Is AI Transcription Illegal Wiretapping? Why Recording Features Are the New Legal Flashpoint

If you're building a product that records, transcribes, or summarizes conversations, the question "is ai transcription illegal wiretapping" is no longer academic. It is the subject of active class actions, expanding plaintiff theories, and a regulatory patchwork that punishes architectural ignorance. This piece covers where exposure actually starts, what courts have ruled so far, and what you can do about it before your next sprint ships something actionable.

Key Takeaways

• Courts are treating AI transcription tools as independent third-party "listeners," not passive extensions of the user who deployed them. This distinction is the core of the legal risk.
• Under California's CIPA, the mere capability to use intercepted data for an independent purpose (like model training) can be enough to survive a motion to dismiss. Actual use is not required.
• California's pending legislative fix (SB 690) would not cover wiretap or recording claims at all. It targets pen-register/tracking-pixel theories only. If you're building recording or transcription features, that bill does nothing for you.
• Auto-generated consent language (e.g., boilerplate in EHR notes claiming a patient "consented") has become evidence against defendants, not a shield. Consent capture is an engineering problem, not a legal afterthought.
• Nearly 4,000 CIPA cases have been tracked through mid-2026, with roughly 83% brought under Section 631(a). This is the fastest-growing category of privacy litigation in the United States.

What Is CIPA and Why Does It Keep Showing Up?

The California Invasion of Privacy Act is a 1967 statute that criminalizes and creates a private right of action for unauthorized interception of communications. Section 631(a) covers wiretapping. Section 632 covers recording of confidential communications. Both carry statutory damages, and both have become the weapon of choice for a plaintiffs' bar that has filed roughly 3,968 tracked cases through mid-2026, with more than 800 in 2025 alone. Florida follows at 811. Illinois at 108. California is the epicenter, and if your product touches audio from a California resident, you are in scope.

The federal analogue is the Electronic Communications Privacy Act (ECPA), specifically Title I (the Wiretap Act). It applies everywhere, but it uses a one-party consent standard: if one participant consents, no violation. CIPA is stricter. California is an all-party consent state. Every participant must consent before recording begins, and "consent" has a specific legal meaning that your onboarding flow probably does not satisfy.

How Are Courts Treating AI Transcription Tools?

As third-party listeners, not as passive instruments of the user. This is the single most important framing shift, and it is the one most engineering teams have not internalized.

In In re Otter.AI Privacy Litigation, filed in December 2025 in the Northern District of California, plaintiffs alleged that the AI notetaker recorded, transcribed, and used conversation contents without consent, violating both the federal ECPA and CIPA. The core legal theory: Otter is not simply a tool controlled by the meeting host. It is a distinct entity that joins the call, captures audio, processes it on remote servers, and retains the output. That makes it a third-party interceptor, not an extension of any participant.

This framing has teeth. If your transcription service is treated as a party to the conversation, one-party consent (federal standard) might save you. If it is treated as a third party listening in, you need the consent of all parties under CIPA, and you need it before collection begins. The Paul Hastings analysis of these cases puts it plainly: when a robot is listening, courts may find it is wiretapping.

Does My Product Actually Have to Use the Data, or Is Capability Enough?

Capability alone can be enough. This is the finding that should change how you architect your pipeline.

In a related ruling, a court found that a company could be treated as an unauthorized third-party listener under CIPA because it had the technical capability to use intercepted call data to improve its AI models. The plaintiffs did not need to prove the data was actually used for training. They only needed to allege the company was "technologically capable" of using it for an independent purpose.

Think about what this means for your stack. If your transcription service sends audio to a cloud endpoint, and that endpoint is operated by a vendor who also trains models, you have a capability argument against you. Your data-use policy says you don't train on customer data. Fine. Policies can change. Architecture is provable in court. The difference between "we promise not to" and "we structurally cannot" is the difference between surviving and not surviving a motion to dismiss.

What Does This Mean for Architectural Decisions?

It means ephemeral processing, architectural separation, and provable data-flow boundaries are more durable defenses than contractual commitments. If audio hits a server where a vendor could retain it, index it, or feed it into a training pipeline, you have exposure. If audio is processed ephemerally and discarded with no persistent access beyond the immediate transcription function, that is a fact you can demonstrate. On-device processing frameworks that keep audio on the device eliminate the third-party data flow entirely, which removes the trigger for wiretap statutes, BAA obligations, and state medical-confidentiality laws. That is a structural defense, not a policy one.

What Happened in the ConverseNow Case?

In August 2025, a federal court in the Northern District of California denied a motion to dismiss a CIPA class action against ConverseNow Technologies. The facts: ConverseNow operated an AI virtual assistant that handled phone orders for a Domino's franchise. A customer called to order pizza. The AI intercepted and recorded the call. No consent was obtained. The court allowed the case to proceed.

