The Consent Trap: Why AI Notetakers Create Ethical Exposure Before You Say a Word

The consent problem starts before you ever click "agree." AI notetakers are now being integrated into Zoom and Teams at what legal analysts describe as warp speed — often auto-activated and hidden, recording meetings without the attorney or client knowing a word was captured. For a criminal-defense attorney, that silent activation is where the legal exposure begins.

Roughly 35 states follow one-party consent, but about 11 require all-party consent — including California, Florida, Illinois, Maryland, and Pennsylvania. If an AI tool records a call with a client or co-counsel in one of those jurisdictions without every participant's informed consent, the exposure isn't just an ethics grievance. Recording without all-party permission can carry criminal or civil liability, and Illinois guidance is blunt that undisclosed recording is inconsistent with the candor lawyers owe clients and conflicts with IRPC Rule 8.4's ban on deceptive conduct.

Consent to "recording" isn't the end of the analysis, either. Traditional consent frameworks don't answer the AI-specific questions: whether transcripts feed model training, who owns the data, and how the vendor retains it. And the privilege risk is real — as Duane Morris attorneys warn, automatic transcription of meetings where sensitive legal strategy is discussed risks exposure to third-party vendors, making vendor due diligence essential to preventing accidental privilege waiver. AI transcripts retained as business records may become discoverable, and they rarely contain attorney impressions that would shield them as work product.

Illinois adds a second layer: under BIPA, if an AI tool processes biometric identifiers like voiceprints, strict notice, consent, retention, and destruction requirements apply.

Defenders actually face two distinct consent scenarios, and the ethics analysis differs for each:

• Client meetings and strategy calls: the conversation is privileged, so the dominant risks are consent statutes, privilege waiver through the vendor, and a chilling effect — clients who know they're being transcribed may withhold sensitive facts, producing weaker advocacy and faulty advice.
• Evidence transcription: bodycam, jail calls, 911 audio, and interrogation footage are already recorded by the state and turned over in discovery. The consent question is largely answered; the duty shifts to accuracy and verification of the transcript you'll rely on.

That second scenario is where NYC Bar Formal Opinion 2025-6 bites hardest: lawyers must independently review AI-generated transcripts for accuracy before relying on them, because errors "can mischaracterize client statements and create content wholly absent from the actual conversation." It's the principle behind how we built Evidence Copilot — every AI output is marked AI-generated and pending human review, and the original recording stays immutable as the authoritative record. When you evaluate any transcription tool for evidence review, require the same: visible human-review flags, citations you can check against the source, and an audit trail.

What the Bar Guidance Actually Says: Informed Consent Is the Baseline, Not the Ceiling

The ethics guidance is no longer scattered — it's converging. In December 2025, the NYC Bar issued Formal Opinion 2025-6 addressing AI notetakers, and Illinois regulators, bar-association practice-management groups, and research-ethics scholars have landed on the same baseline: consent to AI transcription must be explicit, documented, and AI-specific — never a passive default.

The baseline is clear enough. Under Illinois' eavesdropping statute, recording private conversations requires informed consent of all participants, and secret activation of an AI notetaker is "inconsistent with the candor and honesty lawyers owe clients" — a conflict with IRPC Rule 8.4's ban on deceptive conduct, per the Illinois Supreme Court Commission on Professionalism. Roughly 11 states, including California, Florida, and Pennsylvania, require all-party consent, with criminal or civil exposure for violations.

But the guidance goes further than "we're recording." Traditional recording consent doesn't cover how the AI actually processes your words — whether it merely transcribes or runs sentiment and keyword analysis, whether transcripts feed vendor training models, or who owns the data afterward. Scholars publishing in the Journal of Empirical Research on Human Research Ethics recommend consent language that names the AI service and covers retention and training use. The ARDC's Illinois Attorney's Guide to Implementing AI adds a three-step framework: classify the information's sensitivity, determine whether the tool is internal or third-party, and evaluate vendor safeguards.

Two complications deserve attention:

• The chilling effect. Commission Chief Counsel Mark C. Palmer warns that recording can erode trust — attorneys stop brainstorming aloud, clients withhold sensitive facts, and the result is "weaker advocacy and faulty advice."
• The client-side problem. The most novel part of the NYC Bar guidance, as bar practice-management analysis notes, is clients bringing their own AI notetakers. The opinion recommends engagement letters requiring advance notice and stating that client-generated recordings aren't binding unless promptly shared for independent review.

There's also an honest counterpoint: the JERHRE editorial's authors argue that "AI exceptionalism" — demanding AI disclosure while ignoring equivalent human risks, like a human transcriber breaching confidentiality — can shift burdens unfairly. Their conclusion: responsibility is shared, but the party deploying the tool remains accountable.

Consent, in short, is the floor, not the ceiling. It doesn't address the five AI-specific ethical layers — transparency, data ownership, bias, accuracy, and psychological safety — and it doesn't excuse you from the NYC Bar's requirement to independently review AI-generated transcripts before relying on them, because errors "can distort the advice of counsel" and "create content wholly absent from the actual conversation." That's the principle we built into Evidence Copilot: every AI output is marked AI-generated and pending human review, with the original recording kept immutable as the authoritative record. Tools that flag AI output for review — rather than presenting it as finished work — make the ethics compliance largely automatic.

