AI Just Got Much Better at Sign Language

Google DeepMind just published work on a system that roughly doubled the accuracy of sign language translation. That is a harder problem than it sounds: sign language is a visual language with its own grammar, where meaning lives simultaneously in hand shape, movement, facial expression, and body position. Doubling accuracy there is a genuine jump. And it is part of a quieter pattern worth noticing, because it affects your business directly: the accessibility features that were clumsy demos two years ago have become good enough to actually use, and most businesses have not switched them on.

The business case is bigger than most owners assume

Roughly one in five Canadians reports a disability. That is not an edge case you can round off; it is a substantial share of your potential customers and your potential hires. If your booking flow cannot be used with a screen reader, or your videos have no captions, or your phone line is the only way to reach you, you are not making a minor oversight. You are quietly declining business. And unlike most improvements, this one has a second driver: accessibility is a legal obligation in Canada, through the Accessible Canada Act federally and regimes like AODA in Ontario, with other provinces following their own paths.

What is actually usable today

The honest state of play, ranked from mature to emerging. Most of the top of this list is probably already sitting unused in tools you pay for.

Capability	Readiness
Live captions in meetings and webinars	Mature, usually already included
Transcripts of recorded content	Mature, and improves search too
Alt text for images	Good, review before publishing
Plain-language rewriting	Good, helps everyone
Sign language translation	Emerging, improving fast

This is the same pattern as the AI you already pay for and never use, applied to accessibility. The capability shipped. Nobody switched it on.

Raise the floor, keep the humans

Here is the posture that avoids both mistakes. Use AI to raise the floor everywhere: captions on by default, transcripts always available, alt text generated then reviewed, plain-language versions of anything dense. That costs almost nothing and helps far more people than you realize, including everyone watching a video on mute. But do not treat AI as a complete replacement for human accommodation. Automatic captions still stumble on names, technical terms, and accents, and a wrong caption in a medical, legal, or financial setting is worse than no caption. Keep human interpretation available where the stakes are high, the same judgment call we described in where AI can and cannot hold the expert conversation.

On by default, not on request

One small design choice does more work than any tool selection: enable these features by default rather than making people ask. Requiring someone to request captions is itself a barrier, and most people will simply leave instead. Turn them on for everyone, everywhere, and the accommodation stops being a special favour and starts being how your business works. If you run customer-facing AI, apply the same thinking to your chat and voice tools, and keep disclosing when AI is involved so people can decide what they need.

The step everyone skips

Ask. Ask customers and staff with disabilities what would actually help, before you buy anything. It costs nothing, takes an afternoon, and reliably redirects effort away from the impressive-sounding feature toward the mundane thing that was genuinely blocking someone, a form that cannot be tabbed through, a phone-only booking process, a video with no transcript. Accessibility AI has finally got good enough to be worth deploying properly. What makes it land is pointing it at the barriers real people actually hit.

Frequently Asked Questions

What actually happened?

Google DeepMind published work on SL2T, a system that roughly doubled the accuracy of sign language translation. Sign language is genuinely hard for AI: it is a visual language with its own grammar, where meaning lives in hand shape, movement, facial expression, and body position simultaneously, not in a stream of words. Doubling accuracy on that problem is a real jump. It matters beyond the research because it is part of a broader pattern: the accessibility features that were clumsy demos two years ago, live captioning, real-time translation, described images, are quietly becoming good enough for everyday business use.

Why should a small business care about this?

Two reasons, and the second one surprises people. The obvious one is reach: roughly one in five Canadians reports a disability, so inaccessible service is not an edge case, it is a meaningful slice of your market and your potential hires. The less obvious one is that accessibility is a legal obligation here, not a nice gesture. The Accessible Canada Act applies federally, Ontario businesses have AODA requirements, and other provinces have their own regimes. AI is making compliance cheaper and better at the same time, which is an unusual combination worth taking advantage of.

What can we realistically use today?

More than most businesses have switched on. Live captioning in meetings and webinars is mature and often already included in tools you pay for. Automatic transcripts make your recorded content searchable and usable by people who cannot listen. Real-time translation helps both deaf and non-English-speaking customers. Alt-text generation for images is fast and decent. Plain-language rewriting helps people with cognitive disabilities and, frankly, everyone else. Sign language translation is the newest and least mature of these, but it is moving fast enough to plan for rather than dismiss.

Is AI accessibility good enough to rely on?

For assistance, increasingly yes. As a complete replacement for human accommodation, not yet, and treating it that way is the mistake to avoid. Automatic captions still misread names, technical terms, and accents, and a wrong caption in a medical or legal setting is worse than none. The sensible posture is to use AI to raise the floor everywhere, captions on everything by default, transcripts always available, and to keep human interpretation and accommodation available for high-stakes interactions. Ask people what actually works for them rather than assuming the tool has solved it.

Where should a Canadian business start?

Turn on what you already have. Check whether your meeting software, video platform, and website tools include captioning, transcripts, and alt-text generation, and enable them by default rather than on request, because making people ask is itself a barrier. Then look at your highest-traffic customer touchpoints, your phone line, your booking flow, your most-watched video, and fix the biggest gap first. Finally, ask customers and staff with disabilities what would genuinely help. That last step costs nothing and consistently redirects effort toward what matters.