This was not a sophisticated surveillance operation. It was a pizza order. The exposure arose from a straightforward product feature (AI answering a phone call) deployed without consent mechanics that satisfied CIPA's requirements. If you are shipping any product that answers, records, or transcribes a phone call involving a California resident, this case is directly on point.

What About the RingCentral Dismissal? Doesn't That Cut the Other Way?

It does, and it illustrates why jurisdiction and statute selection matter enormously. In January 2026, a federal court in the Northern District of Illinois dismissed a wiretapping claim against RingCentral and Heartland Dental. Plaintiffs alleged that RingCentral's AI phone service used speech recognition and language models to generate transcripts, summaries, and sentiment analysis without caller consent. The court dismissed the claim.

The critical detail: this suit was brought only under the federal Wiretap Act, not CIPA. The federal statute's one-party consent standard likely doomed it. Had the same facts been alleged under CIPA, the result might have been very different. If you are reading this and thinking "we got a favorable ruling," check which statute it was under, and check whether any of your users are in California.

How Is Healthcare Becoming the Sharpest Edge of This?

Ambient clinical AI scribes are the current flashpoint. These tools sit in the exam room (sometimes on a phone, sometimes on a dedicated device), record the patient-clinician conversation, and produce a clinical note. The value proposition is obvious. The legal exposure is severe.

On April 8, 2026, three California patients filed a proposed federal class action (Washington et al. v. Sutter Health) alleging Sutter Health and MemorialCare deployed an ambient clinical AI documentation tool to record patient-clinician conversations without informed consent. The complaint alleges violations of CIPA and California's Confidentiality of Medical Information Act, with the added dimension that audio was transmitted to a third-party vendor's cloud servers.

An earlier case set the template. In November 2025, patient Jose Saucedo filed a proposed class action against Sharp HealthCare in San Diego, alleging the health system used an ambient AI tool to record encounters without consent. The complaint made a specific factual allegation that should concern anyone building consent flows.

How Did Auto-Generated Consent Language Become Evidence Against the Defendant?

Sharp's electronic health record notes reportedly contained boilerplate language stating patients had been "advised" of and "consented" to the recording. According to the complaint, no such advisement or consent ever occurred. The EHR system auto-populated the consent language. The note said the patient consented. The patient says they did not. Now the auto-generated text is evidence of a fraudulent consent process, not evidence of actual consent.

This is a product-design failure masquerading as a legal one. If your system generates text that claims consent was obtained, and consent was not actually obtained through a verifiable, timestamped, human-readable process at the moment of actual disclosure, you have created a liability artifact. The system is producing documentary evidence that you violated the statute. The Ninth Circuit has held that consent under CIPA Section 631 must be obtained before data collection begins, not retroactively through a later-displayed privacy policy. A checkbox that no one checked, populated by default, is worse than no checkbox at all.

Treat consent capture as a first-class engineering problem. Timestamped, auditable, tied to a specific session, with a verifiable record that the disclosure was actually presented and affirmatively acknowledged before recording began. Not a Terms of Service link. Not a banner. Not a default-on toggle buried in settings.

Will SB 690 Fix This?

No. Not for recording or transcription features. This is a widespread misconception that will get teams in trouble.

California's SB 690 stalled in the Assembly in 2025 and became a two-year bill. It was revived in mid-2026, with a California Assembly committee advancing amendments on July 1, 2026. The amendments would eliminate private lawsuits over website "pen register" claims (tracking pixels, cookie-based data collection).

But the bill would leave CIPA's core wiretapping provision (Section 631) and confidential recording provision (Section 632) untouched. The committee explicitly noted that lawmakers "may wish to consider a more comprehensive set of solutions" for those provisions later. The bill must still clear the Assembly floor and return to the Senate before an August 31, 2026 deadline, and even if it passes, it does not address the sections under which recording and transcription lawsuits are brought.

If you are building recording features and waiting for legislative relief, you are waiting for a bill that structurally does not apply to your product category. CIPA is still a mess, and it will remain a mess for recording and transcription builders regardless of what happens with SB 690.

What About Chatbots and Text-Based AI?

The theories are expanding. Plaintiffs have begun testing whether AI-powered chat tools "record" or "repurpose" user inputs in ways that trigger CIPA consent requirements. The argument: when a user types a message into a chatbot, and the chatbot vendor captures, stores, or uses that text for purposes beyond the immediate conversation (training, analytics, personalization), that constitutes interception of a communication by a third party.

These theories are largely untested, and courts have been inconsistent. Some have dismissed at the pleading stage, reasoning a website operator cannot "intercept" its own communications. Others have let cases proceed where the technology captured free-text chat inputs or search queries that arguably constituted the substance of a communication. Generative-AI chatbots are becoming the focus of new demand letters and early complaints arguing these "AI listeners" intercept communications without consent.

