Google shipped its Prompt API in Chrome 148 on May 5, 2026, embedding a JavaScript interface to its Gemini Nano AI model directly into the browser — and doing so after Mozilla, Apple's WebKit team, the W3C Technical Architecture Group (TAG), Microsoft, and a Chrome engineer who helped approve the feature all filed formal opposition. The decision gives any website on the world's most widely used browser the ability to run AI inference on a user's device without asking for permission, using a 4.27 GB model Google downloaded to those devices without consent.
If the critics are right, every Chrome user — roughly 4.16 billion people worldwide — is now a participant in an experiment that could reshape how the web works, who controls it, and whose content policies govern it.
What the Prompt API Does — and What It Takes From You
The Prompt API lets any webpage send natural language instructions to Gemini Nano, the AI model Chrome now embeds locally. The model runs on the user's device, not on a remote server, which Google pitches as a privacy advantage. Supported tasks include text generation, summarization, classification, image captioning, and audio transcription. The API works on Windows 10 and 11, macOS 13 and later, and Linux, but is not yet available on Chrome for Android or iOS.
The hardware requirements are steep: 22 GB of free storage and a capable GPU for audio support. The Gemini Nano model itself occupies approximately 4.27 GB. Chrome downloaded it to eligible users' machines without prompting them. More pointedly, Chrome re-downloads the model if a user manually deletes it. Google maintains a toggle in Chrome's System settings that can switch the local AI off and delete the model, but the default state is on.
Once Gemini Nano is installed, any top-level webpage can silently invoke AI inference — consuming the user's battery, CPU cycles, and GPU compute — without requesting further permission. Apple's WebKit team identified this as a fundamental design flaw: the site that makes the API call gets the benefit; the user pays the hardware cost.
Every Major Stakeholder Opposed It — Chrome Shipped Anyway
Mozilla developer relations lead Jake Archibald summarised the consensus bluntly in a post on May 6, 2026 that gathered over 2,000 likes and roughly 130,000 views: "Mozilla: Opposed. WebKit: Opposed. Microsoft: Several concerns. W3C TAG: Several concerns. Developers: Mostly negative. Chrome: Ships anyway. A sad time for web standards."
Mozilla's formal position, filed at standards-positions issue #1213, states that the Prompt API has "severe negative consequences to the interoperability, updatability, and neutrality of the web platform." Apple's WebKit team filed a parallel opposition at standards-positions issue #495, citing interoperability, privacy, and battery drain concerns. The W3C TAG raised deep concerns in design-reviews #1093.
In the blink-dev Intent to Ship thread, Microsoft Edge representative Alex Russell argued on May 1, 2026, that the feature should revert to Origin Trial status, citing the risks of deploying a versionless multimodal model in production contexts. By May 5, Edge had quietly disabled the feature entirely — notable because Edge is a Chromium-based browser that normally inherits Chrome features automatically.
Then there is Rick Byers, the Chrome engineer who helped approve the API. In a public GitHub post, Byers acknowledged: "As one of the blink API owner approvers for shipping this in Chromium, I admit that I share the concerns here in Mozilla's standards position." He argued, however, that the cost of waiting outweighs the cost of experimenting, and asked the web community to gather evidence of harm. His position, in effect: ship first and see what breaks.
The Interoperability Trap Developers Are Already Walking Into
Traditional web APIs produce identical behavior across Chrome, Firefox, and Safari because their specifications are precise. The Prompt API cannot make that guarantee. AI model outputs are probabilistic and model-specific. A prompt tuned for Gemini Nano may produce garbled or refused output when run against a different model in a different browser.
Archibald explained the cascade to The Register: "Prompts are tightly coupled to models; developers will inevitably tune to the quirks and policies of whatever model they're building against. That's how you end up with model-specific code paths, which is the browser-compatibility problem all over again." Developers who optimize their prompts for Gemini Nano — the model they can actually test against — will find their applications degraded or broken on Firefox and Safari, exactly as "Best viewed in Internet Explorer" sites were broken on other browsers in the early 2000s.
The Prompt API's own proposal repository is explicit about this: it does not intend to guarantee language model quality, stability, or cross-browser interoperability. Critics call that a remarkable disclaimer for something positioned as a web standard.
Chrome holds roughly 65–68 percent of the global browser market. If Gemini Nano becomes the de facto AI inference engine for the web layer, Mozilla and Apple will face pressure to either license Nano or ship their own compatible APIs — effectively ceding the web's AI tier to a single company's model and governance structure.
Agreeing to Google's Content Policy Just to Use a Browser API
A second objection cuts beyond interoperability. The Chrome documentation requires developers to acknowledge Google's Generative AI Prohibited Uses Policy before accessing the Prompt API. That policy prohibits activities that go beyond what is illegal, including generating content Google deems "disturbing" or material it classifies as misinformation.
Archibald called this out directly: "This seems like a bad direction for an API on the web platform, and sets a worrying precedent for more APIs that have UA-specific rules around usage." Matthias Ott, a web designer whose Bluesky post sharing the critique gathered hundreds of responses, framed the implication starkly: accept that the next browser API may come bundled with its own corporate terms, and the one after that. The web platform has historically been implementable by anyone, governed by specifications rather than a single vendor's usage policy.
Gemini Nano Performs Poorly — and That Is Also the Point
Independent benchmarking reported by The Register compounds the concern. A February 2026 benchmark comparing Chrome's Gemini Nano implementation against Microsoft Edge's Phi-4 mini-instruct found that 15.17 percent of Chrome's responses failed to complete generative tasks, and 23.93 percent of its classification responses were incorrect. Chrome's model hallucinated 6 percent of the time, against Edge's 17 percent — the one metric Chrome won. Both numbers raise serious questions about deploying either model as a default web primitive.
A security investigation published to the DEV Community in May 2026 found the model highly susceptible to prompt injection: hostile content embedded in a webpage can redirect Gemini Nano's output entirely, without any privileged separation between system instructions and user content. Sites that use the API to summarize user-generated content — forum threads, support tickets, comment sections — can have those summaries silently rewritten by any user who can post text.
What the Same Release Got Right
Chrome 148 also shipped lazy loading for video and audio elements — a feature that Squarespace engineer Scott Jehl proposed, developed with cross-browser collaboration, and advanced through the WHATWG HTML specification process with positive signals from all three major browser engines. It passed without objection.
Ott drew the contrast plainly: "One feature shows how the standards process can and should work. The other shows what happens when a company with dominant browser market share decides that the process doesn't apply to them." The Prompt API's call for developer interest, Google's own stated justification for shipping, received two GitHub responses in seven months.
What Readers and Developers Can Do Now
Chrome users who want to disable Gemini Nano can navigate to Chrome Settings, open System, and turn off the local AI toggle; Google says the model will be deleted on restart. Developers building against the Prompt API should treat cross-browser availability as zero and implement a hosted-model fallback for all users on Firefox, Safari, and Edge.
The broader stakes are clearer: whether Gemini Nano becomes the de facto AI layer of the web will be determined not by standards votes but by how many developers ship against it. If those applications work well only in Chrome, the outcome resembles an Internet Explorer recurrence — this time governed by a single vendor's AI model and usage policies. Regulators have not yet formally investigated the Prompt API deployment; whether Google's unilateral rollout to billions of devices constitutes a competition or consumer protection concern remains an open question in the EU and the US.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
