The Assumption Worth Dropping First
Most organisations approaching accessibility for the first time assume that captions solve deafness. Add a caption track, obligation discharged.
Captions are essential and they do serve a very large audience. But they rest on an assumption that does not hold universally: that the viewer reads the written language fluently.
For many deaf people, a signed language is their first language and the written language of the surrounding country is a second one, learned without the auditory scaffolding that hearing children get for free. Reading fluency in that second language varies widely. For a signer whose written-language literacy is moderate, dense captions running at broadcast reading speed are hard work in a way they are not for a hearing person reading the same text.
This is why sign language interpretation is a distinct accessibility provision rather than a redundant one, and why WCAG lists it as its own success criterion at Level AAA rather than folding it into captions.
It also cuts the other way. Sign language does not serve everyone who is deaf or hard of hearing. People who lost hearing later in life usually do not sign. People who are hard of hearing rather than deaf typically read the written language fluently. A signed track alone leaves those audiences unserved.
The two provisions are complements, not alternatives.
Signed Languages Are Languages
The second assumption worth dropping is that sign language is a manual encoding of speech.
Signed languages are natural languages with their own grammar, their own vocabulary, and their own histories. Their structure is not derived from the surrounding spoken language.
They are not universal. American Sign Language, British Sign Language, Auslan, Langue des Signes Française, Deutsche Gebärdensprache and dozens of others are mutually unintelligible to varying degrees. ASL and BSL are not variants of each other despite both existing in English-speaking countries; they come from different lineages and a fluent signer of one does not understand the other.
They have their own grammar. Signed languages use spatial relationships, simultaneous articulation, facial grammar, and directionality in ways spoken languages do not. Word order and morphology follow their own rules.
Facial expression is grammar, not affect. Eyebrow position, mouth shape, and head movement carry grammatical information — marking questions, conditionals, negation, and intensity. A rendering that gets the hands right and the face wrong is producing ungrammatical output, not merely flat output.
Signing space carries reference. Signers establish locations in space for entities and then refer back to them by pointing or directing verbs. This is a referential system with no direct spoken equivalent.
The practical consequence for anyone commissioning signed content: this is a translation task between two languages with different structures, performed by a qualified interpreter or a Deaf translator, and it takes the time that translation takes.
What Sign Language Provision Looks Like
Three common production models, with different costs and quality profiles.
Picture-in-picture interpretation. An interpreter is filmed against a plain background and composited into a corner of the frame, or presented in a synchronised second video window. This is the standard broadcast and web approach. It is straightforward to produce, and the main quality variables are the interpreter's size in frame — too small and handshapes become unreadable — contrast against the background, and whether the composite obscures on-screen content.
Presenter-led signing. The person on camera signs, either as a Deaf presenter working in their own language with voice-over added, or as a bilingual presenter. This produces the best experience for signing audiences because the signing is primary rather than an overlay, and it eliminates the split-attention problem of watching two things at once. It requires the content to be planned this way from the start.
Studio-recorded translation. A Deaf translator produces a signed version of the content in a studio, working from a prepared script rather than interpreting live. This produces higher quality than live interpretation, because the translator can consider phrasing, and is the usual approach for content that will be watched many times.
The general rule is that live interpretation is appropriate for live events and prepared translation is appropriate for anything that will be published and rewatched. Publishing a recording of live interpretation is common, cheap, and produces a noticeably weaker result than a prepared translation.
Where AI Translation Actually Helps
There is a lot of noise in this space, and being precise about it matters.
Automated sign language generation — avatars producing signed output from text — has been demonstrated for years and remains, for general content, well short of usable. The reasons are structural rather than a matter of more training data: facial grammar, spatial reference, and the fluid simultaneity of natural signing are hard to synthesise convincingly, and Deaf communities have consistently reported that avatar output ranges from effortful to incomprehensible. Some narrow, highly templated domains — transport announcements, weather — have produced acceptable results because the message space is tiny and fixed. General content has not.
Treating that as a solved problem, or presenting an avatar as equivalent provision, is a mistake that harms the audience it claims to serve.
What AI translation does contribute is everything upstream and adjacent.
Accurate transcripts, fast. Sign language translation starts from a script. Automatic transcription with a review pass produces that script in a fraction of the time manual transcription takes, and the translator starts from clean text rather than working from audio.
Caption tracks that meet the other half of the need. The audience that needs captions and the audience that needs signing overlap only partially. Producing accurate captions in every language you publish serves the hard-of-hearing and late-deafened audience directly, and it does so at a cost that scales.
Multilingual reach for the spoken track. A signed version in one country's sign language serves that country. Localizing the spoken and captioned versions extends reach into other markets, where local signed provision can then be added where the audience justifies it.
Terminology consistency. Signed translation of technical content depends on agreed signs for domain terms. Extracting and locking terminology from the source content gives interpreters a glossary to work from, which measurably improves consistency across a series.
Speaker separation. Multi-speaker content needs clear attribution for the interpreter to work from. Automatic diarization applied to the transcript makes the script usable.
