The Otter ruling turned your training clause into a wiretap question
A federal court let wiretap and biometric claims against Otter.ai proceed because it retains and trains on recordings. Any AI vendor with training-permissive terms should read the order.
On August 13, Judge Eumi Lee in the Northern District of California denied most of Otter.ai’s motion to dismiss in the class action over its meeting notetaker. The federal Wiretap Act claim, the California Invasion of Privacy Act claim, and the Illinois biometric claims all move into discovery. If you build an AI feature that touches other people’s conversations, documents, or voices, and your terms let you train on that material, this order is about you, not just about Otter.
Otter’s main defense was the one every AI tools company has been quietly relying on. The notetaker joins the meeting at the direction of the host, so it is an extension of a party to the conversation, and a party can record its own conversation. Courts have accepted that framework for vendors before. The court accepted it here too, and then held that Otter falls outside it. The reason was not that the bot was uninvited. It was that Otter keeps the recordings and uses them to improve its own speech recognition and machine learning models. A vendor that transcribes and hands the transcript back is a tool. A vendor that transcribes, retains, and trains is acting for its own commercial benefit, and at that point it is a third party listening in.
That is a business model turned into an element of a claim. Nothing in the theory requires the output to be wrong or harmful. The alleged injury is complete when the audio is ingested. Section 230 does not apply because nobody is being treated as a publisher, and the statutes carry fixed damages, five thousand dollars per violation under CIPA and up to five thousand under BIPA, that do not require proof of loss. Plaintiffs’ lawyers noticed this immediately because it routes around the product-versus-content fight that has stalled most other AI litigation. The next complaints will be built on the same template, and the defendant list will come straight from terms of service pages, which anyone can read without discovery.
The operational point is that the exposure comes from a specific combination: your product captures a communication involving someone who is not your customer, you retain it beyond what serving the customer requires, and you use it for your own purposes. Break any one of those links and the theory gets much weaker. Retention limits tied to the customer’s use, training only on data from users who opted in as account holders, and a clean separation between what the customer sees and what you keep are engineering decisions, and they are cheaper to make now than to defend later. Having written code before contracts, I will also say that a training toggle buried in settings that defaults to on will read poorly in a deposition.
The Illinois piece deserves its own attention. The court rejected the argument that speaker identification falls outside BIPA and rejected the extraterritoriality defense. If your product distinguishes speakers by voice, you are plausibly collecting a voiceprint from people who never agreed to anything with you, and Illinois residents show up on calls hosted from anywhere.
This quarter, pull your terms of service and your privacy policy and find every sentence that permits you to use customer content to train or improve your models. For each one, decide whether it covers non-customer data at all, whether it defaults to on, and whether the retention period is tied to anything. Then get your engineers to confirm the code actually does what the document says, because a training clause that is narrower than your pipeline is worse than no clause at all.