BIPA & AI Notetakers: Why Speaker Diarization Just Became a Biometric Liability After the Otter Ruling

Published October 14, 2026

October 14, 2026 · 11 min read

Is an AI notetaker's speaker-identification feature a 'voiceprint' under the Illinois Biometric Information Privacy Act? After August 13, 2026, the answer in federal court is: plausibly yes. On that date, Judge Eumi K. Lee of the Northern District of California allowed both Illinois BIPA claims against Otter.ai to proceed to discovery alongside federal Wiretap Act and California Invasion of Privacy Act claims. That single order reframed every cloud AI notetaker's "who said what" feature as a potential biometric-collection event — and attached statutory damages of $1,000 to $5,000 per violation to it.

This article walks through what the Otter order actually held on BIPA, why speaker diarization is now the key legal trigger, which other vendors are already in litigation, and what the architecture of an on-device alternative looks like. Nothing here is legal advice; it is a mapping of the public record for buyers and in-house counsel evaluating meeting-AI tools.

The August 13 Otter Order: What It Actually Said on BIPA

The consolidated case is In re Otter.AI Privacy Litigation, No. 5:25-cv-06911-EKL, pending before Judge Eumi K. Lee in the U.S. District Court for the Northern District of California. On August 13, 2026, the court granted Otter's motion to dismiss only in part, and the plaintiffs' core theories — federal Wiretap Act (ECPA), CIPA Section 631, both Illinois BIPA counts, UCL, and unjust enrichment — all survived.

As Bloomberg Law reported the next day, the court held that plaintiffs adequately pleaded unjust enrichment and violations of ECPA, CIPA's wiretapping provisions, BIPA, and the California Unfair Competition Law. The BIPA holding is the one that quietly broadens the field: it isn't just about whether a bot "joined" a call — it's about what the vendor's software does with the voice data once it's there.

Why BIPA matters more than CIPA at scale

A sharp analysis in Lawsuit Intelligencer captures the structural point: the BIPA claim is "cleaner" than CIPA at class certification because voiceprint collection is uniform conduct across every meeting and "does not require anyone to characterize the conversation." CIPA Section 631 requires fact-specific analysis of whether each communication was "confidential" — a Rule 23(b)(3) predominance problem. BIPA just asks: did you capture biometric identifiers from Illinois residents without written consent? Yes or no.

What a 'Voiceprint' Is Under BIPA — and Why Speaker Diarization Qualifies

BIPA defines "biometric identifiers" to expressly include voiceprints. As the firm Lewis Rice explains, a voiceprint is not the same thing as an audio recording: it is the analysis of unique vocal characteristics used to identify or distinguish a speaker. The moment software attributes speech to specific individuals using speaker-recognition technology, the statute is in play.

In the Otter ruling, Judge Lee pointed specifically to allegations that the Otter Notetaker tags speakers in real time using Zoom participant names or client-entered data, building a speaker-identification profile so it can recognize the same person in later conversations. ID Tech Wire's write-up of the order notes the court rejected Otter's argument that plaintiffs hadn't alleged the voiceprints could identify anyone. The persistent speaker profile — not the raw audio — is what crossed the biometric threshold.

That matters because speaker diarization is the signature feature of virtually every cloud AI notetaker on the market. If diarization requires building a persistent speaker representation to label future utterances, each new Illinois participant whose voice is sampled is a potential BIPA plaintiff.

The Damages Math: Why BIPA Terrifies Enterprise Buyers

BIPA imposes liquidated damages that can reach $1,000 per negligent violation and $5,000 per reckless or intentional violation, plus attorneys' fees. Those penalties price the harm per recording, which — as one practitioner put it — makes "per recording" a multiplier in a class action, not a rounding error.

Scale it: a sales team of ten reps running five discovery calls a day with Illinois residents using a cloud bot that performs speaker ID — the statutory exposure crosses into seven and eight figures within a single quarter. There is no BIPA cure period for most violations, and plaintiffs' firms have filed hundreds of BIPA cases over the past six years in categories as mundane as fingerprint time clocks and retail photo kiosks.

The Broader 2026 Litigation Wave

Otter isn't an outlier — it is the lead case in a cluster. The AI Lawsuit Tracker currently lists the August 13 Otter order as the most recent dated BIPA/biometric AI development on its docket. Parallel suits are already moving:

Case data compiled from the Little Binger's docket summary. The pattern is the same across all four defendants: cloud ingest of meeting audio, plus a speaker-identification feature, plus default use of transcripts for model improvement. That triangle is what plaintiffs' firms are now treating as a biometric-collection fact pattern.

Our prior coverage of the Chamberlain v. Granola wiretap complaint and the May 2026 Otter motion-to-dismiss hearing traces how we got here.

Cloud AI Notetaker vs On-Device: The BIPA Exposure Comparison

The architectural difference between a cloud notetaker and an on-device recorder is not cosmetic for BIPA purposes — it is dispositive. BIPA Section 15 obligations attach to the entity that "collects, captures, purchases, receives through trade, or otherwise obtains" a biometric identifier. On-device processing changes who is doing the collecting.

None of this makes an on-device app a "compliance" guarantee — Illinois notice obligations still apply to anyone recording Illinois residents. What it does is remove the vendor-side collection surface that drives the biggest BIPA class actions.