The honest framing is that AI handles the language work it can do well and clears the path for human signed translation, which it does not replace.
Deciding What to Provide
Not every organisation needs signed versions of everything, and pretending otherwise leads to nothing being produced at all.
A workable prioritisation:
Provide captions everywhere, without exception. This is the baseline, it is a Level A requirement, and it serves the largest share of the deaf and hard-of-hearing audience.
Provide signed versions where the stakes are highest. Safety information, health information, legal rights, emergency communication, essential public services, and anything a person must understand rather than merely access. These are the contexts where reduced reading fluency has real consequences.
Provide signed versions where the audience is specifically Deaf. Content aimed at Deaf communities should be produced in the relevant signed language as a first-class version, not retrofitted.
Consider signed versions for flagship content. A homepage explainer, an annual report summary, or a major announcement carries symbolic weight beyond its viewing figures.
Work with the community rather than around it. Deaf-led organisations in the relevant market can tell you which signed language is appropriate, what production standards their community expects, and where the real gaps in your provision are. Commissioning without that input regularly produces technically-present provision that the audience does not use.
Production Details That Matter
Interpreter size in frame. Handshape and facial grammar must be readable. A thumbnail-sized interpreter in a corner of a small player is decorative rather than functional. Broadcast guidance in several markets specifies minimum proportions of frame height; follow the local one.
Background and clothing contrast. Plain, non-patterned background, clothing that contrasts with skin tone, no reflective jewellery. This is standard interpreter filming practice and it exists because hands need to be legible against the body.
Framing. Signing uses space around and above the head and out to the sides. Framing that crops the signing space cuts off grammar.
Frame rate and compression. Rapid handshape changes suffer under aggressive compression and low frame rates. Signed content is more sensitive to encoding quality than talking-head video.
Player controls. Where signing is a separate synchronised track, the viewer must be able to enable it, size it, and position it. Where it is burned in, they cannot.
Do not crop or letterbox after the fact. Reformatting signed content for a different aspect ratio without reframing regularly cuts hands out of frame.
A Working Checklist
- Treat captions and signed provision as complementary — neither substitutes for the other.
- Identify the correct signed language for each market rather than assuming one signed language serves all English-speaking or all European audiences.
- Commission qualified interpreters or Deaf translators; signed translation is language work.
- Use prepared studio translation for published content and reserve live interpretation for live events.
- Start from a reviewed transcript so the translator works from clean text.
- Supply a locked terminology glossary for technical or series content.
- Frame the interpreter large enough for handshapes and facial grammar to read clearly.
- Use plain backgrounds, contrasting clothing, and framing that includes the full signing space.
- Avoid heavy compression and low frame rates on signed tracks.
- Deliver signing as a toggleable, resizable track where the platform supports it.
- Prioritise signed provision for safety, health, legal, and Deaf-audience content first.
- Do not present avatar-generated signing as equivalent provision for general content.
- Involve Deaf-led organisations in deciding what to produce and reviewing what you ship.
Frequently Asked Questions
Are captions enough for deaf viewers?
For many, yes. For signers whose first language is a signed language and whose written-language literacy is moderate, dense captions at broadcast reading speed are genuinely hard work. Captions are the essential baseline and serve the largest share of the audience, particularly people who are hard of hearing or became deaf later in life. Signed provision serves a different part of the same audience. Both are needed for full coverage.
Is sign language universal?
No. American Sign Language, British Sign Language, Auslan, and dozens of others are distinct languages, and several of them are mutually unintelligible despite existing in countries that share a spoken language. ASL and BSL come from different lineages. Choosing the right signed language for a market is exactly as consequential as choosing the right spoken one.
Can AI generate sign language automatically?
Not usably for general content. Avatar-based generation struggles with facial grammar, spatial reference, and the simultaneity of natural signing, and Deaf communities consistently report output ranging from effortful to incomprehensible. Narrow, templated domains such as transport announcements have produced acceptable results because the message space is small and fixed. Presenting avatar output as equivalent provision for general content is not defensible.
Where does AI translation help with signed content?
Upstream. Fast accurate transcripts give translators a clean script to work from, terminology extraction gives them a consistent glossary, speaker separation makes multi-speaker scripts usable, and automated captioning and dubbing handle the spoken-language accessibility and reach that sits alongside signed provision. The signed translation itself remains human work.
Should we sign everything?
Almost no organisation does, and treating it as all-or-nothing usually results in nothing. Caption everything, then prioritise signed versions for safety, health, legal, emergency, and essential-service content, for anything aimed specifically at Deaf audiences, and for flagship communications. Expand from there based on what the community tells you is missing.
Can we publish a recording of live interpretation?
You can, and it is better than nothing, but it is noticeably weaker than a prepared translation. A live interpreter is working in real time without the chance to consider phrasing. For content that will be watched repeatedly, commission a studio translation from the script instead — the cost difference is smaller than the quality difference.
Related reading: Video Accessibility Guide | WCAG Video Compliance | SDH Captions Explained
