AI Notetakers for Recruiting Interviews: Candidate Privacy, Illinois AIVIA, and the On-Device Alternative

If a recruiter on your team is running Otter, Fireflies, or Fathom in a candidate interview today, the audio of that interview is sitting on a third-party vendor's servers — and in twelve U.S. states, that architecture may have violated the state's wiretap statute the moment the bot joined. The Illinois Artificial Intelligence Video Interview Act (AIVIA, 820 ILCS 42) layers on additional obligations any time AI "analyzes" the interview — written pre-interview disclosure, explicit consent, and 30-day deletion on request. And the consolidated federal class action against Otter.ai now under submission in the Northern District of California is testing whether an AI bot quietly listening on a call is a "third-party eavesdropper" under decades-old wiretap statutes.

Recruiting is a category where these questions are not academic. Candidate conversations contain protected characteristics, salary history, references to disabilities and accommodations, and negotiation positions the candidate would never volunteer if they knew a vendor's cloud was recording them. This piece walks through the actual legal landscape as of July 2026, benchmarks the leading recruiting notetakers against it, and explains why an on-device architecture is the cleanest way to keep candidate voices out of third-party clouds in the first place.

Why recruiting interviews are the highest-risk category for AI notetakers

An internal team standup with a bot in the corner is a nuisance. A candidate interview with a bot in the corner is a legal and reputational risk stack: consent laws, biometric statutes, anti-discrimination law, and — increasingly — private class actions. As Honeit's interview-intelligence team notes, candidate conversations are private, high-stakes discussions involving personal data, evaluations, and hiring decisions — and if transcripts leak or recordings are mishandled, the company is on the hook.

Three characteristics make interviews unique:

• The other party has no organizational relationship with the notetaker vendor. A candidate has never signed a DPA, agreed to terms of service, or opted into product-training defaults. Whatever the vendor's privacy policy says, the candidate never agreed to any of it.
• The content is disproportionately biometric. Voice, cadence, pauses, and — in video — facial expression are precisely the signals cloud AI tools index. In Illinois, voice patterns can qualify as "biometric identifiers" under BIPA.
• The decision downstream is regulated. A hiring decision informed by AI can trigger Title VII disparate-impact liability if the AI systematically screens out protected groups, per longstanding EEOC guidance.

The state-law consent map recruiters actually have to comply with

Federal law under the ECPA / Wiretap Act permits one-party consent to a recording, but state law is what actually binds most recruiters. Twelve states now require all-party consent for confidential communications, and any interview where a participant reasonably expects the conversation isn't being overheard qualifies. In these jurisdictions, an AI notetaker joining the call without candidate consent is arguably an unlawful interception. We've covered the full state map in our two-party consent compliance guide, but the shortlist is: California, Connecticut, Delaware, Florida, Illinois, Maryland, Massachusetts, Montana, New Hampshire, Oregon, Pennsylvania, and Washington.

Interstate calls make this harder, not easier. Under Kearney v. Salomon Smith Barney, when a caller in a one-party state records a resident of an all-party state, California courts have applied California's stricter statute. For a distributed recruiting team, that means the practical rule is: default to all-party consent for every candidate call.

Illinois AIVIA: the statute that specifically targets AI in interviews

Illinois's Artificial Intelligence Video Interview Act (820 ILCS 42) was one of the first statutes in the country to regulate AI in hiring and remains the most-cited today. The Recording Law summary of Illinois AI statutes lists the four core AIVIA obligations: pre-interview disclosure that AI may be used; a written explanation of how the AI works and what it evaluates; explicit consent from the applicant; and restrictions on who videos can be shared with. Civil penalties are $500 per day per violation.

Two 2026 developments made this materially stricter. First, as RiskTemplates' AIVIA compliance analysis documents, Illinois HB 3773 (effective January 1, 2026) expanded AI employment obligations across hiring, promotions, and discipline decisions — bringing garden-variety AI notetakers used in interviews closer to the statute's reach. Second, a federal court in Deyerler v. HireVue (Feb. 2024) ruled that complying with BIPA does not excuse an employer from AIVIA — the obligations are concurrent, not alternatives.

