Synthetic signing by the minute, and the governance that stays behind when the file leaves
On 9 September 2026, Signapse launched SignStream Chat, a chat-style tool in which a user types an English sentence and receives an AI-generated video of a digital signer rendering it in British Sign Language or American Sign Language, with German Sign Language listed as coming soon. Every user receives sixty seconds of free translation each week and can buy more through pay-as-you-go credits. Alongside it, Signapse has published per-minute pricing for the SignStream API: $2.00 a minute for up to 200 minutes a month, $1.75 for up to 1,000, and $1.50 for up to 4,000, with output generated in fifteen seconds on average. In March, announcing the German signer it is taking to market this month with the streaming integrator G&L, which is showing it at IBC this week as an Innovation Awards finalist, the company’s co-founder Sally Chalk said that “AI sign language is no longer experimental; we can put it in people’s hands.” SignStream Chat is that sentence made literal.
The launch marks a change in the shape of the market, and it is worth being precise about what has changed. Synthetic signing has, until now, reached Deaf people mainly through arrangements an institution chose and paid for: a rail operator’s departure boards, a broadcaster’s stream, a public body’s website. Those arrangements have a buyer who can ask questions, a contract that can carry conditions, and a provider who controls where the output appears. SignStream Chat sells something different. It sells a file, to anyone, by the minute, and the file can go anywhere. This piece argues that Signapse’s published governance, which is more developed than most in the sector, was built for the first kind of arrangement and does not travel with the second, and that the gap is structural rather than a lapse by any one company.
What was launched, in the company’s own words
Signapse is unusually candid about the limits of the new product, and that candour deserves credit before anything else is said. The SignStream Chat announcement states that “because SignStream Chat generates translations in real-time, they aren’t reviewed by our translation team before you receive them,” and recommends the tool “for informal and creative use cases rather than formal or high-stakes content.” The API page is blunter still. Asked whether output is checked by people, it answers: “No, and that’s important. The API is entirely AI-generated, with no human grammar checking or quality assurance.” It describes the service as “not suitable for formal, legal, or official use” and “best for casual messaging or testing new accessibility concepts,” and it directs formal work to a separate, reviewed product. Few vendors in this field state the boundary of their own product so plainly.
The same announcement describes what happens to the output. Videos “can be downloaded in multiple formats, ready to be shared on social media platforms, embedded in content or used wherever sign language accessibility is needed.” The pricing page lists, among the features of the API tier, “Watermark: No.” Signapse’s technical explainer states that its digital signers are drawn exclusively from “native Deaf signers,” used “with expressed written consent” under “a remuneration scheme for use of appearance in commercial settings,” and that the system generates video “indistinguishable from a human signer.” Each of these statements is reasonable on its own. Read together, they describe a product whose output is unreviewed, realistic, unmarked by any visible watermark, modelled on real Deaf people, and designed to be carried out of the environment in which it was made.
Governance built for a controlled deployment
Signapse publishes an AI Governance Framework signed by its chief executive, and much of it is careful. Its use policy sorts sectors by whether human or AI translation is appropriate, states that safeguarding is “always humans,” requires a qualified interpreter for live interactive legal and health communication, and bars AI from assessment, grading or exam content without regulatory approval. Its internal workflow distinguishes reviewed output from instant output, stating that “AI QC is reviewed during the delivery process and Instant AI is reviewed post delivery.” Its labelling commitment reads: “Where we have control, we will label AI generated content for users (e.g. an on-screen icon).” Its ethos states that “signers should retain agency over how any digital likeness is used” and that “AI-generated sign-language output must not imply endorsement, identity, or authorship by any individual signer.”
Every one of these controls assumes something the new product does not supply: that Signapse, or an institutional client bound by contract, controls where the output appears. A sector matrix constrains use when a customer’s deployment can be matched to a sector; it cannot constrain a file after download. Post-delivery review is a meaningful safeguard when the provider can withdraw or replace a clip on a departure board; it cannot recall a copy already posted to a social media account or embedded in someone else’s website. The labelling commitment is qualified, in its own words, by “where we have control,” and a downloaded file is by definition the case in which control has passed. The prohibition on implying a signer’s endorsement is a rule about output, but once the output is a free-standing video the only party positioned to observe the rule is whoever republishes it. None of this makes the framework insincere. It makes it a framework for deployments, now applied to a product that sells files.
A realistic signer without a label
The consequence is sharpest for the people whose likenesses the system uses. A viewer who encounters a downloaded SignStream clip on a social feed sees what the company itself describes as indistinguishable from a human signer, and, on the evidence of the pricing page, no visible watermark telling them otherwise. Whether the files carry an on-screen AI icon or embedded metadata was not stated on the product pages reviewed for this analysis and was not tested. If they carry neither, the viewer has no means of knowing that the person apparently signing never signed those words, and that the words were chosen by an anonymous user who may not know the language. The Deaf signers who licensed their appearance did so under written consent and a royalty arrangement, which is better practice than most of the sector has managed. What that consent covered, and whether it contemplated a self-service tool through which anyone can place any sentence in their hands and publish the result, is not answered in anything Signapse has published about SignStream Chat. The announcement does not say whether the company’s published content moderation policy applies to what users type.
There is a regulatory dimension here as well, though it should be stated carefully. Article 50 of the European Union’s AI Act requires providers of AI systems that generate synthetic video to mark that output in a machine-readable format that allows it to be detected as artificially generated, and requires deployers to disclose clearly when video resembling existing persons would falsely appear authentic. The deployer disclosure obligations applied from 2 August 2026; the provider marking obligation applies now to systems placed on the EU market on or after that date, with a transition to 2 December 2026 for systems already on the market. Whether and how these provisions apply to a United Kingdom provider’s self-service product, whether a machine-readable mark is present in the downloaded files, and whether a digital signer modelled on a real person falls within the deepfake provision are legal questions this analysis does not answer, and nothing here is legal advice. They are, however, questions that any buyer in the European market, including the German broadcasters and institutions to which the forthcoming DGS signer is being marketed, is entitled to put.
