My Vision for AI and the Web
2025-09-01
The way we use the web is shifting fast. Some people are still reluctant to embrace AI, but the reality in 2025 is that it is here to stay. The right response is not to reject it, but to make the best of it and shape it in ways that serve us. Search indices are fading, and chatbots are taking their place. Instead of browsing links, we are asking questions and getting answers directly. That is convenient, but it also moves control away from users and into the hands of a few AI providers.
The Browser’s New Role
Browsers have always been the gateway to the web, but for a long time they have stayed neutral. Now, with AI changing how we find and use information, I believe the browser should become more active. That is why we have started adding local AI features in Firefox that respect privacy by design. For example, the PDF.js alt text generator and the Smart Tab feature both run on small models that we trained, which are downloaded to your device and executed locally. Your data never leaves your machine. This is the kind of approach that keeps people in control, instead of handing that control over to centralized platforms.
The Hybrid Approach
AI will not all run locally yet. Large models still need server-side power. But smaller, specialized models can run on your device today, privately and securely. That is the balance I believe in: local AI for lightweight, privacy-first features, and server AI only when the workload is too heavy. On top of that, privacy on the server side is improving quickly thanks to technologies like GPU enclaves, for example Nvidia’s Confidential Compute. As devices get stronger and these safeguards mature, more intelligence can shift back to the user’s side with stronger guarantees at every step.
Open Questions
There is still much to figure out. One of the biggest changes is that large language models are increasingly acting as intermediaries to the web itself.
Instead of users directly searching, visiting sites, and making their own judgments, LLMs are fetching, ranking, and even rewriting information before presenting it back.
That raises difficult questions:
-
How do we ensure that users remain in control when the LLM is doing the search for them?
-
Can the browser provide visibility into what the model retrieved and why, so users are not left in the dark?
-
What role should open standards play to ensure transparency and user choice when AI providers add search capabilities directly into their platforms?
These are open design questions that don’t have easy answers. But they are central to making sure AI serves people, rather than simply reshaping the web in ways that strip away agency.
The Future
The path forward is clear to me. We need to build powerful hybrid-based features, but always with deep care for security and privacy. That is how we make AI in the browser truly serve the people who use it.