Does AI Transcription Count As Recording Under The Law?

The short answer

Yes. Federal wiretap law bans intercepting a conversation, which means acquiring its contents with a device, and a live transcript is the contents. Illinois names transmitting and California names learning the contents in transit, so a tool that only keeps text needs the same consent as a recorder. Where transcript-only actually helps is afterward, with less to leak, subpoena, or turn into a voiceprint.

Why This Question Keeps Coming Up

Because AI notetakers sell "we don't store your audio" as a privacy feature, and a lot of people hear it as a legal one.

I wear Pocket to work every day and run a software notetaker on my calls, so this question comes up for me every working day. The pitch has shifted. It used to be about encryption. Now it's about what the tool throws away. Plenty of notetakers say they transcribe in real time and never keep the recording, and the implication hangs in the air that if there's no recording, the recording laws don't apply.

That implication is wrong, and it's worth knowing why before you lean on it. My earlier guide on whether it's legal to record conversations covers one-party versus all-party consent, the state list, and calls across state lines. This post picks up the one question that guide skipped. If a tool turns speech into text and deletes the audio, did you record anything?

Short version, the law mostly doesn't care about the file format. It cares about whether a device captured what people said. The long version is below, with links to the statutes so you can read the actual words instead of trusting mine.

Same disclaimer as last time. I'm not a lawyer and this isn't legal advice. I'm a reviewer who reads the statute before trusting a privacy page, and every legal claim here links to its primary source. If a transcript could end up in a lawsuit, an HR file, or a custody fight, pay a lawyer in your state for an hour.

What Federal Law Actually Bans

Federal law bans intercepting a communication, and it defines intercepting as acquiring the contents with a device. Whether anyone keeps the audio isn't part of the test.

The Wiretap Act's definitions live in 18 U.S.C. 2510. Paragraph 4 defines intercept as the "aural or other acquisition of the contents" of a wire, electronic, or oral communication through any electronic, mechanical, or other device. Read that slowly. Aural means hearing. "Or other" means any other way. The question is whether a device acquired the contents, not whether a file sat on a disk afterward.

Then look at what counts as contents. Paragraph 8 says contents includes any information about the "substance, purport, or meaning" of the communication. A transcript is nothing but the substance and meaning of what people said, written down. It's hard to picture a cleaner example of acquiring the contents than a machine that listens and produces the words.

The ban itself is in 18 U.S.C. 2511. Subsection (1)(a) makes it illegal to intentionally intercept a communication. Subsection (2)(d) is the party exception that makes one-party consent work. You can intercept a conversation you're part of, or one where a participant agreed, as long as you aren't doing it to commit a crime or a tort. Notice that both the ban and the escape hatch use the same word. Transcription lives inside that one rule, on both sides of it.

Timing helps the argument, not hurts it. Several federal appeals courts have read intercept to mean grabbing a communication while it's moving, not digging it out of storage later. Live transcription is about as "while it's moving" as it gets. A tool that transcribes as people talk fits the statute more snugly than one that records a file and processes it tomorrow.

For most people this is reassuring. In a one-party state, transcribing your own call is exactly as legal as recording it. Transcription doesn't put you in a worse spot. It just doesn't put you in a better one. If you'd need everyone's consent to record, you need everyone's consent to transcribe, and the civil damages in 18 U.S.C. 2520 don't ask what format the capture was in.

Illinois Names Transmitting Out Loud

Illinois doesn't leave this to interpretation. Its eavesdropping law treats transmitting a conversation as its own act, separate from recording it, and it uses the word transcribe for electronic communications.

Start with the definition of an eavesdropping device in 720 ILCS 5/14-1. It's any device capable of being used to hear or record an oral conversation, or to intercept or transcribe electronic communications, whether the conversation happens in person, by phone, or by any other means. A laptop running a notetaker hits nearly every verb in that sentence.

