is the same discipline this site argues for elsewhere, applied to a duty that is discharged in conversation rather than in a filing.

To be clear about the limits of all this: Opinion 2026-2 is a New York City Bar opinion, it binds no one, and other jurisdictions may reach different conclusions. What makes it worth acting on is not its authority but the fact that most firms have not made these decisions at all, and the tool has been making them by default in the meantime.