The lawsuit, in one paragraph
On September 25, 2026, Ken Sheldon and Shanaya Grant filed a putative class action against Zoom in the U.S. District Court for the Northern District of California. The suit — Sheldon v. Zoom Communications, Inc., No. 3:26-cv-10962 — alleges that Zoom's AI Companion and the newer ZoomMate add-on capture meeting participants' voiceprints without the written consent the Illinois Biometric Information Privacy Act (BIPA) requires. The complaint, as summarized by Settlement Insight, pleads three BIPA counts: collection without written consent (§ 15(b)), no public retention policy (§ 15(a)), and disclosure without consent (§ 15(d)). The proposed class covers people who spoke in a Zoom session with AI Companion or ZoomMate running on or after September 5, 2023 — the day Zoom launched AI Companion.
Why Zoom's speaker labels are the legal problem
The specific feature at the center of this case is the one that makes AI notetakers useful: speaker labels. Transcripts that read "Alice: … / Bob: …" require the system to tell voices apart. According to the Silver Golub & Teitell investigation page, Zoom's AI Companion "analyzes the unique characteristics of each participant's voice" to attribute statements to individual speakers. The plaintiffs' theory — now the central legal argument of 2026-era voiceprint litigation — is that those numeric voice profiles are voiceprints under BIPA.
The complaint goes further. It cites a Zoom patent issued in February 2025 that describes extracting voiceprints and storing them alongside user accounts, and alleges that ZoomMate, launched in June 2026, sends speaker data to outside platforms including Salesforce and AI vendors such as OpenAI. The plaintiffs say they tested the behavior using Zoom's own Scribe developer tool, which allegedly matched short, separately recorded voice samples to the same speakers from a Zoom meeting — by voice alone.
What BIPA actually says — and why damages add up fast
BIPA (740 ILCS 14) is the only U.S. state biometric law with a private right of action. Section 10 defines biometric identifiers as retina or iris scans, fingerprints, voiceprints, and scans of hand or face geometry. As the American Bar Association explains in its litigation newsletter, prevailing plaintiffs are entitled to liquidated damages of $1,000 per negligent violation or $5,000 per intentional or reckless violation, plus attorney fees.
Critically, the Illinois Supreme Court held in Rosenbach v. Six Flags Entertainment Corp. that a person need not allege any actual injury beyond the statutory violation itself to be an "aggrieved" plaintiff. In the voiceprint context, that likely means plaintiffs do not need to prove their voice was used to generate any specific AI output or that they suffered economic harm — the mere collection without consent is enough.
Damages were narrowed by a 2024 amendment (Public Act 103-769), which provides that repeated collection of the same identifier from the same person by the same method counts as a single violation. In April 2026, the Seventh Circuit held in Clay v. Union Pacific Railroad that this limit applies retroactively, as explained in the Secure Privacy 2026 BIPA guide. But "one recovery per person" at $1,000–$5,000, multiplied across every Illinois meeting participant since September 2023, is still a potentially enormous number.
How we got here: the broader voiceprint docket
Zoom is not the first AI meeting tool to face a BIPA voiceprint suit — just the most recent high-profile one. A December 2025 federal complaint against Fireflies.ai alleges that when a host enables Fireflies on Zoom, Microsoft Teams, or Google Meet, it "records, analyzes, transcribes, and stores the voiceprints of all meeting participants, including individuals who never created Fireflies accounts, never agreed to Fireflies' Terms of Service, and never executed any written consent authorizing biometric data collection." The Loevy + Loevy BIPA AI practice page confirms the firm is bringing class and individual claims across the AI stack, from end-user notetakers to vendors of speaker-ID models and voice-recognition infrastructure.
The Sigma Law Group counts "nine AI voiceprint class actions" in flight in 2026, with named plaintiffs including working voice professionals such as Chicago journalist Carol Marin and Pulitzer winners Yohance Lacour and Alison Flowers. These earlier suits target training-data uses; Sheldon is distinctive because it targets the live meeting workflow itself — the ordinary case where a sales team or an HR panel turns on AI Companion and starts recording.
The non-user problem: people who never consented to anything
The feature of BIPA that makes meeting AI particularly exposed is its indifference to whether you are a customer. BIPA protects any "person" whose biometric data is collected in Illinois, which is why the Silver Golub & Teitell investigation notice specifically asks about people who "never joined a Zoom meeting at all" but whose voice was picked up because a spouse, roommate, or coworker was on a call with AI Companion running nearby.
Think about who speaks on a typical sales or vendor call: your prospect's teammates, a client's external counsel, a job candidate, a patient's family member on a telehealth video. None of them clicked through Zoom's terms of service. None of them signed a written biometric release. If AI Companion was labeling who-said-what, they are — under the plaintiffs' theory — in the class.
Cloud speaker-ID vs. on-device transcription: the architectural comparison
The reason this lawsuit matters beyond Zoom is that every cloud AI notetaker that produces speaker-labeled transcripts does roughly the same thing architecturally: upload audio, compute per-speaker voice embeddings on a vendor server, store the embeddings keyed to the account, and attribute each utterance to a labeled speaker. The table below maps where BIPA exposure actually lives.
| Dimension | Cloud AI (Zoom AI Companion, Otter, Fireflies) | On-device (Basil AI on iPhone / Mac) |
|---|---|---|
| Where audio is processed | Vendor's servers | Your device's Neural Engine |
| Voiceprint / voice embedding generated by | Vendor (a BIPA "private entity") | On-device model on your hardware |
| Third-party custodian of the biometric | Yes — vendor stores embeddings | No — nothing leaves the device |
| Subpoena / class-action discovery surface | Vendor can be subpoenaed directly | No vendor to serve |
| Written-consent collection obligation | Falls on the vendor + host under BIPA § 15(b) | No external collection occurs |
| Non-participant exposure (people in the room) | Captured if any mic hears them | Audio stays local; no biometric database built |
| Cross-state data transfer | Yes — Illinois voices land on out-of-state servers | No transfer |
What "on-device" actually means for voice biometrics
Modern iPhones and Macs can transcribe long-form audio locally using Apple's Speech framework. The older SFSpeechRecognizer API exposes a requiresOnDeviceRecognition flag that forces local inference and prevents the request from sending audio over the network. In iOS 26 and macOS 26, Apple introduced SpeechAnalyzer and SpeechTranscriber, a modern replacement that runs only on-device, with no server fallback and no hard duration cap — purpose-built for lectures, meetings, and multi-speaker conversations.
