Working research, shown honestly. See what is verified and what is not →Working research · inspect evidence →
Sign First

State of the field

Can AI understand ASL? The honest answer after the Pixel 11 breakthrough

AI can now translate useful continuous ASL in a consumer product, but no single benchmark proves reliable academic intent for every signer or classroom.

By Sign First Research7 minute read
Two flowing signals converge through a controlled passage
Concept art for the recognition-to-intent boundary; it is not a visualization of a model output.

Direct answer

What the evidence supports

Yes—within tested languages, devices, signers, and tasks. Google’s SL2T release is a major advance in continuous ASL-to-English translation. It is not evidence that every model understands every signer, or that raw translation is safe to route directly into educational AI.

Decision trace

What changed—and what still has to be proved before ASL can safely drive an academic AI?

Google DeepMind SL2T announcement · Sign First evidence register

Observed

Continuous ASL reached a consumer product.

SL2T now supports ASL-to-English input in Gboard and Live Transcribe on Pixel 11.

In brief

Three things to keep straight

  1. Recognition is now a real consumer capability, not only a lab demo.
  2. ASL translation must interpret hands, body, face, space, and language—not isolated hand shapes alone.
  3. Academic use needs a separate decision about consequential uncertainty and signer control.
01

What changed in August 2026

Google DeepMind introduced SL2T in Gboard and Live Transcribe on Pixel 11, beginning with ASL-to-English. Google says the model was trained on more than 100,000 hours across more than 50 sign languages, with roughly a quarter of the data in ASL.1

The model translates pose-landmark sequences directly into text rather than using an intermediate gloss. Google reports a zero-shot score of 70 BLEURT on the FLEURS-ASL sd-test benchmark. That is a benchmark result—not a percentage of classroom questions translated correctly.1

02

What the release does not prove

Google documents remaining errors in rare signs, rapid fingerspelling, passive constructions, classifier depictions, and tense without context. Those are not footnotes for education: a missed operation, quantity, condition, or technical name can change the task the AI answers.1

No public evidence in the announcement establishes age-specific classroom outcomes, signer-independent academic accuracy, or an SL2T developer API. Pixel product access and programmatic model access are different claims.1

03

Where Sign First fits

Sign First does not need to outbuild Google’s recognition model. Its durable role is to accept a candidate from an approved recognizer, isolate details that can change the academic task, preserve the signer’s right to repair meaning, and route only confirmed intent.23

Today that involves visible confirmation because the necessary calibration has not been earned. The target is selective intervention: strong recognition should remove friction while the assurance layer becomes quieter, not disappear.2

Questions answered

The short version

Is Google SL2T available as an API?

Google’s release documents Pixel 11 features in Gboard and Live Transcribe. It does not publish an SL2T API, Vertex model identifier, pricing page, or developer terms.1

Does Sign First translate all ASL today?

No. Sign First is a working assurance platform with bounded recognition paths. It does not claim universal continuous-ASL recognition or validated student outcomes.2

Why not send the translation straight to an LLM?

That may be reasonable for low-stakes personal use. In instruction, a fluent translation can still omit the exact detail that determines the answer, so the product keeps recognition and academic intent as separate decisions.

Source register

What this article relies on

  1. Putting sign language AI into users’ handsGoogle DeepMind ↗

    Primary product announcement and stated limitations for SL2T on Pixel 11.

  2. Evidence registerSign First ↗

    Public, bounded record of internal tests, failures, and current product limits.

  3. Recognition research programSign First ↗

    Public model portfolio, rejected paths, and open recognition questions.