Google Chrome’s AI team has proposed new APIs called Writing Assistance to directly access language models built into web browsers and operating systems. This could lead to more websites and web applications adding AI writing integrations.
Large language models (LLMs) are all the rage, thanks to OpenAI’s ChatGPT, which sent the entire tech industry into an AI frenzy looking for similar features. But building AI web apps that implement LLMs is easier said than done. Software developers can offload all the heavy lifting to cloud APIs, which costs money, or try to set up local AI models using existing technologies like WebAssembly and WebGPU. Thomas Steiner, Google’s developer relations engineer, announced the effort to provide a better solution on August 19, calling it the first step “towards eventual standardization.”
The Writing Assistance API proposal available on GitHub includes APIs for summarizing input text (Summarizer API), writing new content based on a prompt (Writer API), and rewriting input text (Rewriter API) in the desired way. Google provides the APIs as separate proposals so that anyone interested can evaluate them independently.
The APIs are not yet released for delivery in Chrome, but when standardized and adopted by developers, web apps will instruct your preferred web browser to download certain underlying AI models for writing assistance out of the box, “so as not to lock implementers into a single strategy.” This should make AI-powered web apps more flexible when it comes to leveraging language models to provide text summarization, rephrasing, and rewriting capabilities. The company broke down the most common use cases for each API.
Summarizer API
- Creating meeting minutes for people who arrive late to the meeting.
- Summary of support conversations for entry into a database.
- Summarize product reviews in one sentence or paragraph.
- Summarizing long posts to help you decide if they’re worth the time.
- Creating compelling article titles as a form of summary.
- Aggregating questions on Q&A sites so experts can answer them.
Writer API
- Explanation of structured data such as survey results over time, number of defects by product, etc.
- Expand pros and cons lists into full test reports.
- Create a bio from an author’s resume, a list of previous jobs, etc.
- Write articles based on bullet points or your train of thought.
- Writing product posts for social media.
Rewriter API
- Removing redundant or unnecessary information from the text.
- Make your texts more or less formal for specific target groups.
- Rephrase toxic language in a more constructive way.
- Rewriting articles using simpler words and concepts (“explain it like I’m 5”).
Google says the proposal could allow web apps to rely more on your device’s hardware capabilities and less on servers, potentially making them faster and more flexible. The functionality could work even without an internet connection. We’ve become so used to offline support in web apps like Gmail and Google Docs that we take it for granted, but most AI web apps today are crippled without internet access.
Apple has emphasized the importance of encryption and on-device processing with its own Apple Intelligence features coming this fall. In contrast, Google’s proposal promises to combine the best of both approaches. The company also mentions that the Writing Assistance APIs would reduce API costs for web developers and enable hybrid approaches where, for example, users of a free web app would only be able to access a basic AI model (while paying users would get access to a more powerful API-based model).
Chrome’s AI team will gather feedback in the coming weeks and eventually take the proposal to the Web Incubator Community Group (WICG) for further discussion. The WICG is a discussion forum, not a standards body, so the project still has a few hurdles to overcome before it is submitted to the World Wide Web Consortium (W3C) for final approval. The WW3C will either turn the proposal into the official specification or reject it. If it is accepted, it will be up to browser vendors like Google, Apple, Microsoft, Mozilla and others to implement the entire specification or select parts of it.
Source: GitHub, Mastodon