The offense is in 720 ILCS 5/14-2. If you're a party to a private conversation, it's eavesdropping to use one of those devices in a surreptitious manner to transmit or record it without everyone else's consent. Transmit is doing a lot of work there. A cloud notetaker that streams a call's audio to a transcription server is transmitting the conversation, whether or not anyone saves a copy at the other end.

Two subsections later, the statute makes it a crime to intercept, record, or transcribe a private electronic communication you aren't part of. Transcribe is right there in the text. That sentence reads very differently once a vendor's servers are in the loop, because the vendor isn't a party to your call.

Here's the catch that works in your favor. Every one of those offenses requires a surreptitious manner, and the statute defines surreptitious as done by stealth or deception, or through secrecy or concealment. An announced notetaker isn't secret. A recorder clipped to my shirt in plain view, with a sentence at the top of the meeting, isn't concealment. Illinois punishes hiding it, not doing it.

Illinois also makes it a separate offense to use or disclose information you know came from an illegal eavesdrop, unless everyone consents. That matters for transcripts specifically, and I'll come back to it below.

California Covers Learning The Contents

California's two main wiretap sections cover learning a communication's contents while it's in transit and eavesdropping on a confidential conversation. Neither one requires that anybody keep a recording.

Penal Code section 631 is the older and broader one. Among other things, it covers anyone who, without the consent of all parties, reads or tries to read or learn the contents or meaning of a communication while it's in transit. It also covers anyone who uses information obtained that way, and anyone who helps someone else do it. Learning the contents in transit is the whole job description of a live transcription engine.

Courts have generally read 631 as aimed at outsiders rather than the people actually in the conversation. That's exactly why the AI lawsuits point it at the vendor instead of the host. The argument goes like this. The company running the transcription is a third party learning the contents in transit, and the host's consent doesn't cover that third party.

Section 632 is the one most people mean when they call California a two-party state. It makes it a crime to use an electronic amplifying or recording device to "eavesdrop upon or record" a confidential communication without everyone's consent. Two verbs, joined by or. A tool that listens live and writes it down is at minimum eavesdropping with an electronic device. It also captures audio through a microphone and holds it, however briefly, before turning it into text. I wouldn't want to be the test case arguing that isn't a recording device.

The same section has an escape hatch that matters a lot here. Its definition of person leaves out someone known by all parties to be overhearing or recording, and its definition of confidential communication leaves out situations where people reasonably expect to be overheard or recorded. Tell everyone the notetaker is running and both of those start working for you. Skip it, and section 637.2 lets each person you captured sue for $5,000 per violation.

Three Laws, One Answer

Each of these laws turns on capturing or learning what was said, and none of them carves out tools that keep only text. Here's the side-by-side, current as of October 18, 2026.

The first four rows give the same answer in four different vocabularies. The last row is the odd one out, and it's the only place on this table where keeping audio versus keeping text can actually change your exposure. It isn't a consent-to-record rule at all. I'll get to it in the leftovers section.

The Vendor Is Another Listener

A cloud notetaker adds a company to your conversation. Plaintiffs argue that company is a third party that needed its own consent, and two federal rulings so far have let that theory move forward.

The big one is In re Otter.AI Privacy Litigation in the Northern District of California. The plaintiffs say Otter's assistant joined Zoom, Teams, and Google Meet calls, captured people who hadn't agreed, kept the results, and used meeting data to improve its models. According to UC Today's report on the August 13, 2026 order, the judge let the federal wiretap, California privacy, and Illinois biometric claims proceed, and dismissed some others with leave to amend. That isn't a finding that Otter broke the law. It means the case survives to the next round.

Look at how the allegation is framed in that coverage. Plaintiffs say Otter recorded, transcribed, and retained communications without the consent the law requires. Transcription isn't cast as the harmless part of the story. It's listed right alongside recording as part of the alleged interception.

The earlier signal was Ambriz v. Google, decided in February 2025 in the same district. That case went after Google's customer service AI, which listens to calls between businesses and their customers. The court applied what lawyers call the capability test under section 631. If a vendor has the capability to use call data for its own purposes, it can count as a third-party eavesdropper even without proof that it actually did. For anyone choosing a notetaker, that turns the terms of service into a legal document in a very literal way.