Human Review Is Now Mandatory: How to Vet Any Transcription Tool Before You Consent

The ethics rules have caught up: if you rely on an AI transcript without reading it yourself, you're not delegating work — you're gambling on it. NYC Bar Formal Opinion 2025-6, issued in December 2025, now requires lawyers to "independently review AI-generated transcripts and summaries for accuracy" before relying on them, because errors can mischaracterize client statements or create dialogue that never happened at all.

That requirement matters double in criminal defense, where a misattributed line in a jail call transcript or a garbled speaker designation in an interrogation recording can shape a suppression motion or a cross-exam. A hallucinated quote isn't a typo — it's a fabricated record of what your client said. The Duane Morris analysis warns that misheard words, incorrect speaker designations, or hallucinated dialogue can create inaccurate records and even malpractice exposure if you treat a transcript as the authoritative account of client instructions.

So before you consent to AI transcription — of a client meeting, or of the bodycam and 911 audio in your discovery — vet the tool first. The ARDC's Illinois AI implementation guide lays out a three-step framework: classify the sensitivity of the information, determine whether the tool is internal or third-party, then evaluate the vendor's safeguards. Applied to transcription, that checklist looks like this:

• Read the terms of service and data retention policy. Confirm where recordings and transcripts are stored, how long they're kept, and who can access them — third-party storage of privileged strategy sessions risks accidental privilege waiver.
• Verify your data isn't used to train AI models. Some third-party notetakers retain transcripts for training newer models; that should be a dealbreaker for anything touching client communications or case evidence.
• Require visible AI-generated flags on output. Every AI-produced transcript or summary should be marked as such and pending human review — if the tool presents AI output as finished work, it's inviting you to skip the verification step the ethics opinions now mandate.
• Insist the original recording stays the authoritative, immutable record. The transcript is a working aid; the audio and video are the evidence. If a tool modifies originals or can't show its work, don't consent to it.

This is the design logic behind Evidence Copilot: every AI answer is marked AI-generated and pending human review, every citation points to the exact second of video or page of PDF, and originals are never altered. The tool should make human review easy, not optional.

One counterweight belongs in your calculus: transcripts aren't purely risk. The same NYC Bar guidance notes that transcripts help clients who are deaf, hard of hearing, or facing language barriers participate more fully in their own defense. The answer isn't to refuse AI transcription — it's to consent with eyes open, vet the vendor, and read what the machine wrote before you put it in front of a judge.

Putting It Into Practice: Consent Language, Engagement Letters, and a Defense Workflow That Survives Scrutiny

The gap between "we agreed to record" and "we agreed to let a third-party AI process this recording" is where defense teams get burned. Consent language, engagement letters, and your review workflow need to close that gap explicitly.

Start with the consent clause itself. Roughly 11 states — including California, Florida, Illinois, Maryland, and Pennsylvania — require all-party consent before recording, and recording without it can carry criminal or civil liability. So your consent language should do more than get a yes. It should name the specific AI transcription service, link its data governance and retention statement, disclose whether transcripts feed model training, and offer a manual-transcription opt-out for clients who decline. Research-ethics scholars make the same recommendation for any AI transcription: obtain explicit consent for the AI processing itself, not just the recording. Set every AI tool to off by default and toggle on only after consent is documented.

Second, fix your engagement letters. The client-side AI problem is the most novel issue in NYC Bar Formal Opinion 2025-6: when a client's own AI notetaker joins a call, you may never see the recording, and you cannot verify its accuracy or security. Your letter should require advance notice of any client-side AI notetaker and state that client-generated recordings and summaries are not dispositive or binding on you unless promptly shared for independent review. That single sentence protects you from relying on a summary you never saw.

Third, protect work product. Transcripts retained as business records can become discoverable and rarely contain attorney impressions — which makes them vulnerable to production. Add your impressions to a separate protected document, and establish preservation protocols during litigation holds so nothing gets overwritten or deleted inconsistently.

Finally, build the review workflow itself. The NYC Bar opinion requires lawyers to independently review AI-generated transcripts and summaries before relying on them, because errors can mischaracterize client statements or create content that was never said. That means every AI transcript in your file should be a draft, never the record. This is the principle Evidence Copilot operationalizes: every AI output is marked "AI-generated — pending human review," originals stay immutable, and every answer cites the exact video second or PDF page it came from, so verification is a click, not a re-listen. Whether you use a dedicated workspace or a manual process, the standard is the same — the source recording is authoritative; the transcript is a reviewable draft.

A practical checklist for your file:

• Consent clause naming the AI service, its data governance link, and a manual-transcription opt-out.
• Engagement-letter language requiring advance notice of client-side AI notetakers.
• A stated rule that client-generated summaries are not binding until independently reviewed.
• Separate attorney-impressions documents, kept apart from discoverable transcripts.
• A preservation protocol tied to litigation holds.

Vendor due diligence completes the picture: before consenting to any tool, verify its terms of service, privacy policy, and retention practices, and confirm client data is not stored on external servers or used for training. The Duane Morris analysis flags privilege waiver through third-party vendors as the top risk — and it is preventable with documented diligence done before the recording starts, not after.