Read the story the way a Deaf person in Hong Kong might. It tells you that a project was built with your community. It tells you that the videos are culturally and linguistically accurate. It tells you that this is good for you, and that a single video that saves one family from a scam is already a win. Then count who is telling you all of this. A bank’s head of digital accessibility. A bank’s digital governance manager. The founder of the company selling the tool. Three voices, three stakes, and none of them belongs to someone who would watch the videos to find out whether a scam is coming.
NCG’s first piece on this story dealt with the number that does not hold up, the claim that the UK has roughly 80,000 sign language interpreters, and with what Hong Kong’s and the UK’s financial regulators expect from customer-facing AI ([read it here](https://www.novaracg.com/2026/11/16/hsbcs-sign-language-ai-story-put-the-uk-at-80000-interpreters-the-real-number-is-about-1500/)). This piece is about everything else in the reporting. None of it is a finding that the HSBC tool is bad. It is a finding that the story gives readers no way to know whether it is good.
Count the voices
The Standard’s story runs to roughly 750 words and contains quotations from three people: Malintha Fernando, HSBC’s Group Head of Digital Experience and Accessibility; Robert Hsu, a senior digital governance manager in the same team; and Tim Lo, founder and chief executive of Glassbox AI.[1] The rest of the story’s background comes from the same parties and from the Hong Kong Science and Technology Parks Corporation, which incubated Glassbox and brokered the partnership. The closing paragraph reports how many business matchings the Science Park arranged at FinTech Week.
Ver dados
| Categoria | People quoted |
|---|---|
| HSBC staff | 2 |
| Glassbox AI | 1 |
| Deaf community members | 0 |
| Independent experts | 0 |
The story says the project was built in collaboration with Hong Kong’s Deaf community, that Glassbox’s database was built in partnership with that community, and that Deaf employees bring lived experience to the product. Not one of those people is named or quoted. There is no Deaf organisation, no Deaf user, no sign linguist and no interpreter in the story. In a piece about whether machine-generated sign language is good enough, the only people who can judge the signing are described but never asked.
This matters beyond fairness. Deaf communities have spent decades insisting on a simple rule: nothing about us without us. A story that cites Deaf collaboration as a credential, while leaving every Deaf collaborator silent, turns participation into a marketing claim.
| Fonte | Role | Interest in the project |
|---|---|---|
| Malintha Fernando | HSBC, Group Head of Digital Experience and Accessibility | Leads the bank’s accessibility ambition |
| Robert Hsu | HSBC, Senior Digital Governance Manager | Works in the same HSBC team |
| Tim Lo | Glassbox AI, founder and CEO | Sells the tool |
| Hong Kong Science and Technology Parks | Incubator and broker | Supported Glassbox and arranged the partnership |
A headline that promises more than the project does
The headline is “Banking in every language, including sign.” The body describes something much narrower: an experiment that turns written web content into Hong Kong Sign Language video, shown publicly as a demonstration on HSBC’s fraud and cybersecurity pages at a fintech event.[1]
That is one direction of communication, from the bank’s text to a customer’s eyes. It is not a Deaf customer asking a question in sign language and getting an answer. It is not a branch conversation, a complaint, a dispute about a transaction or an application for a loan. It is one sign language, not every language. And the story itself calls it an experiment. A reader who sees only the headline, which on social media is most readers, comes away thinking Deaf customers in Hong Kong can now bank in sign language. Nothing in the reporting supports that.
Accuracy as an adjective
The story says Glassbox has built one of the most extensive HKSL databases, that the database accounts for gestures to ensure cultural and linguistic accuracy, and that the models were trained from scratch on local data.[1] Each of those may be true. None comes with a number.
The story gives no error rate, no test result, no sample of users who watched the videos and were asked what they understood, and no comparison with a human-produced HKSL video of the same text. “One of the most extensive” is a ranking with nothing to rank against. Researchers have published HKSL datasets, including a 2024 corpus of about 16 hours of video from sign-interpreted television news, so there is a field to compare against if anyone asks.[6] Nobody in the story did.
Accuracy in sign language translation is also not one thing. A video can use the right signs in the wrong grammar, get the vocabulary right and the facial grammar wrong, or render each word faithfully and lose the meaning of the sentence. The World Federation of the Deaf and the World Association of Sign Language Interpreters make the point directly: because signed languages have their own structures, a word-for-sign exact translation is not possible, and translation has to account for context and culture.[3] A claim of accuracy that does not say what was measured cannot be checked.
Which Hong Kong signing?
