Mozilla Opposes the Prompt API

本站收录: 2026-05-18

[00:00:00]Google recently shipped an LLM prompt API to the web platform in Chrome48. At Misilla, we oppose this. So, I wanted to explain why. Now, the API itself is welldesigned. You create a model and you prompt it for a response. There's more to it, but this is the basic usage. In Chrome, when you create the language model, it will either use Google's cloud API or download a 4 GB Gemini model to run locally. Here are our main concerns.

[00:00:22]First up, interoperability. When you create a language model, you'll usually give it a system prompt to define its behavior. And if you've done this before, you'll know you'll end up iterating on it based on the quirks of the model you're testing. Based on Chrome's market share and that they use their Gemini model exclusively, usage of the prompt API on the web will be entirely tailored to Google's model.

[00:00:39]Like another browser could ship a superior model, but it will be perceived as worse by users because the system prompts in the wild are tailored for Google's model. This puts other vendors in a situation where they have to either license Google's model or ship a model that is somehow quirks compatible with Google's. The next issue is neutrality.

[00:00:57]According to Chrome's documentation, to write code with the prompt API, you have to acknowledge Google's generative AI prohibited uses policy. And this policy goes beyond law. For example, it bans generating sexually explicit content.

[00:01:09]And it bans content in vague terms that could be misleading about things like governmental or democratic processes.

[00:01:15]Here, Google is setting a precedent where as a browser maker, they could take action against you because someone visited your website and as a result, a supposedly open web API was called with perfectly legal content, but content that Google as a company doesn't like.

[00:01:33]We believe this to be against the mission of the open web and an abuse of power. The final issue is developer sentiment. Part of the reason Google gave for shipping this API is that developers are strongly positive about it. The key evidence they gave for this was a survey. They don't state the number of developers that were surveyed or how they were selected, but the data they did provide shows that they were asking about this API existing in browser extensions, not on the open web.

[00:01:59]It feels like if you summarize this evidence using the prompt API, you would be facilitating misleading claims and therefore breaking Google's terms of service. LLMs are clearly powerful, but the usage we've seen in the wild is usually using a predictable model chosen by the developer. Now, some developers are positive about this API, but many of you are as concerned about it as we are.

[00:02:20]WebKit folks have objected to this API for similar reasons. Microsoft have similar concerns and have decided even though Edge is based on Chromium, not to ship this API in Edge. The W3C tag raised several concerns. So, it's pretty sad that despite all of this, Google just went ahead and shipped this API in Chrome just in time for Google IO. We are going to continue representing web developers and fighting for the health of the open web by opposing this API.

[00:02:44]Now, if we've gotten this wrong, let us know in the comments. We are always open to feedback. And if you agree with us, I mean, let us know that, too. It's always nice to hear. So, that's where things are now. If anything major changes, I will keep you updated.