Did Zoom train AI on meeting recordings?

Zoom's March 2023 terms update granted a perpetual license to use meeting content for AI training. After significant public backlash, Zoom CEO Eric Yuan revised the terms to exclude audio, video, and chat content from AI training without user consent. However, 'service-generated data' - usage patterns and metadata - remained outside the revised commitment. Whether Zoom trained on meeting content during the period the original terms were in effect is not publicly known.

Is Zoom safe for confidential meetings?

Zoom has updated its terms to exclude meeting content from AI training without consent, and has end-to-end encryption available for some call types. However, all meeting audio is processed through Zoom's cloud infrastructure - it leaves your device. Terms can change, ownership can change, and 'service-generated data' remains outside the current consent commitment. For meetings involving legal strategy, medical discussions, financial planning, or journalist sources, local voice transcription tools that process audio on-device are the appropriate choice.

What is the difference between Zoom's 'customer content' and 'service-generated data'?

Zoom's revised terms committed to not training AI on 'customer content' - the recordings, transcripts, and files you actively create in meetings - without consent. 'Service-generated data' refers to usage patterns, feature interactions, timing data, and metadata about how you use Zoom. The revised consent commitment applies to the first category but not the second, meaning Zoom can still use behavioral metadata from your calls for AI development under the original terms.

Zoom Tried to Train AI on Your Meetings. Here's What Happened.

Zoom became essential infrastructure during the pandemic. Hundreds of millions of people used it for doctor's appointments, therapy sessions, legal consultations, business negotiations, and personal conversations. Then in 2023, Zoom updated its terms of service to claim a perpetual, worldwide license to use all of that content for AI training.

The backlash was significant enough that Zoom CEO Eric Yuan publicly called it a mistake. The terms were revised within weeks. The story still matters - both for what it reveals about the default posture of cloud meeting tools toward user data, and for the loophole that survived the revision.

What did Zoom's terms actually say?

Zoom's March 2023 terms update granted the company a "perpetual, worldwide, non-exclusive, royalty-free, sublicensable, and transferable license" to use customer content for "machine learning, artificial intelligence, training, testing" and product development. The terms applied to meeting content including video, audio, transcripts, and files shared in calls.

The original terms also contained language that appeared to contradict the AI training provision - a clause stating that Zoom would not use audio, video, or chat content for training without consent. Gizmodo's reporting on the contradictory language brought it to wide attention. The problem wasn't that Zoom buried a disclosure - it's that the terms said two opposite things, and the version that served Zoom's interests was the broader one.

The reaction and the revision

The public reaction was fast and negative. Security researchers, legal commentators, and enterprise customers raised concerns. Zoom CEO Eric Yuan posted a public response acknowledging the company had "created confusion" and committed to not training AI models on customer content without consent.

Axios reported that Yuan described the terms update as a mistake in how it was communicated. The revised terms added explicit language: "Zoom will not use audio, video or chat Customer Content to train our artificial intelligence models without your consent."

That sentence is the right sentence. The question is why the original terms said the opposite.

The loophole that remained

The revised consent commitment applied to "Customer Content" - the meeting recordings, transcripts, and files that users actively create. It did not apply to "service-generated data" - usage patterns, feature interactions, timing data, and metadata generated by how users use Zoom.

TechCrunch's analysis identified this gap: Zoom could still use behavioral and metadata from your calls to train AI models without the revised consent requirement applying. The meetings themselves were protected by the new language. Everything around them remained fair game under the original terms.

This is a common pattern in cloud service terms revisions following privacy backlash. The revision addresses the specific complaint precisely enough to quiet the reaction, while leaving adjacent data collection in place. The result looks like a complete fix while actually being a partial one.

Zoom's prior settlement

The 2023 controversy was not Zoom's first privacy settlement. In 2021, Zoom paid $85 million to settle a class-action lawsuit over sharing user data with Facebook, Google, and LinkedIn without user permission - and for falsely claiming the service provided end-to-end encryption when it did not.

$85 million for sharing data without permission and misrepresenting encryption. Then, two years later, terms that claimed a perpetual license to use meeting content for AI training. These are not isolated incidents. They represent a consistent orientation toward user data that treats it as an asset to be used, with user consent as an obstacle to be managed.

What this means for confidential meetings

Most Zoom users don't read terms of service updates. Most Zoom users conducting confidential calls - with lawyers, doctors, investors, journalists, sources - didn't know that for some period in 2023, Zoom's terms claimed the right to use those calls for AI training. They may not know today that the revision left service-generated data outside its scope.

The meeting content on a cloud platform is subject to that platform's terms, which can change. A revision can re-expand what was narrowed. An acquisition can bring new owners with different priorities. A regulatory change can compel disclosure that current terms prohibit. The meeting exists on infrastructure you don't control, under terms you may not have read, governed by policies that can change.

For genuinely confidential meetings - legal strategy, medical consultations, financial planning, journalist sources, business negotiations - this isn't a theoretical risk. It's a structural exposure that exists as long as the meeting audio leaves your device.

ToolPiper's voice transcription processes audio locally on your Mac. A meeting you transcribe with ToolPiper stays on your machine. No cloud terms apply. No revision can change that. The Zoom terms controversy is irrelevant to a tool whose audio never reaches Zoom's servers - or any server.

Download ToolPiper at modelpiper.com.

Part of the Voice AI Privacy series. Related: AI Meeting Recorders and the Consent Problem. Amazon Alexa's Listening Program. Is ToolPiper Safe?