Plaintiffs continue to bring wiretap claims against AI products that capture and process conversations, including voice agents, meeting notetakers, and built-in assistants. Three recent cases—Thompson v. SoundHound AI, Chamberlain v. Granola, and Thele v. Google—illustrate how gaps in notice and consent can give the plaintiffs’ bar opportunities to apply the California Invasion of Privacy Act (CIPA), the federal Wiretap Act, and other state laws to these products. As some early defenses have succeeded, subsequent complaints are adjusting pleading strategies.
The central issue is familiar: what participant consent is needed for companies to collect and use communications. But AI tools introduce additional risk because many times, they are not only recording or transcribing conversations to provide a service, but also using the content of those conversations to train and improve their own AI models.
Thompson v. SoundHound AI: When an AI Agent Answers the Phone
In Thompson, filed in California state court, the plaintiff alleges that she called a Chipotle restaurant; SoundHound’s voice agent answered the call, and unbeknownst to her, transcribed the conversation in real time, collected her name and credit-card information, and retained the information to improve its AI models. SoundHound allegedly also shared call data with outside contractors and processors. Plaintiff argues that consumers have no reason to know SoundHound is involved, much less that their conversations may be used to train AI. The complaint emphasizes that the notice plaintiff heard—that the call “will be recorded”— provides no notice that a third-party AI company is receiving and using the information for its own purposes.
The complaint asserts claims under three CIPA provisions: Section 631, which prohibits unauthorized interception of communications in transit; Section 632, which prohibits recording or eavesdropping on confidential communications without all parties’ consent; and Section 632.7, which prohibits recording certain cellular and cordless telephone communications without all parties’ consent. It also asserts common-law intrusion upon seclusion and California Unfair Competition Law (UCL) claims.
Two features stand out. First, plaintiff uses SoundHound’s privacy policy offensively. Plaintiff quotes statements that SoundHound uses conversational data to improve its services and shares some information with data partners. Based in part on those disclosures, she argues that SoundHound is a third party using communications for its own purposes, rather than merely a tool serving as an extension of the restaurant, which is important for her Section 631 claim. Second, plaintiff alleges only state-claims even though allegations like hers are often used to underly a federal Wiretap Act claim as well. She likely did this to optimize her chances of keeping the case in state court, where standing requirements can be less strict.
Chamberlain v. Granola: An Invisible Notetaker Is Still a Third Party
Chamberlain, filed in the Northern District of California, concerns an “invisible” AI meeting assistant that operates differently from products like Otter.ai.
Rather than sending a visible bot into a Zoom or Microsoft Teams meeting, Granola allegedly captures audio directly from the user’s microphone and computer speakers and generates a live transcript. The Complaint claims that Granola markets the absence of a visible bot as a feature, explaining that other participants will not know the software is there. That selling point animates plaintiff’s theory: if meeting participants do not know an AI notetaker is present, they cannot consent to its collecting their communications. While the Complaint concedes that Granola offers some notification features, including meeting-chat messages and video watermarks, it alleges that these are optional. Customers must affirmatively enable them, leaving other participants dependent on the customer’s decision to provide notice.
As in Thompson, Chamberlain also alleges that Granola uses anonymized meeting data to improve its models by default for certain subscription plans unless the customer opts out. Participants who do not have Granola accounts have no way to exercise that choice themselves.
The complaint asserts claims under the federal Wiretap Act, CIPA Sections 631 and 632, California’s computer-access statute, the UCL, and common-law privacy and unjust-enrichment theories. The Wiretap Act generally permits interception with one party’s consent, but that exception does not apply where the interception is undertaken for the purpose of committing an independent criminal or tortious act. The complaint asserts that Granola’s alleged violations of CIPA and other laws defeat any reliance on the consent of the customer who activated the software.
In other words, whether a vendor is participating in a meeting isn’t necessarily the key question. If a third-party agent is capturing the contents of the meeting in real time, notice and consent may be required to mitigate wiretap and CIPA liability.
Thele v. Google: Standing Remains a Useful Defense
Thele, pending in the Northern District of California, concerns Google’s integration of Gemini AI into Gmail, Chat, and Meet. Plaintiffs allege that Google enabled Gemini-powered smart features by default, allowing the technology to access and analyze private communications without users’ knowledge or consent.
In July 2026, the court dismissed the first amended complaint for lack of Article III standing. Although plaintiffs alleged that Gemini could access sensitive information, they did not identify any personal information it actually accessed or explain how they had been concretely harmed. The court also found that plaintiffs lacked standing to seek an injunction because they could disable the challenged features.
The plaintiffs have since amended their complaint to address those deficiencies. One now alleges that he received Gmail search suggestions that appeared to be AI-generated. Others allege that they discovered smart features enabled without having opted in. The amended complaint also attacks Google’s consent interface as a “dark pattern.” Plaintiffs allege that disabling Gemini requires navigating two separate settings and sacrificing familiar features, including spellcheck and the tabbed inbox. They contrast Google’s default-on approach in the United States with its allegedly default-off approach in Europe.
Google has moved to dismiss again, arguing that plaintiffs still have not identified a concrete injury and that its terms and privacy policy authorize the challenged conduct. That motion remains pending.
The case illustrates both the value and limitations of standing defenses. Allegations about what AI could access are not necessarily enough to establish injury in federal court. But plaintiffs can respond by alleging more specific experiences with the technology, making the product’s actual operation increasingly important.
Takeaways
These cases reinforce four lessons for companies building or deploying AI products:
- Know what your vendors do with customer communications. A vendor’s contractual right to use conversations for model training or product improvement can support an argument that it is an independent third party rather than a passive service provider. Review vendor agreements and understand whether communications are retained, shared, or used beyond providing the immediate service.
- Provide notice before collection begins. Generic recording disclosures may not adequately explain that a third-party AI system is receiving communications and using them for its own purposes. Companies should identify the technology’s role and ensure that users have a meaningful opportunity to consent before the challenged collection occurs.
- Do not assume customers will obtain everyone else’s consent. The person who activates an AI notetaker may agree to its terms, but other meeting participants ordinarily have not. Making notification optional shifts responsibility to users and creates litigation risk. Products that depend on third-party consent should build notice and consent into the workflow rather than leaving them to chance.
- Document what the AI actually accesses and uses. As Thele demonstrates, the difference between hypothetical access and actual collection may determine whether a federal privacy case survives. Companies should be able to explain what their systems access, when they access it, and what happens to the resulting data.