Research journal
Show the source. Keep the boundary.
Evidence-reviewed answers about ASL, AI, and education—paired with the failed experiments and open questions that shaped the product.
Primary sources firstIllustrations labeledClaims reviewed Aug. 15, 2026
Verified explainers
Answers built from primary sources and the work we can actually prove.
8 articles01
State of the field
Can AI understand ASL? The honest answer after the Pixel 11 breakthrough
Google reports its consumer feature now translates useful continuous ASL, but that is a vendor benchmark: no independent evaluation yet proves reliable academic intent for any signer or classroom.7 minEvidence reviewed Aug 15, 2026
→02Release analysis
Google released sign-language AI. What does it actually change for Deaf education?
SL2T makes recognition more available. Education still needs age-aware interaction, academic-risk controls, answer checks, and evidence from Deaf learners.7 minEvidence reviewed Aug 18, 2026
→03Academic assurance
What happens when sign-language AI gets the detail that changes the answer wrong?
Translation error rate alone does not measure academic harm. Sign First separates fluent language from the numbers, operations, conditions, and names that determine the task.6 minEvidence reviewed Aug 15, 2026
→04School decision guide
Can AI be an IEP accommodation for a Deaf student?
AI may support an individualized program, but it is not automatically a legally sufficient accommodation and it does not replace effective communication or qualified services.8 minEvidence reviewed Aug 15, 2026
→05Learner experience
Why “ages 5+” cannot mean one AI interface
A beginning reader and a teenager should not perform the same confirmation work. Age changes presentation and support—not intelligence or signer authority.6 minEvidence reviewed Aug 15, 2026
→06Role boundary
The interpreter shortage is real. AI still does not become an interpreter by default.
Workforce pressure can justify responsible automation research. It cannot erase competency, context, legal duties, or the difference between a brief transaction and instruction.6 minEvidence reviewed Aug 15, 2026
→07Foundation
Language access comes before AI access
AI can widen access to information, but it cannot substitute for a child’s early access to a full language or turn a technology choice into a language-development plan.7 minEvidence reviewed Aug 15, 2026
→08Research governance
Why Deaf-led evaluation is a product requirement—not a launch credential
Community participation must shape what gets measured, what failure means, and who can stop deployment. A testimonial after development is not governance.6 minEvidence reviewed Aug 15, 2026
→Field notes
What failed, what changed, and what remains open inside Sign First.
8 notes01
Product principle
Recognition is a proposal
Why fluent text must not become authoritative academic intent without a signer-controlled assurance boundary.0 / 91Critical academic terms recovered by direct video routing in the locked comparison.
→02Research practice
What a failed model teaches
Rejected candidates narrow specific claims without closing the broader recognition program.98.44% → 0 / 23Official-shard top-1 accuracy did not transfer to critical continuous-ASL terms.
→03Camera finding
Confidence is not calibration
A confident camera label can be less trustworthy than an honest abstention.84–86%Wrong live confidence assigned to “F” on unsupported poses.
→04Interaction design
Repair the risk—not every word
The path beyond exhausting full-transcript confirmation is a smaller, academic-risk decision.34 / 34Critical-pair checks recovered in the sealed selective-intervention evaluation.
→05Governance
Learning without silent collection
A future improvement loop is valuable only if ordinary student use never becomes invisible model contribution.38 / 38Exact repair suggestions on covered rows; five revoked rows removed on rebuild.
→06Learner experience
Ages 5+ is not one interface
Reading load can change without changing the learner’s intelligence, meaning, or authority.4 modesImplemented presentation targets from visual-first through advanced review.
→07Academic safety
The answer needs its own boundary
A perfectly confirmed question can still receive a fluent, irrelevant, or wrong answer.2 gatesIntent assurance before the model; response assurance after it.
→08Architecture
Build for the models to change
Recognition should become a commodity without making learner authority one.1 contractCandidate in; signer-confirmed intent out; raw media stops at the boundary.
→