Why We Need Web Standards for AI (Before It's Too Late)

Last week, OpenAI's CEO Sam Altman stood on stage at Dev Day and announced something that should concern anyone who values the open web: ChatGPT is becoming an "operating system."

Not a chatbot. Not a tool. An operating system where apps run entirely inside the chat window, where commerce happens through ChatGPT's "agentic commerce protocol," and where—in Altman's words—users "start your day with ChatGPT" before being guided to other apps.

OpenAI isn't alone in this vision. Google's Gemini ambitions, Apple's Intelligence integration, Microsoft's Copilot expansion, and Amazon's Alexa ecosystem all point in similar directions: AI as the new platform layer that sits between users and the open web.

If this sounds familiar, it should. In the 1990s, we had AOL, CompuServe, and Prodigy—walled gardens where you accessed content through proprietary interfaces. Then came the web, with open standards like HTTP and HTML that let anyone publish, anyone browse, and no single company control the experience. Now, thirty years later, we may be watching the same pattern repeat with AI: ChatGPT's plugins, Claude's MCP, Gemini's extensions, Copilot's integrations—each requiring developers to learn different APIs, keeping users locked in, and acting as gatekeeper for what tools users can access.

What We're Losing

Without standards, users get stuck with whichever AI platform has the tools they need, unable to switch providers. Developers must integrate with every platform separately, subject to changing terms and policies. And the open web itself risks becoming a second-class citizen—a source of raw data to feed AI models, but no longer the primary interface through which people interact with digital services. As users spend more time inside AI platforms' chat windows, the web gets marginalized to backend infrastructure.

The Browser Alternative

Here's a different vision: What if browsers—not AI companies—mediated between AI models and web services?

Imagine if web services registered their capabilities using standard APIs, just like how websites use standard HTML. Imagine if AI models accessed those capabilities through the browser, the same way websites load via HTTP. Imagine if users could choose their AI provider without losing access to tools, just like you can switch browsers but still access the same web. And imagine if the browser enforced security and privacy, exactly as it does for web pages today.

This approach offers significant advantages over current "agentic browsing" implementations that parse DOM trees or use vision models to interpret web pages. Websites explicitly define what capabilities they expose through standard APIs—giving them control rather than having AI scrape everything. Direct API calls are faster and more reliable than parsing HTML or simulating clicks. AI models receive only the specific data needed—a medical site could let AI book appointments without exposing patient records visible on the page. And because capabilities use standard browser APIs rather than proprietary integrations, they work with any AI model, not just the one your platform vendor prefers.

This isn't hypothetical. The technology exists. Service Workers, Permissions API, CORS—the web platform already has the building blocks for browser-mediated AI.

Why This Matters Now

The trend is clear across the industry. When OpenAI's head of product says "we never meant to build a chatbot; we meant to build a super assistant," it echoes similar ambitions from other tech giants. Google is integrating Gemini deeper into Chrome and Android with "agentic browsing" capabilities. Apple is weaving Apple Intelligence throughout iOS with its Foundation Models framework. Microsoft is making Copilot central to Windows and Office with direct app integration. Amazon has Alexa+ with its autonomous action APIs.

Every major platform appears to be racing to build their proprietary AI ecosystem, lock in developers, and own the user relationship.

Here's the technical lock-in: Each of these platforms is creating proprietary APIs that only their AI can use. Google's Chrome integration APIs work only with Gemini. Apple's Foundation Models framework works only with Apple Intelligence. Microsoft's Copilot integration APIs work only within their ecosystem. Without open standards, if you want AI that can interact deeply with Chrome, you must use Gemini. If you want AI integrated into iOS apps, you must use Apple Intelligence. If you want AI that controls Office applications, you must use Copilot.

This isn't just about business strategy—it's about technical architecture. The absence of standards means each vendor builds their own incompatible interfaces, and users get locked in at the technical level, not just the commercial level.

But there's a window. AI integration into browsers is still early. Users haven't formed unbreakable habits. Network effects haven't fully kicked in.

If browser vendors and standards bodies act now—if we push for open APIs that let any AI model access any web service, mediated by the browser—we can preserve the open web's core principles: user choice over vendor lock-in, permissionless innovation over gatekeeping, decentralization over platform control, and interoperability over fragmentation.

The Choice We Face

The question isn't whether AI will integrate deeply with our digital lives. It almost certainly will.

The question is: Will that integration happen through open standards that preserve user control and developer freedom, or through proprietary platforms that could concentrate power in a few companies?

History teaches us that once platforms become entrenched—once network effects lock in—they're nearly impossible to dislodge. Just ask anyone who's tried to leave Facebook, or switch from iPhone to Android with all their app purchases.

We're at the fork in the road. The path we choose now could shape the next thirty years of computing.

Do we want the AI era to look like the open web—where anyone can innovate, users control their experience, and no single company sets the rules?

Or do we want it to look like the pre-web era of walled gardens—just with better natural language interfaces? In that future, the web could become marginalized: a backend data source that AI platforms scrape, but not where users actually spend their time or where developers build their businesses.

The choice is ours. But only if we make it soon.

The future of AI doesn't have to be a collection of walled gardens. But making it open requires action—before the platforms become too powerful to challenge.

References