Granola Suit Challenges an AI Notetaker Built to Stay Out of Sight

Those descriptions have not been tested in court. Nor does the filing establish that every participant was identified, recorded or subjected to the same settings. But its theory is that speakers could be captured before they knew an AI notetaker was present.

The filing also alleges that the product’s chat notice and camera watermark were optional rather than enabled by default. Granola’s consent guidance confirms that users can enable a watermark and an automated meeting-chat message; it calls it best practice to make participants aware that an AI notetaker is in use and cautions that recording and transcription rules vary by region. The same guidance says Mac and Windows versions do not store audio, only transcripts, while the iPhone version temporarily caches audio during a meeting and deletes that cache when transcription is complete.

That is an important limit on the complaint’s framing. Granola documents notice tools and describes its own handling differently from a permanent audio archive. It does not, however, say that a non-user receives a notice by default or has an independent control over capture.

The dispute is also about what happens after transcription

Chamberlain alleges that model training was enabled by default for Free and Business accounts, that an account holder had to opt out for future use, and that non-account-holders whose voices were collected had no equivalent setting. The complaint also alleges that data already used for training could not readily be removed from a resulting model.

Granola’s current model-training page makes a more qualified claim. It says Granola may use anonymized data from non-Enterprise users to improve its services; all users can opt out in settings; and Enterprise organizations are opted out by default. It also says agreements with its service providers prohibit them from using customer data for their own model training.

Neither account, on its own, resolves the central factual questions. The documentation does not establish the plan, setting or data flow in Chamberlain’s alleged meeting. The complaint, meanwhile, is a plaintiff’s account rather than a finding that Granola used her data for model improvement. The difference between Granola’s own use of anonymized data and a provider’s prohibited use is material, not semantic.

Granola’s product page depicts AI-assisted editing of meeting notes. Source: Granola product page.

A product advantage becomes evidence in the case

Granola says notes are private by default and that a user controls what is edited, shared and sent into team tools. The complaint alleges a broader set of risks: public sharing links, integrations, and summaries that could extend information beyond the original conversation. Those are allegations that will require evidence about the actual sharing configuration, not proof that every Granola note was exposed.

The company’s no-bot approach is therefore both its commercial pitch and the feature at the center of the case. A visible bot can give other participants a chance to object or leave; Granola says avoiding one produces a more natural meeting. The complaint says that trade-off deprived people of notice and consent. It also says competing notetakers commonly make themselves known, but that market comparison remains an assertion in the filing, not a court-established industry standard.

The case asserts seven causes of action, including claims under the federal Electronic Communications Privacy Act and California’s Invasion of Privacy Act, as well as privacy, unfair-competition and unjust-enrichment theories. The complaint says the proposed classes may include millions of people and asserts more than $5 million in controversy. Those figures are jurisdictional and class allegations, not a damages award or a finding of liability.

What the litigation needs to establish next

The next consequential evidence is operational: product settings at the time of the alleged recordings; what each participant could see or hear; whether notices were activated; the account plans involved; and records of retention, anonymization, sharing and any model-improvement use.

The court will also have to decide whether the claimed consent and interception rules apply to this kind of local-capture workflow and whether Chamberlain can prove common facts across a proposed class. Until then, the case is not a verdict on AI meeting notes. It is a focused challenge to whether the person who enables an invisible notetaker can supply the consent and data-use choices for everyone else speaking.