Sign Chat 0.2.0 · experimental beta. Camera and video input for ChatGPT. General ASL understanding and natural-speed conversation are not validated.
Six positive samples from public ASL University teaching photos: three clear correct readings (THANK YOU, NO, YES), two correct tentative readings (MOTHER, FATHER), and one partial reading (HELP, without recovering YOU HELP ME). A blank frame and an ambiguous single-fist still both produced abstentions.
A separate 1.5-second PLEASE animation produced a correct tentative reading. The Pro setting took 79 seconds for it. These were manually attached samples, mostly one instructor, not live camera trials or a representative benchmark. Public images may be in training data.
On 14 November 2026, two 1.5-second clips were imported, reviewed, and sent through the actual ChatGPT widget on the Instant setting. Each delivered 12 sampled frames. The model declined to interpret the known PLEASE sample and correctly identified the blank gray control. A follow-up described visible content from the first clip, supporting image delivery but not recognition accuracy. Both sends left capture paused for the user to start the next phrase.
First complete replies were observed within about 26 seconds and 15 seconds of Send. These are observation upper bounds from two trials, not a latency benchmark. No personal webcam footage or fluent-signer session was tested.
Independent fluent-signer review, new signers, varied lighting, fingerspelling, directional meaning, facial grammar, sentences, natural signing speed, false sends, and supported-device compatibility. The app remains a beta until these are measured. Do not rely on it as an interpreter.
Video stays local until you send, or explicitly enable live mode. Live mode sends one phrase, then waits for you to read the reply and start the next phrase. No model-completion detection or guaranteed latency is claimed.