The University of Illinois HR AI guidelines — one of the few public-sector implementations — spell out what compliance actually looks like in practice: consent must be in writing and obtained before the interview, not embedded in a post-hoc terms-of-service flow.

What AIVIA does not cover — and why that's a trap

AIVIA is narrow. It applies specifically to AI analysis of candidate-submitted video interviews, not to general-purpose transcription of a Zoom conversation. A recruiter who uses Otter to transcribe a live interview but doesn't run separate AI "analysis" on the video may fall outside AIVIA — and then walk straight into BIPA (for voiceprints), the Illinois Human Rights Act as amended by HB 3773, and the state's all-party consent statute. "Not covered by AIVIA" is not the same as "legal."

The Otter.ai litigation and why every recruiting team should be tracking it

The consolidated In re Otter.AI Privacy Litigation (No. 5:25-cv-06911, N.D. Cal.) is the first federal test of whether an AI notetaker is a wiretap. Plaintiffs allege OtterPilot recorded meetings without all-party consent and used the resulting audio to help train the company's models. Judge Eumi K. Lee heard the motion to dismiss on May 20, 2026, and as of publication no ruling has issued — but the exposure is real. As UC Today's coverage of the platform response notes, the case has already pushed Microsoft, Google, and Zoom to rewrite how their meeting platforms handle third-party bots. Microsoft Teams now flags external bots as "Unverified" in the meeting lobby and requires explicit organizer admission (Message Center notice MC1251206, March 13, 2026).

For a deeper walkthrough of the statutes at issue and the potential damages exposure, see our In re Otter.AI hearing analysis. The short version: CIPA allows $5,000 per violation, ECPA allows $10,000 per violation, and BIPA allows $1,000–$5,000 per biometric identifier collected without consent. A recruiter running an AI notetaker on 200 candidate calls a quarter is a meaningful damages surface.

Bot vs. bot-free vs. on-device: the architecture differences that actually matter

The recruiting AI notetaker market is usually described in two buckets — bot-based (Otter, Fireflies, Fathom) and bot-free (Granola, Jamie, Krisp). This framing is misleading. The privacy-relevant question is not whether a bot joins the call. It is where the audio is processed.

Granola's own participant-privacy explainer is candid about this: it captures device audio locally, but it still transcribes in the cloud and offers built-in ways — like a video watermark or automatic chat message — to notify other people in the meeting. That's better than a visible bot in the participant list, but the candidate's voice still leaves the recruiter's device. Jamie and similar "bot-free" tools operate the same way. Only a fully on-device architecture keeps the audio on the recruiter's laptop or phone from capture through transcription and summary.

What Otter's, Fireflies', and Zoom's own privacy policies actually say

Recruiting leaders often assume vendor DPAs solve the candidate-data problem. They don't. Otter.ai's privacy policy grants broad rights to process and use conversation content, and the plaintiffs in the pending federal litigation allege training-on-content is an opt-out default, not an opt-in — an issue we broke down in our Otter de-identification analysis. Zoom's privacy statement covers AI Companion features whose retention and analytics scope have been contested in enterprise contracts.

None of these vendors are lying about their practices. But no candidate has read any of these policies, and no candidate has consented to any of these terms. That gap is what the wiretap and BIPA class actions are testing.

What good candidate-consent capture looks like

Consent is not a legal fig leaf. Done right, it's an actual workflow. The AIVIA-compliant version has four steps: (1) written pre-interview notice describing that AI will be used, what it evaluates, and how it feeds hiring decisions; (2) explicit affirmative consent from the candidate before recording starts — not a buried checkbox in a scheduler; (3) a non-AI alternative pathway for candidates who decline; and (4) a documented deletion workflow that honors requests inside the 30-day statutory window.

The practical problem with cloud notetakers is (4). "Delete" in a cloud transcription service typically removes the primary record but leaves copies in backups, training corpora, subprocessor systems, and any ATS integrations the recruiting team has wired up. On-device architectures make deletion a single-device operation — the audio never existed anywhere else to begin with.