This is where "we don't keep your audio" stops helping. A vendor can transcribe in real time, drop the audio a second later, and still be the party that learned the contents in transit. Deleting the evidence of an interception doesn't undo the interception.

So read a notetaker's data terms as part of the consent question, because under the capability theory they are. Check whether the vendor trains on your data by default, whether you can switch it off, and which outside providers touch the audio. Granola is a useful example because it's upfront about it. Its security page says it doesn't store meeting audio, transcribes in real time on desktop through providers like Deepgram and Assembly, and trains on anonymized data unless you opt out in settings. That's an honest description, and it proves the point of this post. No stored audio, but the audio still travels to a transcription provider while people are talking. My Granola review covers the product side, and Market Correct's AI note taker roundup compares the rest of the field.

On-Device Vs Cloud Transcription

On-device transcription takes the vendor out of the conversation, which takes the third-party theory off the table. It doesn't take you out, so your own consent obligations stay exactly where they were.

Where the text gets made matters more than whether the audio gets kept. Roughly three setups are on the market right now, and each carries a different kind of risk.

On-device is the cleanest answer to the vendor problem, because there's no vendor in the conversation. Some phones now transcribe voice memos locally, and Market Correct's guide to transcribing voice memos on iPhone walks through the built-in route. The tradeoff is usually features. The summaries, search, and speaker labels people actually buy notetakers for mostly run in the cloud. If you're converting files you already have, Market Correct's guide to transcribing audio to text covers the main routes.

Cloud with audio dropped is where most bot-free notetakers sit. It beats keeping audio for several reasons, all of them in the next section. It just doesn't touch the consent question.

Cloud with audio kept is where wearables usually land, and where meeting tools built around video clips land too. My Pocket review spells it out. The device uploads audio to Pocket's cloud for transcription, and there's no local-only mode. My Fathom review covers a recording-first tool that keeps the video so you can clip it. Neither is a problem if you announce them. Both mean more data sitting somewhere after the call ends.

Bots versus bot-free matters here too, just not legally. A visible bot in the participant list does part of your announcing for you. A bot-free tool like the Wispr Flow notetaker captures from your computer's audio, so nobody on the call knows unless you say it. The law is the same either way. The job of telling people shifts entirely to you.

Where Transcript-Only Actually Helps

Everything after the conversation. Less data kept means less to leak, less to hand over in a lawsuit, and in Illinois, possibly less biometric exposure.

I don't want to oversell this post into "transcript-only is pointless." It isn't. It just solves a different problem than consent. Consent is about the moment of capture. Retention is about everything that happens to the capture afterward, and that's where text-only is genuinely safer.

Leaks. A transcript is less dangerous than audio in one specific way. It doesn't carry anyone's voice. A leaked recording can be clipped, replayed out of context, and used to clone someone's voice. A leaked transcript is embarrassing, but it's words on a page, and anyone can claim a transcription error.

Subpoenas and discovery. Anything you keep can be requested in a lawsuit. Audio adds tone, hesitation, and the exact thing someone said versus what the summary says they said. If the audio never existed past the live stream, nobody can demand it. The transcript is still fair game, so this shrinks the exposure instead of removing it.

Voiceprints. This is the Illinois one. BIPA lists voiceprint in its definition of a biometric identifier, and section 15 says a private company can't collect one without first telling the person in writing and getting a written release. It also requires a public retention schedule and a plan for destroying the data. Damages under section 20 run $1,000 per negligent violation and $5,000 per intentional or reckless one.

Speaker identification is where this bites. In the Otter case, the BIPA claim that survived centers on speaker-identification features built from voice data. A tool that drops audio and never builds a voice profile has a much thinner connection to BIPA than one that keeps audio and fingerprints speakers. Some tools label speakers from the calendar invite instead, which is how my Wispr Flow review describes its notetaker. Whether that sidesteps the voiceprint question depends on how a tool matches voices to names, and vendors rarely say. Whether speaker labels legally create a voiceprint at all is unsettled. That's the honest answer, and it's also the reason to care if anyone on your calls lives in Illinois.