If your product includes a chat interface powered by a third-party AI provider, and user messages are transmitted to that provider's servers, you have a data flow that looks structurally identical to the one alleged in the Otter.AI and ConverseNow cases. The medium differs (text vs. audio), but the legal theory is the same: unauthorized third-party interception of the contents of a communication.

What About States Beyond California?

Twelve states plus the District of Columbia require all-party consent for recording. California gets the most litigation because CIPA provides statutory damages and a private right of action, which makes class actions economically viable for plaintiffs' attorneys. But the risk is not California-only.

Florida is second in tracked cases (811 as of mid-2026). Illinois is third (108), though Illinois litigation is more commonly brought under BIPA for biometric data than under wiretap statutes specifically. Florida's wiretap statute classifies unauthorized interception as a felony, which means physicians using ambient AI scribes without proper consent face potential criminal exposure, not just civil liability.

Texas is taking a disclosure-focused approach. As of January 1, 2026, the Texas Responsible AI Governance Act requires physicians to disclose to patients, by the date of service, that an AI system is involved in their care, in clear and plain language. This is a narrower requirement than all-party consent, but it creates a new compliance obligation that recording-feature builders need to surface in their product.

What Should Engineering Teams Actually Do?

Here is what the case law suggests, stated as architectural and product requirements rather than legal advice (you still need a lawyer, and specifically one who has read these cases).

1. Treat consent as a real-time, auditable, pre-collection event. Not a privacy policy. Not a terms-of-service checkbox at account creation. A specific, timestamped disclosure presented to every participant before recording begins, with an affirmative acknowledgment that is stored independently from the recording itself. If your system auto-generates language claiming consent was obtained, make sure the consent actually happened first. The Sharp HealthCare case is a direct warning.

2. Architect for structural inability, not policy prohibition. If you can demonstrate that your pipeline processes audio ephemerally and discards it after transcription, with no persistent access by any party (including your vendor), that is a provable architectural fact. A policy saying "we don't train on your data" is a promise. Architecture is evidence. Courts are evaluating capability, not stated intent.

3. Understand your data flows to the provider level. If audio leaves the device and hits a cloud endpoint operated by a vendor who also trains models, you have a capability argument against you, even if your contract says training is prohibited. The question courts ask is whether the vendor is "technologically capable" of using the data, not whether your MSA says they won't.

4. Do not assume SB 690 or any pending legislation will help. The pending California reform does not touch the wiretap or recording provisions. Your compliance posture should be durable under the current statute, not contingent on a bill that might pass, might not, and wouldn't apply to your feature category regardless.

5. Scope your exposure by state. If your product can be used by or in the presence of a resident of any all-party consent state, you need consent mechanics that satisfy the strictest applicable standard. Building for one-party consent (federal) and assuming it covers you everywhere is how the ConverseNow defendant ended up in a class action over pizza orders.

6. Surface AI involvement explicitly. The trend, in both litigation and legislation, is toward mandatory disclosure that AI is present in the interaction. Even in states without all-party consent requirements, failing to disclose that an AI system is recording and processing the conversation creates exposure under emerging disclosure mandates and general consumer-protection theories.

How Big Is This Going to Get?

The trajectory is clear. CIPA wiretap claims are one of the fastest-growing areas of privacy litigation, with roughly 83% of all cases brought under Section 631(a). The plaintiffs' bar has a working template (Otter.AI, ConverseNow, Sharp, Sutter), a receptive set of courts in the Northern District of California, statutory damages that make class certification economically attractive, and a growing universe of defendants as every SaaS product adds "AI meeting notes" or "call summarization" to its feature list.

Healthcare ambient scribes are the current high-profile target, but the theory applies to any product that records or transcribes a conversation and involves a third-party data flow. Customer service AI. Sales call recording. Interview transcription. Meeting summarization. If audio (or even text) goes from one participant to a third-party server without all-party consent, and that server is in a jurisdiction or involves a party subject to all-party consent requirements, you have the same exposure the defendants in these cases have.

The teams that will avoid this are the ones that treat consent as a product feature, data flows as an architectural constraint, and "capability to use" as a higher bar than "policy not to use." The teams that will not are the ones building the feature first and asking legal second.

If you're building an AI product and want to see how we think about privacy at the architectural level, start a free 7-day trial, no card required.

Michael C.

Founder & Principal Engineer, Selina Labs

Michael builds Selina, a privacy-first AI that remembers you across conversations. He ships security-sensitive AI in production — real attacks, real fixes, measured in minutes and dollars — and writes about privacy, security, and LLMs from that seat. Top Rated Plus and expert-verified on Upwork.

• www.upwork.com
• x.com
• medium.com
• selinalabs.io