Six ways hearing-built sign-language AI fails the Deaf community

I keep a running catalog of how hearing-built sign-language AI fails the Deaf community. It is not a list of other people's sins. It exists so Parley can catch itself the moment it starts to look like one of them.

Parley is my sign-language computer-vision research arm. It is deliberately small, deliberately public, and deliberately not a product yet. Every hearing-led sign-language AI project I have studied fails at some subset of six patterns. The only honest question is whether we notice when we are in one.

Pattern 1 — Capturing motion and calling it translation

The archetype is the ASL glove. It reads finger position, maps to words, outputs text, and gets called a translator. What it misses is most of the language. ASL carries grammar in the face: eyebrows, mouth shape, eye gaze. It carries meaning in body lean, in the use of space, in the direction a sign travels. A system that reads the hands and ignores the rest is processing a fraction of the language and presenting it as the whole. The failure is not that the tech is incomplete. It is that incomplete tech gets sold as complete, and a hearing user has no way to know what is missing.

Pattern 2 — Accuracy theater

This is the one I think about most, because it is the easiest to commit by accident. A project reports a high number, ninety percent, on a test set that shares signers with the training set. The model learned how specific people sign and gets rewarded for recognizing them again. Put it in front of a deaf person it has never seen and the accuracy falls through the floor. The community has watched this cycle for over a decade. The fix is to evaluate on held-out signers and report it plainly, even when the honest number is half the impressive one.

Pattern 3 — The one-way street

Most accessibility products convert speech to text for deaf users and stop there. That is genuinely useful, and many deaf people rely on it. The gap is that the hearing person still cannot understand the deaf person signing. The market decided one-way captioning was good enough, which solved the hearing side's comfort and left the harder problem unsolved. A tool that only runs in one direction is not bidirectional communication. It is captioning with better marketing.

Pattern 4 — Selling it as an interpreter

Some products market themselves as an AI interpreter that replaces a human. What they actually do is pattern-match isolated vocabulary, sometimes fluently on scripted content, never reliably in a medical or legal or emergency setting. Human interpreters are trained professionals with cultural judgment and real liability. The Registry of Interpreters for the Deaf has published positions on AI in interpretation, and they are worth reading before anyone uses the word. Parley is not an interpreter replacement, and the moment its copy implies otherwise it has joined the pattern.

Pattern 5 — Nothing about us without us

A hearing engineering team builds the product, then runs "Deaf user testing" after launch. What got skipped is the part that matters: Deaf people shaping what gets built, which signs, which framing, whether the thing should exist in this form at all. Testing-after is retrofitted accessibility. Co-design is the disability-rights standard, and it has to happen before the architecture is set.

Pattern 6 — The advisor with no authority

This is the late-stage version of pattern five. A single Deaf advisor appears on the About page, named, sometimes paid, with no actual power to block a decision. The name prevents the exact criticism it would otherwise draw. The test is simple: does the advisor have written authority to veto a framing or a deployment, publicly, and to resign with the reason visible? If not, the role is decoration.

The catalog as a checklist

I run these as a self-audit, not a sermon. Do we ever quote accuracy on a split that shares signers? Have we ever called Parley a translator? Does the advisor role carry real veto authority? Is the first capture shaped by Deaf input or by my guesses? When we describe what Parley does, do we describe what it cannot do with equal weight?

None of this makes Parley exempt. Reading the patterns is not the same as avoiding them. The catalog earns its place only if I actually stop when the work starts to match it. That is the whole reason it is written down.