Google's experimental AI Edge Foresight app transcribes meetings entirely on-device

Google ships an offline AI note-taking app

Google has released an experimental note-taking application that can transcribe meetings and audio files without any of the processing happening in the cloud. According to The Verge, which cites an earlier TechCrunch report, the app is called Google AI Edge Foresight, is free to use, and runs on macOS powered by the company's compact EmbeddingGemma 2 model executed directly on the machine.

The release puts Google into an increasingly crowded category of meeting transcription and note-taking tools, but with a twist that most rivals lack: nothing is sent to a server for inference.

What the app actually does

Foresight combines several functions that AI note-taking products typically offer separately. The Verge compares it to existing apps such as Granola and Wispr Flow, which also summarize meetings alongside a workspace for manual notes.

The core workflow works like this: the app transcribes a meeting as it happens and produces a summary, while also giving the user a place to write. Google says a user can type quick shorthand bullet points during a conversation, and the app will expand them into what the company calls "polished notes" using the full transcript as context. In effect, the transcript fills in the gaps the note-taker skipped.

Beyond transcription, Foresight includes a built-in assistant that can answer questions about what was said in a meeting. Users can also connect local files to the app, and those documents are consulted when the assistant formulates its answers.

Privacy as an architectural property

The most notable aspect of the release is not any single feature but where the computation happens. Because the model runs on-device, the privacy claim rests on architecture rather than policy. Google states that files, meeting audio, and notes "never leave your computer." That framing matters for users in legal, medical, financial, or corporate settings where recording a meeting and uploading it to a third-party service may be restricted or simply unacceptable.

Cloud-based transcription services generally require audio to be transmitted to servers, which means trusting a provider's data handling, retention terms, and processing agreements. An on-device approach sidesteps that chain of custody entirely, though users still have to trust that the app behaves as described.

Availability is narrow for now

The Verge notes that Foresight is currently optimized only for Macs with Apple Silicon. The app is also explicitly labeled experimental, which typically signals a limited feature set, possible rough edges, and no guarantee of continued development or eventual wider rollout. There is no indication in the reporting about versions for Intel Macs, Windows, or mobile platforms at this stage.

Why it matters

Foresight is a meaningful data point in the shift toward on-device AI. Small language models such as EmbeddingGemma 2 are now capable enough to handle genuinely useful tasks like transcription, summarization, and question answering without cloud GPUs, and hardware like Apple Silicon provides the memory and neural acceleration to run them comfortably.

For Google specifically, the app shows the company exploring a privacy-first positioning it has not always been associated with. Meeting audio is among the most sensitive data categories an AI product can touch, and demonstrating that a competent assistant can operate entirely locally is a statement about where consumer AI tooling is heading.

It also raises competitive pressure on startups in the note-taking space. Granola and Wispr Flow already operate in a market where privacy differentiation is valuable, and a free entry from Google that keeps data on the user's machine — even an experimental, Mac-only one — could pull attention away from subscription-based rivals. Whether Foresight graduates from experiment to product will depend on adoption, but the underlying message is clear: local inference for everyday productivity tasks has arrived.