Here is the question the story never raises, and it may be the most important one. Hong Kong has more than one way of signing in common use. Natural Hong Kong Sign Language conveys meaning visually and directly. Signed Chinese follows the sentence structure of Cantonese, one sign per word. A principal of a Hong Kong Deaf school told CUHK’s student journalists in 2018 that teachers must use signed Chinese in lessons because they speak and sign at the same time, and Deaf signers interviewed for the same piece said natural sign language is easier to understand.[4]
The HSBC tool starts from written web content, which in Hong Kong means written Chinese or English. A system that turns written Chinese into signs can produce something that follows Chinese word order and looks like signing, while reading to a Deaf HKSL user the way a word-for-word machine translation reads to anyone else. HSBC’s own accessibility lead explained the project on the basis that many Deaf people read Chinese as a second language.[2] If that is the problem, signed Chinese does not solve it. Only natural HKSL does.
Glassbox’s chief executive told FinTech Week that ASL and HKSL are completely different languages.[1] The same logic applies inside Hong Kong. The story should have asked which signing the system produces, who decided, and how that was tested with the people who use it.
The broadcast comparison argues the other way
Robert Hsu framed the project as part of wider social progress: the government uses sign language in public broadcasts and on television, and now HSBC is bringing that inclusivity into financial services.[1]
The government’s sign language in broadcasts is human interpreting. By 2024, the Hong Kong government and the Legislative Council had been providing live sign language interpretation for important meetings for nearly a decade, with the Hong Kong Society for the Deaf playing a pioneering role.[5] The model HSBC points to is people. Even that model has its critics: a Deaf viewer interviewed in 2018 said Legislative Council interpretation still used signed Chinese and that she missed about half of it.[4] If human broadcast interpreting can lose Deaf viewers when it follows Chinese word order, a machine working from written Chinese faces the same risk with less room to adjust. That is a reasonable model, and the comparison is a fair argument for HSBC to commission human-interpreted or Deaf-translated videos for its most important content. It is not an argument that automated signing has reached the same standard.
Designed to scale, but sign languages are not universal
The story reports that while the current focus is Hong Kong, Glassbox’s technology is designed to scale, and that the company is exploring the Chinese mainland, Singapore and Japan, each with its own sign language system.[1]
Earlier in the same story, Lo explains why his company trained from scratch: sign languages are not universal. Both statements cannot carry equal weight. If every sign language needs its own data, its own Deaf contributors and its own testing, then the thing that scales is the company’s process, not the product. Each new market is a new language, and the evidence from Hong Kong, whatever it turns out to be, does not transfer. A reporter could have asked what scaling means in practice: new data from Deaf signers in each country, with what consent and what review, before any video reaches a customer.
Counting only the wins
Lo closed his remarks with an appealing line: fraud does not discriminate, and if one sign language video helps a Deaf customer or their family avoid a scam, that is already a win, because the mission is about impact rather than profit.[1]
That framing counts the upside and leaves out the downside. A fraud warning is short, conditional and specific. If one video gets a condition wrong, telling a viewer when to share something rather than never to share it, that is also one family, and the harm runs in the other direction. A responsible accounting of impact needs both columns: how many people understood the warning correctly, and how many came away with the wrong instruction. The story offers neither. HSBC’s own statement that accessibility helps differentiate its offerings also sits beside the claim that the mission is not about profit.[1] Commercial motive is fine. It means commercial standards of evidence apply.
The data question nobody asked
HK01’s Chinese-language coverage reports that the translation module was trained with local Deaf signers, and that Science Park support helped Glassbox hire Deaf staff to build its sign language database.[2] A sign language dataset is a set of recordings of people’s faces, hands and bodies. Recordings of identifiable people are likely to be personal data under Hong Kong’s Personal Data (Privacy) Ordinance, and the Hong Kong Monetary Authority’s 2024 principles for customer-facing generative AI list data privacy and protection as one of four core areas for banks.[7]
The story does not ask who appears in the training data, what they agreed to, whether they were paid, whether their likeness could be reproduced in a generated video, or whether the same data will train models sold to other banks and other countries. For a community as small as Hong Kong’s signing Deaf community, these questions are not abstract. A recognisable signer is a recognisable person.
What the English story left out
There is one more gap, and this one cuts in HSBC’s favour. HK01 reports that, to ensure the accuracy of the videos, the project engaged SLCO Community Resources, an organisation that provides sign bilingual training and communication support, as a consultant.[2] That is the closest thing in the public record to independent review by people who know the language, and The Standard’s English story does not mention it.
Leaving it out made the project look less accountable than it may be. Including it would have raised the right follow-up questions: what the consultant reviewed, whether every video or a sample, and whether it could stop a video from going live. Good reporting would have helped HSBC here, not only scrutinised it.
