Sign-language AI leaves the lab and lets Deaf users sign instead of type

The launch turns signing into a system input

On Pixel 11, users can choose American Sign Language in Gboard and sign directly into supported text fields. The same capability can support search, messaging, documents, and requests to Gemini. Live Transcribe can translate a nearby signer into English text.

The initial product boundary is narrow: ASL to English on one device generation. Training across more than 50 sign languages should not be confused with launch support for those languages.

Privacy starts before the video leaves the device

MediaPipe Holistic processes camera frames on the phone and extracts geometric pose landmarks. DeepMind says only those coordinates reach the server and raw video is discarded immediately.

That architecture reduces exposure but does not eliminate privacy questions. Pose data, translated text, retention rules, cloud processing, failure logs, and downstream application access still deserve clear controls and independent review.

Community participation is part of technical quality

The project began with a Deaf Googler and incorporated Deaf data partners, user research, language experts, an advisory committee, and an impact report. That process helps surface product requirements a benchmark alone can miss, including left-handed signing and using a phone with one hand.

DeepMind also names failure modes instead of marketing universal translation. Rare signs, rapid fingerspelling, some grammar, fine-grained hand details, and tense remain difficult. Expansion should be paced by performance evidence and participation from each language community.

• Report performance separately for every supported sign language and user group.
• Keep raw video processing on device wherever possible.
• Give users control over storage, correction, and downstream sharing.
• Treat Deaf community governance as an ongoing product function.