Retention settings are the lever you actually control. Check whether your notetaker lets you set how long audio and transcripts are kept, and whether it lets you delete recordings by hand. Shorter is better on all three counts above.

Using The Transcript Is Its Own Problem

If the capture was illegal, using the transcript can be a second violation. Federal law, Illinois, and California all say so in plain text.

Federal law covers this in subsections (1)(c) and (1)(d) of section 2511. It's illegal to intentionally disclose or use the contents of a communication if you know or have reason to know they came from an illegal interception. Contents, again, means the substance and meaning. A transcript forwarded to your team is disclosing the contents. A summary pasted into a CRM is using them.

Illinois has the same idea in 14-2(a)(5), which covers using or disclosing information you know came from an illegal eavesdrop. California's section 631 reaches anyone who uses or tries to use information obtained by learning a communication in transit. So the transcript isn't a safe harbor you retreat to after a bad capture. It's the thing you'd get in trouble for spreading.

The flip side is good news. A transcript from a properly announced meeting is clean. You can share it, summarize it, and feed it into whatever AI tool you like, subject to any confidentiality you promised the people on the call. Get the capture right and the downstream use mostly takes care of itself.

What I'd Actually Do

Treat transcription exactly like recording. Announce it, get the yes into the transcript itself, then use the transcript-only advantages for what they're good at, which is keeping less.

The script is the same one from my recording laws guide. Something like "Mind if I run my notetaker? I'll send you the summary after." Say it after the tool has started, so the question and the answer both land in the transcript. That puts the consent in writing, timestamped, in the same document as the conversation it covers.

The rest is a short routine.

1. Say "recording" or "transcribing" out loud, never just "taking notes." A tool that streams audio to a server isn't you holding a pen, and soft wording is what a secrecy argument feeds on.
2. Know where your text gets made. On-device, cloud with audio dropped, and cloud with audio kept carry different risks.
3. Turn off training on your data if the tool allows it. That's the fact pattern at the center of the Otter case and the capability test in Ambriz.
4. Set retention short, and delete audio you don't need even if the tool keeps it by default.
5. Think twice about voice-based speaker identification if anyone on your calls is in Illinois.
6. Don't forward a transcript from a call where someone said no. Using and disclosing it are their own violations.

None of this is a reason to skip the tools. An accurate transcript beats anyone's memory of a meeting, and most people on the other side agree once they're asked. For the hardware half, Pocket is the one I wear, with cloud transcription I announce at the top. My best AI voice recorder ranking compares it with the other device I've tested, and Market Correct's Pocket AI review goes deeper on the hardware.

Frequently Asked Questions

Does AI transcription count as recording?

For consent purposes, yes. Federal law bans intercepting a conversation, which means acquiring its contents with a device, and a transcript is the contents. Illinois and California use similar language. If you'd need consent to record, you need it to transcribe.

Is it legal to transcribe a call without telling the other person?

Under federal law and in one-party consent states, generally yes if you're on the call and aren't doing it to commit a crime. In all-party states like California and Illinois, everyone has to agree or at least know. Cloud notetakers add a vendor to the call, which is the theory behind the Otter lawsuit.

If my notetaker deletes the audio, do I still need consent?

Yes. Consent rules apply at the moment of capture, and deleting the audio afterward doesn't undo it. Deleting audio does shrink what can leak or be subpoenaed later, and it may lower biometric privacy exposure in Illinois.

Is on-device transcription safer than cloud transcription?

For third-party risk, yes. If the audio never leaves your device, no vendor learns the contents in transit, which is the core claim in the notetaker lawsuits. Your own consent obligations are the same either way.

Can I share a transcript of a meeting?

If everyone consented to the capture, generally yes, subject to any confidentiality you agreed to. If the capture was illegal, federal law and states like Illinois and California treat using or disclosing the contents as a separate violation.