What a stronger story would have asked
Coverage of sign language AI is going to grow, and most of it will come from launch events and partner announcements. Reporters, editors and communications teams can use the same short list.
- Quote at least one Deaf user and one Deaf-led organisation who are not paid by the project.
- Ask for an accuracy measure, how it was tested, and with how many Deaf users.
- Ask whether the output is natural sign language or a signed form of the written language.
- Ask whether customers are told the video is AI-generated, and what to do if they do not understand it.
- Ask who appears in the training data and what they consented to.
- Check every comparison figure against a primary source before printing it.
- Ask whether there is a route to a human interpreter.
- Describe a demonstration as a demonstration in the headline as well as the body.
Perguntas frequentes
Who was quoted in The Standard's story on HSBC's sign language AI?
Three people: Malintha Fernando and Robert Hsu of HSBC’s digital experience and accessibility team, and Tim Lo, founder and chief executive of Glassbox AI. No Deaf person, Deaf organisation or independent expert was quoted.
Can Deaf customers bank in sign language with HSBC in Hong Kong?
The public reporting does not show that. It describes an experiment that turns written fraud and financial education content into Hong Kong Sign Language videos, demonstrated at Hong Kong FinTech Week 2025. It does not describe two-way signed communication with the bank.
What is the difference between Hong Kong Sign Language and signed Chinese?
Hong Kong Sign Language is a natural language with its own grammar. Signed Chinese follows the sentence structure of Cantonese, one sign per word, and has been used in Hong Kong classrooms. Content produced from written Chinese can drift toward signed Chinese, and Deaf signers interviewed in Hong Kong have said natural sign language is easier to understand.
Is the government's use of sign language on TV the same as AI signing?
No. Hong Kong government and Legislative Council broadcasts use human sign language interpreters. They are an example of human interpreting, not automated signing.
Was anyone from the Deaf community involved in HSBC's project?
According to the reporting, yes. Glassbox AI hired Deaf staff, and HK01 reports that the organisation SLCO Community Resources was engaged as a consultant on accuracy. The English-language story describes Deaf collaboration but does not quote anyone from the community.
Back to the reader
Go back to the Deaf reader in Hong Kong. Nothing in this critique says the HSBC videos are wrong. They may be very good. The point is that the story gave that reader no way to tell, and gave everyone else, from bank boards to regulators to other banks shopping for the same tool, the impression that the question had already been answered. It had not been asked. The next story about sign language AI, in Hong Kong or anywhere else, can start by handing the camera to the people who sign.
Heather Grizzle-Odland is the founder of Novara Consulting Group, which evaluates sign language AI before institutions deploy it. She is DeafBlind and lives in Minnesota.
Referências
- The Standard. (2025, December 12). Banking in every language, including sign. https://www.thestandard.com.hk/news/article/319092/Banking-in-every-language-including-sign
- HK01. (2025, December 12; updated 2026, June 16). HSBC partners with Glassbox AI to explore use cases for an AI sign language tool to promote accessible finance [in Chinese]. https://www.hk01.com/%E7%86%B1%E7%88%86%E8%A9%B1%E9%A1%8C/60293598/%E6%BB%99%E8%B1%90%E8%88%87glassbox-ai-%E5%90%88%E4%BD%9C-%E6%8E%A2%E7%B4%A2ai%E6%89%8B%E8%AA%9E%E5%B7%A5%E5%85%B7%E7%9A%84%E7%94%A8%E4%BE%8B-%E4%BB%A5%E6%8E%A8%E5%8B%95%E7%84%A1%E9%9A%9C%E7%A4%99%E9%87%91%E8%9E%8D
- World Federation of the Deaf and World Association of Sign Language Interpreters. (2018, March 14; updated April 14). WFD and WASLI issue statement on signing avatars. https://wfdeaf.org/?p=7496
- Leung, J., & Shiu, S. (2018, April 3). Will emphasis on speaking and hearing lead to demise of sign language in Hong Kong? Varsity, The Chinese University of Hong Kong. https://varsity.com.cuhk.edu.hk/?p=12354
- Xinhua. (2024, October 8). Sign language speaks volumes, conveying care in Hong Kong. China Daily Hong Kong. https://chinadailyhk.com/article/594823
- Niu, Z., Zuo, R., Mak, B., & Wei, F. (2024). A Hong Kong Sign Language corpus collected from sign-interpreted TV news. arXiv:2405.00980. https://arxiv.org/abs/2405.00980
- Hong Kong Monetary Authority. (2024, August 19). Consumer protection in respect of use of generative artificial intelligence [Circular to all authorized institutions]. https://brdr.hkma.gov.hk/chi/doc-ldg/docId/getPdf/20241107-1-EN/20241107-1-EN.pdf
