At 7 p.m. in a Changsha living room, a father who cannot hear watches the local news through a small avatar signing in the corner of the screen.
In a county hospital hundreds of miles away, a deaf patient points at a symptom while a nurse types a question into a phone and waits for the reply.
A national audit released this month says that avatar is now common enough to count. It also says the waiting is still the hardest part.
The audit behind the numbers
On November 3, 2026, China's National People's Congress (全国人民代表大会) received a report on enforcement of the Barrier-Free Environment Construction Law (无障碍环境建设法). Delivered to the 25th session of the 14th NPC Standing Committee by vice chairperson Wang Dongming (王东明), it is the most detailed official snapshot yet of how the country's accessibility rules work in practice — including where artificial intelligence is helping and where it is not.
The law took effect on September 1, 2023. The inspection ran from June to July 2026, with field visits to six provinces and municipalities and a third-party assessment by Tsinghua University. That provenance matters: this is not a vendor's press release. It is a legislature's own accounting of what got built.
The AI layer that scaled
The report's most concrete AI finding is short: an artificial-intelligence sign-language television broadcast system (人工智能手语电视播报系统) has been tested at more than 500 units across 30 provincial-level administrative regions, with 150 units in live operation.
Those figures appear in the same document on the official NPC website, and a Hunan provincial government page describing the developer's product independently reports coverage of more than 500 units across 30 provinces. Two public sources, one consistent number.
Behind the avatar is a supply chain that rarely gets mentioned. A Hunan government description credits the developer — a Changsha-based firm linked to the local broadcaster — with a sign-language annotation team of more than 750 people, organized through schools and deaf associations to keep the corpus current. The company has also worked with a university and a major technology firm on a multimodal sign-language model.
- Tested across 30 provincial-level regions
- More than 500 units tested, 150 in live service
- More than 750 sign-language annotators behind the training data
- Over 200 sign-language columns (手语栏目) run by broadcasters nationwide
When the emergency is real
A signed newscast is one thing. A signed emergency is another. The same report notes that the 12110 national public-security short-message alarm number (全国公安机关公益性短信报警号码) continues to give people with hearing and speech impairments a text channel to report crimes, and that help mini-programs (小程序) for deaf users and 120 emergency-alarm platforms have been activated in multiple places.
Shanghai is cited as a model: since February 2020, the city has added a sign-language interpreter to every epidemic-prevention press conference, a practice that became permanent. The report also credits the national broadcaster with producing audio-described and sign-language versions of the Spring Festival Gala for two consecutive years.
What the report says is still broken
The audit is unusually blunt about the gap between installation and use:
- Some government website and app retrofits are cosmetic — hidden entry points and long, convoluted flows leave screen-reader users unable to finish a task.
- Registration pages commonly rely on visual image verification (视觉图片验证) with no alternative, a hard stop for blind users.
- Self-service terminals generally lack voice output and large-text display; touch-only interfaces leave visually impaired users "able to touch but unable to use."
- Rural areas lag sharply, and many older or disabled residents "do not know how, or do not dare," to use smart devices.
- Only a handful of leading companies are willing to invest in accessibility research, because the development cost is high and certification is not mandatory.
The report also found that public understanding of the law is thin: in a Tsinghua survey, more than 90 percent endorsed the idea of accessibility, but only 19.5 percent could describe the law's actual contents.
Why this is an AI story, not a charity story
Accessibility is one of the few areas where a Chinese AI deployment can be judged by a plain metric: can a person who could not previously do a thing now do it? A signing avatar that lets a deaf viewer follow a weather warning passes that test. A government app that fails at the CAPTCHA screen does not.
The same audit points at where the next gains lie — voice recognition, AI sign-language translation, and smarter terminals — while warning that products remain homogeneous and that many sit at exhibitions rather than in daily life.
Honest limitations
This article relies chiefly on the NPC enforcement report, an authoritative government document, plus one provincial government product description. The report gives deployment counts but no accuracy rates for the signing avatars, no viewer numbers, and no evidence that emergency response times improved for deaf callers. "In live operation" describes a deployment state, not a measured outcome. The 750-person annotation figure comes from a government capability listing and should be read as a developer-provided claim. We have not independently verified how the 150 live units are distributed or whether all 30 regions use the same vendor.
What readers can do now
- If you run a public-service website, test a full task with a screen reader and without a mouse, and remove visual-only CAPTCHA first.
- If you work in broadcasting, publish your signing avatar's error rate and keep a human interpreter available for breaking news.
- If you are deaf or hard-of-hearing, save 12110 in your phone today and check whether your city has activated a text-based 120 platform.
