Foresight processes meetings, notes and personal documents directly on the computer, without sending data to the cloud: this is how “on-device” AI (for now only on Mac) guarantees privacy and autonomy even offline.
When we use ChatGPT, Gemini, Claude and other AI-based assistants, normally what we write or upload is sent over the internet to powerful computers hosted in data centers, which process the request and return the response. Google AI Edge Foresight tries to do something different, that is, carry out the entire process inside our computer.
Currently, it is a free experimental application for Mac, designed especially for models with Apple Silicon chips, capable of listening to and transcribing a meeting, sorting out notes and searching for information in personal documents even without a connection. It is not, therefore, a complete version of Gemini that can live without the internet, but a concrete example of so-called AI on-device: Models compact enough to run directly on the user’s hardware, without continually sending data to the cloud.
All local
Foresight uses the microphone and audio produced by the Mac, so it can work alongside any video conferencing platform or during an in-person conversation, without needing a virtual “bot” to enter the meeting. While two or more people talk, the app creates a transcript on the computer and allows the user to quickly take concise notes: a note like “delivery Friday”, for example, can be automatically enriched with details retrieved from what was said a few moments before.
The work is mainly done by Gemma 4 and, above all, from EmbeddingGemma 2a 740 million parameter Google model designed specifically to run on relatively small devices. In practice, instead of sending the audio to a remote server, the Mac performs the necessary calculations itself and once the templates are installed, the transcription and AI functions can continue even with the connection disabled.
Inside the files
Perhaps the most interesting part comes, however, after the meeting. Foresight can build a sort of personal library on your computer made up of documents, images, notes and old transcriptions and allow you to query it using normal sentences. You can then ask, for example, “What did we decide on price last month?” and let the AI search for the answer in the available material.
To achieve this, EmbeddingGemma 2 transforms texts, images and audio into numerical representations called embedding: simplifying, a kind of mathematical imprint of their meaning. Conceptually similar information thus ends up “close” to each other, allowing the program to retrieve a relevant passage even when the question does not contain exactly the same words present in the document. Google calls this function “multimodal search”, because the same system can relate content of different nature.
Privacy and limits
The most obvious benefit is confidentiality: according to Google, recordings, transcripts, notes and personal files used by Foresight are processed locally and do not need to be sent to the company’s servers. This can be important for business meetings, business documents, or simply for those who don’t want to trust private information to cloud services. Processing everything on the device also reduces the delay of communicating with a server and allows you to work on the train, on the plane or anywhere without a network.
There is a price, however: a local model must adapt to the power and memory of the computer and cannot rely on the enormous computing capacity of a data center, so it tends to be more specialized and less versatile than large online assistants. Foresight is still an experiment for macOS and not a new product, but it indicates a possible direction of AI in the coming years, i.e. tools small and smart enough to work directly on the devices in front of us: today a computer, tomorrow a smartphone.
