The phone was still a camera in my hand, not a window into anything else, when I first tried to understand what Google had built. The Pixel 11 Pro’s screen was showing a live wireframe of my own upper body—130 glowing points tracking the arch of my eyebrow, the plane of my shoulder, the angle of a wrist. I was in a conference room in New York, watching as my clumsy attempt at fingerspelling “Hello” was translated, almost instantly, into text in the search bar. The raw video feed was never leaving the device. Only that moving constellation of coordinates was being sent out for translation. It felt less like being watched and more like being mapped. For a moment, the technological achievement was all that mattered.
Then I thought about the history of the language it was trying to read.
The announcement of SL2T, the AI model powering this new sign-to-text feature in Gboard and Live Transcribe, is the kind of news that usually gets filed under incremental tech upgrades. It is not. This is the first time a major consumer product has been architected from the ground up to reflect a fundamental linguistic truth that hearing institutions spent over a century denying: American Sign Language is not English performed with the hands. It is a complete, natural language where grammar and meaning are built from the simultaneous movement of hands, face, and body. By designing a system that tracks 130 landmarks across the face, torso, and hands, Google has finally built a piece of consumer technology that aligns with what Deaf linguists have asserted since 1960.
To grasp why this architectural choice is a quiet revolution, you have to understand what came before. The most persistent misunderstanding in tech has been the sign language glove. It reappears cyclically in engineering labs and crowdfunding campaigns, hailed as a breakthrough. Deaf linguists patiently explain each time why it fails. A glove captures only handshapes, missing the facial expressions, head tilts, and use of space that form ASL’s grammar. It is like trying to understand a spoken sentence by only reading the vowels. Google’s team, advised by a committee of Deaf experts, rejected this reductive approach from the start. They abandoned the old method of first converting signs into written English glosses, arguing that glosses flatten the very non-manual markers that carry meaning. Instead, their on-device computer vision creates that kinematic map, treating the entire signing space as the source material.
This technical respect is paired with a rare and necessary caution in the launch. Google is explicit about what SL2T is not. It is an assistive input tool for low-stakes, daily communication—a text message, a web search. It is not a replacement for qualified human interpreters and does not fulfill legal obligations for reasonable accommodation in settings like courtrooms, classrooms, or doctor’s offices. The model, trained on over 100,000 hours of data across 50 sign languages, has clear limits. It can struggle with:
- Regional signs
- Slang
- Rapid fingerspelling
- Complex grammar
- Ghost text generation
- Low light or insufficient contrast
These caveats are not weaknesses in the press release. They are the product of a more honest conversation about what technology can and should do. “This is groundbreaking technology, and as a Deaf ASL user, I’m excited to see ASL recognized computationally as the complex language it is,” says Chrissy Marshall, a Deaf writer and director. “But I can’t help thinking about how this technology fits into our existing accessibility landscape.” That landscape, she notes, has always run on Deaf labor. “Deaf people already carry so much of the communication burden: watching captions, filling in what was missed, correcting errors, advocating for accommodations.” As a society, I think we are far too comfortable expecting Deaf people to rely on imperfect communication tools. Especially when they are ‘futuristic’ or ‘cool.’
Marshall’s insight cuts to the core question for any assistive technology: Who adapts to whom? “Will an AI model understand the way I sign, or will I have to adapt my language to be understood by the technology?” she asks. The difference is everything. A system that performs best when the user standardizes their signing toward its training data has merely shifted the burden of accessibility back onto the Deaf person. That is a pattern as old as Deaf education itself, tracing back to the 1880 Milan Conference where hearing educators voted to ban sign language in schools, insisting speech was the only path to enlightenment.
There is also the practical matter of access. The most sophisticated sign language model ever shipped is currently exclusive to the Pixel 11 lineup. “Innovation only goes so far if the people who could benefit most can’t afford or access it,” Marshall observes. She points to the parallel history of automatic captions, which were launched with similar disclaimers but are now routinely treated by institutions as “good enough” compliance, forcing Deaf individuals to constantly advocate for the professional captioning they are legally owed.
What makes SL2T different, and worth a deeper look, is that it emerges from a collaboration that centers Deaf expertise rather than assuming it. The concept originated with Sam Sepah, a Deaf Googler. The model was trained to handle one-handed signing—crucial when your other hand is holding the phone—and to recognize both left and right-handed signers. This is technology built with, not just for. Its greatest promise may not be in flawless translation, but in offering another tool for spontaneous, everyday communication where no interpreter is present. “I know imperfect technology can still be incredibly useful because I don’t have an interpreter with me every day,” Marshall says.
The story of technology and sign language has too often been one of well-intentioned hearing people building simplified solutions to a complex reality they did not fully grasp. Google’s SL2T is not the end of that story, but it may represent a turning point. It acknowledges the language’s complexity in its engineering, acknowledges its own limits in its messaging, and places the tool in users’ hands with a measure of humility.
| Feature | Description |
|---|---|
| Technology Type | AI model for sign-to-text translation |
| Tracking Points | 130 landmarks |
| Training Data | Over 100,000 hours |
| Supported Languages | 50 sign languages |
| Device Exclusive | Pixel 11 lineup |
| Intended Use | Low-stakes, daily communication |
The model will improve. It will expand to more devices. The boundaries of what it can handle will grow. But the standard it sets today—for technical fidelity, for ethical transparency, for collaborative development—is what matters. For once, the technology is not asking the language to become more legible to the machine. It is trying, earnestly and imperfectly, to learn how to read.