Biometric exposure: why voiceprints are the sleeper issue

Speaker diarization — the feature every notetaker markets as "knows who said what" — works by extracting voice embeddings, a form of biometric identifier. In Illinois, BIPA requires informed written consent before collecting biometric identifiers, plus a published retention schedule and destruction plan. Class actions against Fireflies and against Microsoft Teams' speaker-recognition feature have already tested this theory. Fireflies was hit with a BIPA class action in Illinois in December 2025 (Cruz v. Fireflies.AI Corp., 3:25-cv-03399), and the case has drawn a second suit in the Northern District of Illinois.

For recruiters interviewing Illinois residents — including remote candidates who happen to live in the state — the safest posture is to avoid architectures that extract voiceprints on a vendor server at all.

How Basil AI solves this for recruiting teams

Basil AI is a fully on-device AI notetaker for iPhone and Mac. It uses Apple's on-device Speech Recognition and the Apple Neural Engine to transcribe audio locally — no vendor server, no cloud transcription API, no training corpus. Everything about the audio-processing pipeline is documented in Apple's platform-level privacy commitments: the audio never leaves the device.

For a recruiting workflow, this changes the compliance posture in three specific ways:

• No third-party processor. The candidate's voice is captured, transcribed, and summarized on the recruiter's own device. Otter's, Fireflies', and Granola's clouds are all out of the loop.
• No bot in the participant list. The candidate sees a normal Zoom or Teams call with the recruiter. Consent is captured verbally at the top of the call and logged in the notes — the way most recruiters already run interviews.
• Single-device deletion. If a candidate requests deletion under AIVIA, GDPR, or CCPA, the recruiter deletes the note from their device. There is no backup fleet, no training-set copy, no third-party ATS integration silently holding audio.

Basil is not a magic wand. It doesn't relieve the recruiter of the underlying obligations under the Illinois Human Rights Act as amended by HB 3773, AIVIA's disclosure and consent requirements, or an employer's Title VII obligations if AI is used in hiring decisions. Those are the customer's determinations, not the vendor's. What Basil does is remove the architectural pieces — the third-party vendor server, the training default, the visible bot, the diffuse copies — that a hiring-team lawyer has to write policy around.

A checklist for recruiting leaders evaluating AI notetakers in 2026

For a General Counsel, CCO, or Head of Talent evaluating tools, the questions to ask any vendor:

1. Where is candidate audio processed? On the recruiter's device, or on your servers?
2. What is the default on training? Is customer content used to improve your models unless the customer opts out?
3. What is the deletion SLA on a candidate request, and how do you verify deletion across backups and subprocessors?
4. Does the tool add a visible participant to the call in Zoom, Teams, or Google Meet?
5. How do you handle speaker diarization / voiceprints in Illinois? Do you extract biometric embeddings on your servers?
6. Do you offer an audit log of candidate consent capture, tied to the specific interview?
7. If our AIVIA disclosure obligations change, will your product support written pre-interview disclosure and per-candidate consent tracking without a workaround?

Any vendor who can't answer these in writing is not ready for regulated recruiting. For a broader background on why the industry is splitting between bot-based and bot-free architectures — and why the platforms themselves are cracking down — see our analysis of the 2026 Google Meet and Microsoft Teams bot policies.

The bottom line for recruiting teams

Recruiting interviews are the wrong place to be the vendor's beta tester on a wiretap statute. The 2026 legal landscape — 12 all-party consent states, AIVIA's explicit-consent requirement, BIPA's voiceprint rules, HB 3773's expansion of AI employment obligations, and the pending federal Otter litigation — has moved fast enough that most large recruiting teams are already reviewing their notetaker stack. The tools that best survive that review are the ones that don't create a third-party processor of the candidate's voice in the first place. That is the entire point of an on-device architecture.

Interview candidates without a third party listening

Basil AI transcribes candidate calls entirely on your iPhone or Mac. No cloud upload, no bot in the participant list, no vendor server storing candidate voices. Try it free.