PSA: Join Chrome's built-in AI early preview program

[Kenji Baheux]

Hi everyone,

As you may have heard, Chrome is integrating AI directly into the browser, and we'd like to invite you to be part of Chrome’s built-in AI early preview program to help shape the future of on-device AI in extensions, and the web.

Here's why you might want to join the program:

• Explore: discover innovative AI use cases with Chrome's built-in Gemini Nano model. Your insights will help us prioritize the most impactful on-device AI capabilities.

• Preview: be the first to know about, and experiment with upcoming AI APIs.

• Influence: your feedback will directly shape the APIs, ensuring they best meet your needs and expectations.

If you’re interested, you can sign up here. Once you’ve signed up, you can check out our sample Chrome Extension to see the new built-in prompt API in action.

Looking forward to collaborating with you to enable practical use cases for on-device AI in your extensions! We’ve been working closely with the Extensions DevRel team on this and will be looking out for questions in the mailing list.

Cheers,

Kenji on behalf of the Built-in AI team in Chrome

[Stefan Van Damme]

Hi there,

I signed up, but should there not be permission to inform users that this Chrome extension uses AI, considering privacy concerns?

Thanks,
Stefan vd

[Rajat Paharia]

Hi Stefan - I'm curious what the privacy concerns are since the AI model is running in the user's Chrome browser on their local machine. Nothing is being sent to the cloud. 

- rajat

[Stefan Van Damme]

Hi Rajat,

When a new user installs a Chrome extension from the Chrome Web Store, it is crucial to inform them if the Chrome extension uses AI technology locally on their machine. This transparency ensures users are aware of the functionalities and potential data interactions of the Chrome extension, promoting trust and informed consent. Currently, the demo Chrome extension shows nothing to the user:
https://github.com/GoogleChrome/chrome-extensions-samples/blob/main/functional-samples/ai.gemini-on-device/manifest.json
This practice is similar to disclosing AI-manipulated content in YouTube videos. When using the permissions entry in the "manifest.json" file (such as "ai" as the permission name), this approach is more effective for static analysis tools that assess an extension's capabilities.

Thanks,
Stefan vd

[Jackie Han]

Hi Stefan,

This is a Chrome experimental API for Web, not just for extensions. So there is no new permission in the manifest file.

I know a few Chrome extensions already use this feature.

For privacy concerns, developers can add an on-board page for users.

--
You received this message because you are subscribed to the Google Groups "Chromium Extensions" group.
To unsubscribe from this group and stop receiving emails from it, send an email to chromium-extens...@chromium.org.
To view this discussion on the web visit https://groups.google.com/a/chromium.org/d/msgid/chromium-extensions/07c6218e-81ed-4363-a36a-a1fca959ca63n%40chromium.org.

[Cuyler Stuwe]

AI isn't inherently a data exfiltration threat, so there should be no inherent privacy issue from merely using an on-device AI model.

At the risk of sounding maybe a little reductive: An LLM is no different in nature from a linear regression model (y=mx+b).

Would you warn the user of privacy concerns if you were making a Y prediction given an X coordinate on a linear regression model?

To view this discussion on the web visit https://groups.google.com/a/chromium.org/d/msgid/chromium-extensions/CAAgdh1KMnHBfA%3DRzFGr3x%3DF0K%2BE7EvHr-WqjJFq%3Dy-0_4mLquw%40mail.gmail.com.

[Cuyler Stuwe]

The only scenarios I can think of off the top of my head where it would make sense to give privacy warnings for on-device AI are:

* If data is being sent from the local machine to a server in order to continually train the model.

* If analytics or other telemetry which might contain sensitive information are being sent to a server.

I haven't checked whether this on-device model is doing either of those things, but I'm inclined to believe that it probably isn't and won't (since that would defeat much of the purpose of having an on-device model in the first place).

[Kenji Baheux]

Hi,

Thanks for signing up and bringing this topic to our attention.

Disclaimer: The Prompt API is in an experimental phase and may change based on community feedback. We are committed to fostering responsible use of AI, so our policies and terms of service may evolve over time.

Currently, the Prompt API in itself doesn't have privacy implications because:

• The local LLM does not use any user data, and isn't personalized.
  
• It operates entirely on the user's device.
• We only collect standard usage statistics and crash reports, as per user preferences (see help article).

However, extension authors may choose to include disclaimers, user controls, or privacy policy clarifications if their use of the Prompt API (or any other API) requires it.

I hope this helps.

--

Kenji BAHEUX (my how-to)

Product Manager - Chrome

Google Japan

[Cuyler Stuwe]

On-device AI has no inherent need to transmit any personal data to a server, so as long as they aren’t collecting analytics that could potentially include personal data, there shouldn’t be a need for a privacy warning.

--
You received this message because you are subscribed to the Google Groups "Chromium Extensions" group.
To unsubscribe from this group and stop receiving emails from it, send an email to chromium-extens...@chromium.org.

To view this discussion on the web visit https://groups.google.com/a/chromium.org/d/msgid/chromium-extensions/07414657-56c5-4148-8689-842888e4106an%40chromium.org.