Picture a Deaf retiree in Kwun Tong who grew up signing HKSL. Written Chinese is not her first language, and the warning that arrives on her phone about a “security verification” is full of terms she half knows. HSBC’s own accessibility lead has described the problem this way: for many Deaf people, reading Chinese or English feels like reading a second language, and complex fraud information can slip past them.[2] She is the person this project is meant for. She is also the person who carries the cost if a machine-generated signing video gets the warning wrong.
That is why the details matter, starting with a number printed in the story that introduced the project to English-language readers, and continuing with what the public record does and does not show about how the bank intends to make sure the signing is right.
What HSBC and Glassbox AI announced
On 12 December 2025, The Standard reported that HSBC, Hong Kong’s Deaf community and the start-up Glassbox AI were “experimenting with” an AI-powered text-to-sign language generation tool. The aim is to turn complex web content, such as financial education and anti-fraud information, into HKSL videos.[1] The collaboration debuted at Hong Kong FinTech Week 2025 with a demonstration of sign language videos on HSBC’s Cybersecurity and Fraud Hub. Glassbox AI was founded at the Hong Kong Science and Technology Parks Corporation, which provided computing power, mentorship and funding access, and the company has hired several Deaf employees through the Science Park’s inclusive hiring programme.[1]
The Chinese-language coverage in HK01 adds a detail the English story left out. To check the accuracy of the sign language videos, the project brought in SLCO Community Resources, an organisation that provides sign bilingual training and communication support, as a consultant.[2] HK01 also reports that the translation module was trained with local Deaf signers.
Credit belongs where it is due. Glassbox’s chief executive, Tim Lo, made a point many vendors skip: sign languages are not universal, and American Sign Language and HKSL are completely different languages. The company says it trained its models from scratch on local data.[1] Building for HKSL rather than adapting a model built for another sign language is the right starting point. Hiring Deaf staff and engaging a Deaf-community organisation are real commitments, not decoration.
The story also tells us what the project is for. HSBC’s Group Head of Digital Experience and Accessibility, Malintha Fernando, said the bank’s ambition is “to become the world’s most digitally accessible bank,” and that accessibility “allows us to differentiate our offerings.”[1] Glassbox says its technology is designed to scale, and that it is exploring mainland China, Singapore and Japan, each with its own sign language.[1] This is a group-level ambition with a regional pilot attached, which is why the questions below matter well beyond Hong Kong.
The number that does not hold up
The Standard justified the project with a comparison. Hong Kong’s Deaf community faces “a stark shortage of sign language interpreters,” the story said, with “just 50 to 60” serving the entire city, “compared with roughly 80,000 in the UK.”[1]
The Hong Kong half of that sentence holds up as a registry count. A 2024 study by the Centre for Sign Linguistics and Deaf Studies at the Chinese University of Hong Kong found 55 interpreters on the register kept by the Hong Kong Council of Social Service.[3] The UK half does not. RNID, the UK’s national hearing loss charity, puts the number of registered sign language interpreters in the UK at around 1,500, set against up to 87,000 Deaf BSL users, which it calculates as roughly one interpreter for every 60 Deaf signers.[4]
The printed figure of 80,000 sits close to the number of Deaf people who use BSL, not the number of people who interpret it. The likeliest explanation is that a population figure was carried over as a workforce figure. Whatever the cause, the story overstates the UK interpreter workforce by a factor of about 50.
Saib Cov Ntaub Ntawv
| People | Nqi |
|---|---|
| UK interpreters as printed in The Standard | 80000 |
| UK Deaf BSL users (RNID estimate) | 87000 |
| UK registered sign language interpreters (RNID) | 1500 |
| Hong Kong interpreters on the HKCSS register (CUHK, 2024) | 55 |
This is not pedantry. The comparison does work in the story: it makes Hong Kong look like an outlier and automation look like the obvious answer. The accurate picture is less convenient. Both places are short of interpreters. The UK has roughly one registered interpreter for every 60 Deaf BSL users. A shortage is a reason to look for help. It is not evidence that a particular tool works.