Why the "Single-Host Consent" Defense Failed

A standard argument from cloud vendors has been that the meeting host's acceptance of terms of service covers everyone — the "single-host consent" model. Judge Lee's order rejected that theory for ECPA and CIPA: because Otter plausibly "independently collects, retains, and uses communications for its own commercial purposes," it looks like a third-party eavesdropper, not a passive tool.

The BIPA theory is even harder to defeat because the statute requires a written release — not an implicit license, not a visible participant badge, not a Zoom chat message. As Fisher Phillips noted in its analysis of the Cruz v. Fireflies complaint, the plaintiff was never informed in writing that her biometric data was being collected, never told how long it would be retained, and never given the option to provide written consent. Those are the Section 15(b) elements, and they are not satisfied by a bot showing up with a name tag.

Who Is on the Hook: Vendor or Employer?

Both. As Amundsen Davis warned employers, although initial litigation often names the AI technology provider as a defendant, employers that deploy these tools are not insulated from liability — Illinois courts have held that multiple entities can be responsible for the same biometric collection.

The practical implication for buyers: a Data Processing Addendum from your notetaker vendor does not transfer BIPA risk to the vendor. If your sales team enables Otter or Fireflies in a Zoom call with an Illinois participant, your company is a potential co-defendant on the Section 15(b) count.

What Buyers and GCs Should Ask Vendors

Below is a checklist drawn from the Otter order's reasoning and from firm-side guidance in the Sheppard Mullin practice note on the Fireflies case. Bring these to your vendor demo:

1. Does your product build a persistent speaker-identification profile across meetings?
2. Is that profile stored on your servers, on the participant's device, or both?
3. Do you have a BIPA-compliant written-release workflow for non-user meeting participants?
4. Where is your publicly posted biometric-data retention and destruction schedule?
5. Are meeting audio and transcripts used to train your ASR or LLM models by default? Is opt-out available at the free tier?
6. What is your notice mechanism for Illinois residents who join a meeting without being on the host's account?
7. In the event of a litigation hold or subpoena, how do voiceprints and transcripts flow out of your system?
8. If I disable speaker ID, does your product still function, and does that eliminate voiceprint creation?
9. Will you contractually indemnify my company for BIPA Section 15(b) claims arising from your product's speaker-recognition features?
10. Can you provide a technical attestation that no voice data leaves the participant's device? (This is the only answer that eliminates the vendor-collection theory entirely.)

Related Issues: Discovery, Retention, and the Transcript as Exhibit A

Even setting BIPA aside, the Otter ruling cemented another concern that corporate counsel began writing memos about in mid-2026: the transcript itself is discoverable. Multiple firms have independently landed on the same phrase — "the meeting bot becomes Exhibit A". Every stored cloud transcript is a potential subpoena target, a potential HR evidence file, and a potential input to opposing counsel's narrative.

For a deeper treatment, see our guide to AI notetakers for lawyers handling client intake and depositions and our analysis of data-privacy architecture for asset managers handling MNPI.

How Basil AI Solves This

Basil AI is an iOS and Mac app that never sends meeting audio to a vendor server. The architecture is deliberately boring in the ways that matter for BIPA:

1. Apple Speech framework with on-device recognition flag

Basil uses Apple's Speech framework (SFSpeechRecognizer on legacy devices, SpeechAnalyzer/SpeechTranscriber on iOS 26+) with on-device recognition enforced. As the Apple documentation specifies, setting requiresOnDeviceRecognition = true prevents an SFSpeechRecognitionRequest from sending audio over the network. The transcription runs on the Apple Neural Engine; the raw audio and the resulting text stay in the app's sandbox.

2. No vendor-side voiceprint collection

Because there is no Basil cloud ingesting audio, no cross-meeting speaker-recognition profile is ever assembled at the vendor layer. Any speaker labeling happens locally against the user's own device-held references. The BIPA Section 15(b) theory that drives the Otter voiceprint claim — a vendor obtaining a biometric identifier without written release — does not attach to a vendor that never receives the voice data.

3. Zero training-data egress

Because audio and transcripts never leave the device, there is nothing to feed back into model training. Apple's broader on-device privacy architecture is the design philosophy; Basil inherits it. For a technical comparison, see our prior write-up on AI meeting notetaker architectures.

None of this makes Basil's deploying customer immune from Illinois notice obligations — if you're recording Illinois residents, you still need appropriate notice and, where applicable, written consent. But the vendor-collection prong of a BIPA class action — the one Judge Lee just let proceed against Otter — does not exist against a tool that never obtains the biometric data in the first place.

What Happens Next in the Otter Case

Otter received leave to amend on the dismissed counts (CFAA, CDAFA, Washington Privacy Act, and most common-law privacy claims). As UC Today reported, the plaintiffs had 14 days from the August 13 order to submit an amended complaint. The surviving ECPA, CIPA, BIPA, UCL, and unjust-enrichment claims proceed into discovery, where plaintiffs will seek Otter's internal documents on how speaker-identification profiles are built, retained, and used in training — exactly the record that will set the pattern for Fireflies, Microsoft, and Granola.

For deploying organizations, the practical calendar is tighter than the litigation calendar: your next vendor renewal is a good moment to put the ten questions above in front of the sales rep in writing.

Try the on-device alternative

Basil AI records up to 8 hours continuously, transcribes in real time on your Apple Neural Engine, and never uploads your audio to a vendor server. No voiceprint collection at the vendor layer. No training data egress. Your meetings stay on your device.

Nothing in this article is legal advice. Consult a licensed privacy attorney in your jurisdiction before relying on any statement here for a compliance decision.