From a BIPA standpoint, the critical fact is that when the model runs on your own hardware, no "private entity" external to you collects a voiceprint. The computation that distinguishes speakers happens in process, in memory, on silicon you own, and discards the intermediate state when the session ends. There is no vendor database of embeddings to subpoena, breach, or share with a downstream AI platform.
Why this is a bigger governance problem than most companies realize
Legal commentators have been warning about this exposure for most of 2026. The analysis in Startup Fortune put it bluntly in June 2026: a vendor-operated AI notetaker in a sensitive meeting "is not just furniture" — if the bot records a lawyer, a customer, or a deal counterparty without clear consent and controls, you may have built a transcript your company cannot defend. The Darrow Everett advisory from mid-2026 adds that AI-produced content is typically treated as a neutral document and "easily discoverable by opposing parties" — compounding the biometric risk with a conventional eDiscovery risk.
For an in-depth look at how those same cloud transcripts become evidence in litigation, see our earlier analysis of whether AI meeting notes are discoverable, and our coverage of the May 2026 BIPA voiceprint suits against Big Tech, which first set up the "voice embeddings = voiceprints" theory now being pressed against Zoom.
What this means for different roles
For GCs and compliance officers
- Inventory every meeting tool that generates speaker-labeled transcripts, including default-on features like Zoom AI Companion and Microsoft 365 Copilot summaries.
- Confirm whether any Illinois participant (customer, candidate, employee, vendor, family-member-in-room) could be in your call stream.
- Update AI-usage policies and litigation-hold templates to include voiceprint-generating features, not just transcripts.
- Negotiate explicit biometric data-processing addenda with your meeting-platform vendors — or move sensitive calls to a tool that doesn't generate off-device embeddings at all.
For privacy-conscious individuals
- If AI Companion is on in a meeting you join, you'll typically see a pop-up reading "AI Companion has been turned on" and a small star/sparkle icon in the meeting UI.
- You can decline to speak, ask the host to disable AI Companion, or — if you need a record of the meeting — run your own on-device recorder locally rather than letting the cloud summarizer attribute your words.
- If you spoke in an Illinois Zoom meeting with AI Companion since September 2023, you may be in the proposed Sheldon class.
For legal, HR, M&A, and clinical teams
- Treat voiceprint exposure like you'd treat PHI exposure: your default assumption should be that cloud speaker-ID = regulated biometric collection.
- For privileged conversations, prefer architectures where no third-party vendor computes a voice embedding of anyone in the room.
How Basil AI solves this
Basil AI is a privacy-first meeting recorder built around a simple architectural rule: audio never leaves your device. Transcription runs on Apple's on-device Speech framework (and, on iOS 26 / macOS 26, SpeechAnalyzer), executing on the Apple Neural Engine inside your iPhone or Mac. Speaker separation happens locally. Summaries and action items are produced locally. There is no cloud ingest step, no vendor-side voice embedding database, no cross-state biometric transfer, and no third-party custodian whose servers can be subpoenaed in the next BIPA class action.
For the specific concern raised by Sheldon v. Zoom: because Basil AI never transmits meeting audio to an external "private entity," the architectural precondition for a BIPA § 15(b) collection-without-consent claim against the vendor doesn't exist in the first place. Whether a given deployment meets a specific regulatory standard is still your organization's determination, but the architecture fact — no vendor server holding your recording — is on your side rather than against you.
If you want to go deeper on the on-device model that powers this, see our walkthrough of bot-free, on-device meeting notes and our offline transcription deep dive.
The bigger pattern: biometrics are the next frontier of AI meeting regulation
Discovery risk, privilege risk, and training-data risk have dominated AI-meeting coverage in 2025 and the first half of 2026. The Sheldon case opens a fourth, structurally harder front: biometric risk. Discovery you can mitigate with retention policies. Privilege you can try to protect with careful tool selection. Training-data use you can sometimes opt out of. But a BIPA voiceprint claim attaches at the moment a voice embedding is computed by a third party without written consent — and no retention policy, no opt-out toggle, and no post-hoc deletion fixes a violation that already occurred.
The only durable mitigation is architectural: don't let a third party compute the biometric in the first place.
The bottom line
Sheldon v. Zoom is at the complaint stage; Zoom has not yet filed a substantive response, as noted in the Voibe AI Privacy Tracker. The case could narrow on standing, on BIPA's extraterritoriality limits, or on how courts classify numeric voice embeddings. But the trajectory is unmistakable: when a cloud AI meeting tool says "who said what," plaintiffs' firms are increasingly going to argue it computed a voiceprint — and under BIPA, that triggers consequences whether or not any specific person was harmed.
Any organization that cares about not being the next defendant should be asking a simple question before every meeting: does our architecture require a third-party vendor to compute a biometric identifier of the humans in this call? For an on-device recorder, the answer is no.