The Hong Kong side has a second problem the story did not mention. The same CUHK study found that Hong Kong has no interpreter accreditation system. The register counts hours worked, at least 200 hours over two years, and the researchers note that it does not guarantee quality. Of 185 interpreters surveyed, fewer than 10 percent had ever worked full-time and more than 70 percent had interpreted on a voluntary basis.[3] That matters for an AI project in a specific way. If there is no accredited standard for human HKSL interpreting, there is no ready-made benchmark for checking whether machine-generated HKSL is good enough. Someone has to build that benchmark, and the public record does not say who did.
What the public record does not show
Read The Standard and HK01 side by side and a pattern appears. The project is described in terms of intent: community collaboration, cultural accuracy, inclusion, scale. It is not described in terms of evidence. Neither story reports any of the following.
- Accuracy figures, error rates or any measure of translation quality.
- Comprehension testing with Deaf HKSL users, including older signers, signers with low vision, and DeafBlind people who use tactile or close-vision signing.
- A human review step before each video is published, or who holds sign-off.
- Whether the signer in the videos is a synthetic avatar, a generated human likeness, or something else.
- Whether the videos are labelled as AI-generated, and what customers are told about their limits.
- A route to a human: an interpreter, a video relay call, or a staffed channel for Deaf customers who do not understand a video.
- How the content will be kept current as scam patterns change, and who re-checks a video when the underlying text changes.
The SLCO consultancy is the closest thing to quality assurance in the record, and it is a good sign. But a consultant engaged “to ensure accuracy” is a promise, not a method. What did the consultant review: every video, a sample, or the vocabulary database? Could the consultant stop a video from going live? Those are the questions that separate Deaf involvement from Deaf authority.
Hong Kong’s own regulator already asked these questions
This is a Hong Kong pilot, so the first rulebook to check is Hong Kong’s. On 19 August 2024 the Hong Kong Monetary Authority wrote to every authorized institution, which includes The Hongkong and Shanghai Banking Corporation, with guiding principles for generative AI in customer-facing applications. The circular extends the HKMA’s 2019 principles on big data analytics and AI across four areas: governance and accountability, fairness, transparency and disclosure, and data privacy.[5]
A system that generates sign language video from text and shows it to customers appears, on the circular’s plain terms, to be a customer-facing generative AI application. Whether HSBC and the HKMA treat it that way is for them to say. But the circular’s expectations read like a checklist for this project.
| HKMA expectation | What the public record shows |
|---|---|
| Board and senior management remain accountable for GenAI-driven decisions and processes | Senior HSBC staff are quoted on ambition. No accountable owner for output quality is named. |
| Validate models so outputs are accurate and not misleading | No accuracy figures or validation method published. An outside consultant was engaged to check accuracy. |
| Keep a human in the loop during early deployment | No human review step described. |
| Where practicable let customers opt out of GenAI and request human intervention | No human or interpreter route described. |
| Disclose the use of GenAI and explain its purpose and limitations | No labelling or disclosure plan described. |
| Avoid unfair bias against any customer group | No testing across age groups or sign language variation described. |
The HKMA has also expected banks since 2018 to implement the Hong Kong Association of Banks’ Practical Guideline on Barrier-free Banking Services, which covers customers with hearing impairment. A second edition, endorsed by the HKMA, was issued in January 2026.[6,7] An AI video on a web page does not replace the obligations in that guideline. It sits on top of them.
What the FCA’s October 2026 paper adds
HSBC Holdings is headquartered in London, and the executive quoted in The Standard holds a group role. If this tool travels with HSBC’s group-wide accessibility ambition into the UK, it will have to work in BSL, and it will land in a regulatory system that has already described what good and poor practice look like.
