Mozilla's Opposition to Chrome's Prompt API

Mozilla noted that Chrome's Prompt API exposes an in-browser language model as a web API, which could encourage prompt optimization tailored to a specific model's quirks. If developers tune features to match implementations like Edge's Phi-4-mini, this could lead to quality degradation and site blocking on other browsers or models, so Mozilla marked its standards position as negative.

Instead, Mozilla suggested longer validation in userland and web extensions rather than shipping directly as a web API. Mozilla's trial web extension API and the Web Machine Learning group's web extension proposal were mentioned as early feedback paths.

The core concerns are as follows.

• If system prompts become fixed to a specific model, new models could end up looking worse
• Developers could identify specific implementations through LanguageModel.create() and model ID queries, creating browser branching
• Attaching the Google Generative AI Prohibited Uses Policy to the terms of use for a browser API could set a precedent for user-agent-specific rules

On the Chrome side, it was explained that they are trying to reduce these risks through JSON schema/regex-based response constraints, WebML CG discussions, and experiments with alternative models. The Prompt API has shown some level of compatibility across Google and Microsoft models, and Chromium-family projects are also exploring Android MLKit, Apple foundational model integration, Gemma 4 open models, and categorical sampling modes.