Artificial Intelligence

Google's Sign Language AI Ships, ASL Only

Google DeepMind's SL2T reads sign language from hand and body coordinates alone, sidestepping GDPR biometric-data rules. It ships in American Sign Language only, leaving BSL and DGS users out for now.

Artificial IntelligenceBy Servola Tech Desk2026-08-135 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Google's Sign Language AI Ships, ASL Only

Key takeaways

Google DeepMind's SL2T model translates sign language directly to text and now powers dictation in Gboard and conversation replies in Live Transcribe on the Pixel 11, trained on more than 100,000 hours of data across 50-plus sign languages.
Rather than sending video to the cloud, SL2T runs Google's MediaPipe Holistic pose tracker on-device, transmits only geometric hand, body and face coordinates for translation, and discards the original recording immediately - a design choice that avoids treating a signer's face and body as biometric video data under EU law.
The feature launches with American Sign Language to English translation only; British Sign Language, German Sign Language (DGS), French Sign Language (LSF) and every other national sign language used across the EU and UK are not supported, and Google has not published a timeline for adding them.
ASL and BSL are separate, mutually unintelligible languages despite sharing English as the surrounding spoken language, so an EU or UK Deaf user gets nothing usable from this release even though Google frames it as reaching '70 million' Deaf and hard-of-hearing people worldwide.
The launch lands weeks after the European Accessibility Act's enforcement window opened across member states, raising the compliance bar for digital products sold into the EU without this feature actually closing any of it for the sign languages EU users actually sign.

What Google Actually Shipped

SL2T (sign language to text) is a Google DeepMind model that converts the simultaneous hand, face and body movements of sign language directly into written text, without passing through the intermediate 'gloss' annotations that limited earlier research systems. It launched on 12 August 2026 inside two consumer products on the Pixel 11: sign-to-text dictation in Gboard, and sign-based conversation replies in Live Transcribe. The model was trained on more than 100,000 hours of signing data spanning over 50 sign languages, roughly a quarter of it American Sign Language (ASL), and scores 70 BLEURT on the FLEURS-ASL benchmark, which Google describes as well ahead of any previously reported result. It was built with an AI Sign Language Advisory Committee of Deaf organizations and specialists, and the project was conceived by Sam Sepah, a Deaf Googler, under a stated principle of building with the Deaf community rather than just for it.

Google is upfront about where the model still struggles: rare signs, rapid fingerspelling, passive grammatical constructions, classifier depictions and tense inference without surrounding context all appear on its own published limitations list. That candor matters, because this is not a research paper or a demo video - it is the first sign-language translation model shipped inside a mainstream phone's default keyboard, available to any Pixel 11 owner at no extra cost from day one.

The Privacy Engineering Choice Nobody's Headline Caught

Under EU data protection law, biometric data used to uniquely identify a natural person sits in a special category under GDPR Article 9, with tighter conditions on when it can be processed at all. Video footage of a person's face and hands, captured specifically to analyze their body movements, sits close to that line. SL2T's architecture is built to stay well clear of it: the phone runs MediaPipe Holistic, an on-device pose tracker, and only the resulting geometric coordinates of hands, face and body position are sent to Google's servers for translation. The original video is discarded on the device immediately and never transmitted or stored.

Whether or not Google's engineers framed it in GDPR terms internally, that is a textbook data-minimization design: strip a camera feed down to the smallest abstraction that still does the job, and drop everything else before it ever leaves the device. Any EU-facing product team building a camera-driven accessibility, gesture or dictation feature now has a concrete, shipping example to point to in a data protection impact assessment, and a strong argument for why landmark coordinates, not raw video, should be the default architecture rather than an afterthought.