In October 2026 the Financial Conduct Authority published AI in financial services: good practice and areas for improvement.[8] Two cautions keep the comparison fair. The paper says it creates no new requirements, and it was published about ten months after the HSBC story. Holding a December 2025 pilot to an October 2026 paper would be unfair. But the paper rests on rules that were already in force: the Consumer Duty, in force since July 2023, and the FCA’s 2021 guidance on the fair treatment of vulnerable customers.[9,10]
Several of the paper’s examples of poor practice describe gaps visible in the HSBC record. The FCA lists assuming AI outputs are clear without testing comprehension across groups, including people with accessibility needs. It lists overstating AI accuracy or reliability, and being unclear whether AI or a person produced something. It lists automated journeys with no way to reach a human and no channel switching. And it lists relying on untested vendor assurances.[8] On the good-practice side, it describes AI drafting clearer communications with staff approving all final wording, telling consumers clearly when they are using an AI tool and what it cannot do, and giving consumers a choice between AI and non-AI support.[8]
Translate the staff-approval point into sign language and its meaning sharpens. A hearing compliance officer cannot approve a signed video. Approval of signed content requires a fluent Deaf signer, and in HKSL, in the absence of any accreditation system, the bank would need to define what fluent means and who decides.
Fraud warnings are the hardest place to start
The project began with fraud content for a sympathetic reason. HK01 reports that Hong Kong recorded about 44,000 fraud cases in 2024 with losses of about HK$9.15 billion, and that Deaf people are more easily targeted because of information gaps.[2] Fraud education is a genuine need.
It is also the place where a translation error does the most harm. A fraud warning is short, precise and conditional: never share this, always check that, call this number and no other. A mistranslated condition does not produce a clumsy sentence. It produces wrong instructions about money. The World Federation of the Deaf and the World Association of Sign Language Interpreters said in 2018 that they are particularly concerned about signing avatars where information is complex or of significant importance to the lives of Deaf citizens.[11] Fraud prevention for a bank’s customers fits that description.
There is a second risk worth testing, and it is a hypothesis rather than a finding. HSBC’s own deepfake guidance tells customers to watch for unnatural mouth shapes, limited facial expression and odd lighting as signs that a video may be fake.[12] A bank that teaches Deaf customers to trust synthetic signing video for fraud warnings may also be teaching them to trust synthetic signing video in general. If HSBC-branded signing videos become familiar, a convincing fake one becomes a new route for the scams the project is meant to prevent. Before scaling, the bank should be able to say how a Deaf customer can tell a real HSBC signing video from a counterfeit.
What a defensible version looks like
None of this requires HSBC to stop. It requires the bank to publish the evidence that would let Deaf customers and regulators trust the output. A pilot that wanted to meet both the HKMA’s principles and the FCA’s description of good practice would be able to show the following.
- Comprehension testing with Deaf HKSL users across age groups and education backgrounds, with results published in plain language and in HKSL.
- Error rates for the generated signing, measured against a benchmark that Deaf HKSL experts defined.
- A named Deaf HKSL reviewer who approves every video before release and can block one.
- Clear labelling of each video as AI-generated, with a short statement of what it cannot do.
- A visible route to a human: a booked interpreter, a video relay option or a staffed channel.
- A process for re-checking videos when the underlying fraud guidance changes.
- A way for customers to verify that a signing video really comes from HSBC.
- Contract terms that require Glassbox AI to notify HSBC of significant model changes, and an exit plan if the vendor changes course.
- A Deaf advisory body with decision rights, not only consultation rights, before the tool moves into another sign language.
Questions HSBC is positioned to answer
- Are the HKSL videos live for customers, or only demonstrated at FinTech Week?
- Is the signer a synthetic avatar or a generated human likeness?
- What did SLCO Community Resources review, and could it stop a video from being published?
- How many Deaf HKSL users have tested the videos for comprehension, and what were the results?
- Who at HSBC is accountable for the accuracy of the signed content?
- Are the videos labelled as AI-generated?
- What should a Deaf customer do if a video is unclear?
- Has HSBC assessed the pilot against the HKMA’s August 2024 principles on generative AI?
- What evidence will HSBC require before the tool is used in another sign language, including BSL?
Cov lus nug uas nquag nug
What is HSBC doing with AI sign language in Hong Kong?
