Chrome offers several built-in AI APIs. Translator, Language Detector and Summarizer are stable from Chrome 138; other capabilities are at different development stages. Support varies by API, browser version, operating system and device. Treat it as progressive enhancement, not a universal capability.
1. What's actually inside Chrome
Stable task APIs include Translator, LanguageDetector and Summarizer. Chrome uses the LanguageModel API for generative prompts, whose availability can differ by context. Always check the availability state before creating a session.
2. Where you should use it
- Form assist: create a first draft when the required API is available
- Live translation: product descriptions, FAQs, support chats multilingual without an agency
- Summaries: long documents/reviews reduced to one sentence for the visitor
- Language detection: recognize input and route it to the right translation or support flow
3. Where it doesn't (yet) work
Built-in AI is best suited to bounded text tasks. Do not use it as the sole decision maker for financial, legal or strategic advice, current knowledge or other high-risk output. Those cases need additional sources, validation and often a server model.
4. Cost comparison for small business
For a webshop translating or summarizing product copy, the trade-off involves more than the price per call:
- Cloud AI: usage costs by model and volume, with broad browser support
- Professional translation: price per word plus editorial quality control
- Built-in browser AI: no separate per-call fee when available, but development, fallback and review still cost time
5. Browser support and fallback
Status differs by API. Check, for example, await Summarizer.availability() and expose the feature only when it is usable. A server fallback or regular non-AI flow should cover other browsers, mobile devices and unsupported hardware.
6. Privacy is the real selling point
On-device processing can significantly reduce the data sent to an external AI service. That does not make an application automatically GDPR compliant: still review logging, storage, consent, the complete data flow and how generated output is used.
7. What you can build today
Start with one bounded task, such as translation or summarization. Design feature detection, model download, error handling, quality control and fallback from the start. Expand only when usage and results are measurably sound.