What exists
A working, certified control path anyone can inspect today.FAQ
Clear answers before a demonstration.
What Sign First is, what it is not, what works today, and what still needs proof.
11 / The boundaries
Trust starts with straight answers to the hard questions.
Questions people should ask
No fine print required.
Is Sign First an ASL translator?+
No. It is a meaning-protection layer. It takes what a recognizer thinks was signed, flags the parts that could change the answer, and asks the signer to confirm before AI responds.
Does it replace an interpreter?+
No. It does not replace interpreters, educators, language access, or human relationships. It solves one narrow problem: asking an AI a question in ASL.
What can the demonstration prove today?+
It can show the working control path: sign, review, confirm, answer. It can show the privacy controls with test data. It cannot prove student outcomes or full ASL recognition.
Is the product only fingerspelling?+
No. The in-browser recognizer handles fingerspelling only, as one demo path. The design also accepts full signing, but no public model has yet passed our release gate for student use.
Why does the learner confirm?+
Because smooth text can hide an error that changes the meaning. Confirming everything is the current safety step. The goal is to pause only when the risk is real—not on every question.
Does ordinary use train a model on a child?+
No. Private use is the default. Personal support is a separate choice. So is helping improve the model. Real sign-video contribution is switched off until Deaf-led governance and real approval exist.
Is raw camera video stored?+
No. The camera stream is used for the moment of interaction only. No video or audio is kept.
Which AI models are connected?+
Google Vertex Gemini is the tested live route today. Other providers can plug in, but we do not call them live until they are built and tested.
Is the strongest recognition model hosted online?+
No, by choice. The strongest current model knows only 24 signs and runs on a local machine for in-person demos. There, its measured limits can be shown honestly. Hosting it could make it look like a finished service, and it is not one.
Is the site ready for a school pilot?+
Not yet. A stakeholder demo exists. A real pilot needs Deaf-led governance, proof the recognition works for many different signers, school and family approval, legal review, and age-specific testing.
01
Is this an ASL translator?
No. It is an academic assurance layer.
No. Sign First checks and protects meaning. It does not claim to translate all ASL, and it does not replace interpreters.
02
Is it ready for children?
The technical research build is demonstrable; student outcomes are not validated.
The research build works and can be shown. But no child outcome is proven. Real student use needs Deaf-led governance, school and family approval, and much more evidence.
03
Does use train the model?
No. Private use is the default.
Personal support and model contribution are separate choices. Video is not kept, and real visual-model contribution is switched off.
From today’s boundary to tomorrow’s evidence
The honest answer should change when the evidence does.
A research platform should be precise enough to say no today and structured enough to show exactly what would justify a different answer tomorrow.What unlocks
Bounded personalization and quieter assurance, as evidence and permission arrive.Where this goes
Direct signed access to approved AI, with authority always recoverable.Continue exploring