HSBC is testing an AI tool from the Hong Kong start-up Glassbox AI that turns written content into Hong Kong Sign Language videos. The first use shown publicly was anti-fraud content on the bank’s Cybersecurity and Fraud Hub, demonstrated at Hong Kong FinTech Week 2025. Public reports describe it as an experiment and do not say whether it is live for all customers.
How many sign language interpreters are there in the UK?
RNID puts the number of registered sign language interpreters in the UK at around 1,500, against up to 87,000 Deaf BSL users. The figure of roughly 80,000 UK interpreters printed in The Standard in December 2025 is wrong.
How many sign language interpreters are there in Hong Kong?
A 2024 study by the Chinese University of Hong Kong counted 55 interpreters on the register kept by the Hong Kong Council of Social Service. The register is based on hours worked, and Hong Kong has no interpreter accreditation system, so the number says little about quality.
Does the FCA's AI guidance apply to HSBC in Hong Kong?
No. The Hong Kong pilot is supervised by the Hong Kong Monetary Authority. The FCA’s October 2026 paper on AI in financial services applies to FCA-regulated firms in the UK and states that it creates no new requirements. It becomes relevant if the tool is used with UK customers.
What does the HKMA expect from banks using generative AI with customers?
The HKMA’s circular of 19 August 2024 expects board and senior management accountability, validation so outputs are accurate and not misleading, a human in the loop during early deployment, an option to opt out and reach a human where practicable, disclosure that generative AI is in use, and safeguards against unfair bias.
Do Deaf organisations accept AI signing avatars?
The World Federation of the Deaf and the World Association of Sign Language Interpreters caution against using signing avatars as a replacement for human signers, and are particularly concerned where information is complex or important to Deaf people’s lives.
Back to Kwun Tong
Go back to the retiree with the phone. HSBC is right that she deserves fraud warnings in her own language, and right that Hong Kong does not have enough interpreters to provide them by hand. What she does not have yet is any public evidence that the video she would be shown says what the text says, any label telling her a machine made it, or any number to call if it does not make sense. The project started with the correct language and the correct community. It now needs the evidence that turns a demonstration into something a Deaf customer, and a regulator, can rely on.
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.
Cov chaw siv
- 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
- The Chinese University of Hong Kong. (2024, May 7). CUHK surveys the status of sign language interpreters in Hong Kong. https://cpr.cuhk.edu.hk/en/?p=172605
- RNID. (n.d.). Inclusion facts. Retrieved November 16, 2026, from https://rnid.org.uk/what-we-do/data-and-evidence/inclusion-facts/
- 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
- Hong Kong Monetary Authority. (2018, March 23). Practical Guideline on Barrier-free Banking Services of Hong Kong Association of Banks [Circular]. https://brdr.hkma.gov.hk/eng/doc-ldg/docId/20180323-1-EN
- Hong Kong Monetary Authority and Hong Kong Association of Banks. (2026, January 21). Banking industry introduces enhancements to elderly-friendly and barrier-free services [Press release]. https://www.hkma.gov.hk/eng/news-and-media/press-releases/2026/01/20260121-3
- Financial Conduct Authority. (2026, October). AI in financial services: Good practice and areas for improvement. https://www.fca.org.uk/publications/good-poor-practice/ai-financial-services
- Financial Conduct Authority. (2022, July). PS22/9: A new Consumer Duty. https://www.fca.org.uk/publications/policy-statements/ps22-9-new-consumer-duty
- Financial Conduct Authority. (2021, February 23). FG21/1: Guidance for firms on the fair treatment of vulnerable customers. https://www.fca.org.uk/publications/finalised-guidance/guidance-firms-fair-treatment-vulnerable-customers
- World Federation of the Deaf and World Association of Sign Language Interpreters. (2018, March 14; updated April 14). WFD and WASLI statement on use of signing avatars. https://wfdeaf.org/wp-content/uploads/WFD-and-WASLI-Statement-on-Avatar-FINAL-14032018-Updated-14042018.pdf
- HSBC Hong Kong. (n.d.). Deepfake scams. Retrieved November 16, 2026, from https://www.hsbc.com.hk/help/cybersecurity-and-fraud/deepfake-scams/