Why 'ASL Only' Is a Bigger Gap in Europe Than It Sounds

Sign languages are national and regional, not universal, and they do not track the spoken language of the country that uses them. British Sign Language and American Sign Language are separate languages with different grammar and largely different vocabulary, despite both surrounding a country that speaks English. Germany has DGS, France has LSF, Italy has LIS, Poland has PJM, and none of them are dialects of ASL or of each other. Google's '50-plus sign languages, 100,000 hours' figure describes the breadth of SL2T's training data and research base; the feature that actually shipped translates ASL to English only. That means zero EU or UK sign languages are usable in this release, on day one or for an unstated period after it.

The timing sharpens the gap. The European Accessibility Act's obligations became enforceable for many digital products and services across member states from 28 June 2025, with national market surveillance authorities now empowered to act on complaints. The Act does not mandate a specific mobile keyboard feature like this one, so its absence is not a violation on its own. But it does mean a Deaf user in London, Berlin or Paris is reading Pixel 11 accessibility marketing that describes a capability their own sign language cannot use, at exactly the moment digital accessibility scrutiny across the EU and UK is rising rather than easing.

What This Means for EU and UK Deployment Planning

For product and compliance teams, the landmark-only architecture is the pattern worth copying for any camera-driven accessibility, gesture or dictation feature aimed at EU or UK users. It narrows the biometric-data exposure question before legal has to answer it under pressure, and it is a concrete example worth citing in a data protection impact assessment as data minimization by design rather than an afterthought bolted on for compliance.

For accessibility or procurement teams evaluating this specific feature, the advice is more blunt: do not budget on British Sign Language, DGS, LSF or any other EU national sign language arriving soon. Google's own limitations section commits only to 'additional sign languages' with no date attached, and the company's own sequencing here, ASL first with everything else described as following later, suggests EU and UK sign-language support is a second-wave rollout measured in quarters rather than weeks. Treat the Pixel 11 accessibility marketing and the feature that actually ships in a given market as two different things when advising a client on device procurement for Deaf staff or customers.

Frequently asked questions

What is SL2T?

SL2T is a Google DeepMind AI model that translates sign language directly into text. It launched on 12 August 2026 inside Gboard's dictation feature and the Live Transcribe app on the Google Pixel 11, trained on more than 100,000 hours of data spanning over 50 sign languages.

Does SL2T support British Sign Language or other European sign languages?

No. The current release covers American Sign Language to English translation only. British Sign Language, German Sign Language (DGS), French Sign Language (LSF) and other national sign languages used across the EU and UK are not yet supported, and Google has not given a release date for adding them.

Does a camera-based feature like this raise GDPR biometric-data concerns?

Potentially, if it processes video of a person's face and body to identify them or analyze their movements. SL2T avoids much of that exposure by running pose-tracking entirely on-device, sending only abstract geometric coordinates to servers for translation and discarding the original video immediately, rather than transmitting or storing footage of the signer.

SL2T is a real engineering achievement: the first sign-language model shipped inside a mainstream phone rather than a lab demo, built with an architecture that quietly solves the biometric-data problem before any EU regulator has to raise it. None of that changes what a British, German or French Deaf user gets from this release today, which is nothing usable. The feature that will matter to most of Europe's sign-language users is the one Google has not shipped yet, and has not promised a date for.

Google DeepMindSL2TAccessibility AIPixel 11GDPREuropean Accessibility Act

DeepMind's CEO Steps Aside as Three Scientists Exit

Demis Hassabis moves to chairman and Alphabet chief scientist as Jeff Dean, Sanjay Ghemawat and Oriol Vinyals leave to found Discovery Loop. Alphabet stock fell nearly 4 percent, and Koray Kavukcuoglu now runs Gemini day to day.

5 min read

Artificial Intelligence

The Model You Can Use Cannot Move A Robot

DeepMind released Gemini Robotics 2 on 30 July as three models. Only the planning model is generally available, and the published success rates range from 32 to 92 percent.

4 min read

Decisions

The Third Pivot Was the Warning Magic Leap Missed

Magic Leap filed a WARN notice for 193 layoffs effective 1 October 2026, its fourth strategic pivot after $4 billion raised. The lesson is the pattern, not the pivot.