Google just made signing an input method on a phone. It did not make a human interpreter.
What shipped. Google DeepMind's new SL2T model powers sign input in Gboard and Live Transcribe on the Pixel 11. A user can sign a search, draft a message or document, ask Gemini a question, or compose a response during a face-to-face conversation. The first product release translates American Sign Language into English; more devices and sign languages are promises for later, not current features.
That distinction matters because sign languages are not spoken languages performed with the hands. ASL has its own grammar, including meaning carried by hand shape, body position, facial expression and space. SL2T tracks 130 points across a signer's hands, face and body, then translates that sequence directly into English rather than first reducing each sign to a written label.
The privacy boundary is useful, but not fully on-device. According to the joint technical and impact report, the phone converts camera video into abstract pose coordinates and immediately discards the video. Those coordinates go to Google's servers for translation. Google says it does not retain the signed input or generated text unless a user explicitly joins an evaluation study. Raw images stay on the phone; body, hand and face geometry does not.
The headline benchmark is not the product guarantee. Google reports a zero-shot BLEURT score of 70 on FLEURS-ASL. That is evidence of progress on one ASL-to-English test, not a production error rate across dialects, lighting, devices and people. Independent sign-language-translation research has found that common benchmarks can reward memorization and repeated phrases, while recommending broader datasets including FLEURS-ASL. The safer read is that Google chose a stronger test direction, but I did not find an independent real-world evaluation of the shipping model.
The most important part of the release is the line around it. Google co-authored the impact report with the AI Sign Language Advisory Committee, whose contributors include the National Association of the Deaf, the World Federation of the Deaf, Deaf Professional Arts Network and RIT's National Technical Institute for the Deaf. The report calls SL2T an assistive drafting tool for low-stakes communication. Users preview and edit the English output before sending it, and decide when Live Transcribe shows it to someone else.
The same report explicitly rules out medical care, legal proceedings, police interactions, classroom instruction, job interviews, workplace discipline and government-benefit decisions. It says the tool does not satisfy legal obligations for reasonable accommodations and must not replace qualified interpreters. Institutional cost-cutting is listed as the first critical hazard, not a remote misuse scenario.
Those limits come from observed failure modes. In pre-release testing, evaluators saw invented “ghost text” when a signer paused or another person entered the frame, distortion of number sequences, missed facial grammar, uneven regional-sign recognition and failures under poor lighting or framing. The system follows one signer, handles clips of at most 60 seconds, and does not yet retain conversation history. It was not trained or formally evaluated on people under 18, and testing across motor disabilities remains limited.
The next test is not another demo. It is whether Google publishes error rates for the release people actually use, expands beyond one language pair with the same community oversight, and keeps institutions from treating a cheap input tool as permission to remove human access.
Source graph: Semble source collection