Feedback reaches the wrong person
SignStream Chat has a feedback system, and its design reveals the same structural gap from another angle. Users can rate each translation out of five stars, submit written feedback, or record video feedback, and the announcement presents this as a way for users to “help shape the tool as it grows.” The difficulty is who the user is. A tool that turns typed English into signing is most useful to someone who cannot sign, and a person who cannot sign cannot judge whether the signing is correct. The people best placed to judge it, fluent Deaf signers, are generally not the person at the keyboard. They are the audience, and they encounter the output downstream, as a downloaded file, with no rating button and no route back to the provider. A quality signal collected from the party least able to evaluate quality, and unavailable to the party most able to, is not a substitute for review, and the company’s own framework does not claim it is. It does mean that for this product, errors are most likely to be seen by people with the fewest means of reporting them.
Terms that predate the product
The published terms of service on Signapse’s website were last updated on 1 October 2023, nearly three years before SignStream Chat launched, and do not mention it. They grant users a licence to download content “solely for your personal, non-commercial use or internal business purpose,” define content to include video “in the Services,” and bar republishing or distributing content for any commercial purpose without the company’s written permission. The product announcement, by contrast, invites users to share downloaded videos on social media and embed them in content. It is not clear from the public record whether generated videos fall within the terms’ definition of content, whether a small business posting a signed promotion is making commercial use of it, who holds rights in the output, or who bears responsibility when an unreviewed rendering is wrong and someone relies on it. Signapse may well have product-specific terms presented at sign-up that resolve these questions; if so, they are not public, and a buyer cannot read them before buying. The point is not that the terms are defective. It is that the documents a buyer would ordinarily consult to understand rights and responsibility were written for a different product.
From pilots to per-minute
The wider significance of the launch is a matter of interpretation rather than fact, and it should be labelled as such. Signed language AI has moved over a short period from research to demonstration, from demonstration to institutional pilot, and from pilot to enterprise contract. Per-minute self-service pricing is a further step, toward synthetic signing as a commodity input that individuals and small organisations buy without a procurement function standing between them and the product. That shift matters for governance because much of the assurance in this field has been designed to operate at the point of institutional purchase. Procurement questionnaires, contract conditions, deployment reviews and the kind of evidence standards NCG has argued for all assume a buyer who asks. A community group making a signed event notice, a small retailer adding signing to a promotion, or a teacher preparing a class announcement may not know that the product’s own provider recommends against formal use, and nothing in the file they download will tell their audience.
The recommendation that SignStream Chat be used for informal and creative purposes is sensible. But a recommendation is advice to the buyer, not a property of the output. It does not travel with the file, and the people it is meant to protect never see it. Signapse’s own view, stated in its March announcement, is that its technology is “designed to complement, not replace, qualified sign language interpreters,” with human professionals remaining essential in “medical, legal, and highly sensitive contexts.” In a self-service market, whether that boundary holds depends on thousands of individual decisions made by people who are not experts, about content that no one reviews, for audiences who cannot tell what they are watching.
What would close the gap
None of the following presumes that SignStream Chat is harmful, and several of the measures are already anticipated in Signapse’s own framework; the question is whether they reach the file. The first is provenance that survives download: a visible indication that the signer is AI-generated, placed so that ordinary editing does not remove it, together with machine-readable marking of the kind the EU AI Act now requires and emerging content credential standards support. The second is use conditions that travel with the output, meaning licence terms, written for this product and published before purchase, that require AI-generated signing to be identified as such when republished and state plainly which uses are excluded. The third is a reporting route for the audience rather than only the author, so that a Deaf viewer who encounters an erroneous or misused clip can reach the provider without an account.
The fourth concerns the signers themselves: a public statement of whether the consent given by the Deaf people whose likenesses are used extends to open self-service generation, what content is refused at the point of input, and how a signer can see and object to uses of their appearance. The fifth is measurement: published error rates for unreviewed real-time output, disaggregated from reviewed output, so that the difference between the two products is a number rather than a warning. These are questions a buyer, a regulator or a Deaf organisation can reasonably put to any provider selling synthetic signing as a downloadable file. They are put to Signapse first only because Signapse has moved first, and because its published framework already contains the principles against which the answers can be read.
The limits of this analysis
This analysis rests entirely on Signapse’s public web pages, press releases, published governance framework and terms of service as they stood on 11 September 2026, together with public sources on the EU AI Act. NCG has not used SignStream Chat, has not created an account, and has not examined any downloaded output; statements about labelling and marking describe what the published pages say and do not say, not what the files contain. Signapse is among the systems NCG has reviewed under its Sign Language Access Trust (SLAT) Index. Nothing in this piece reports or draws on that review, and SignStream Chat, which launched after it, has not been assessed. The author has previously exchanged public comments with Signapse’s chief executive on the question of appropriate-use limits. If Signapse publishes product-specific terms, labelling practice or evaluation data that answer the questions above, the right response is to read them, and this piece will be updated to reflect them.
Signapse built a governance framework on the premise that the company, or its client, would know where its signing appeared. It has now launched a product whose value lies in that premise no longer holding. The governance did not fail. It stayed where it was written, which is where the company has control, and the signing left.
Sources
1. Signapse, “SignStream Chat: Instant Messaging Translator for Sign Language,” 9 September 2026. https://www.signapse.ai/new/signstream-chat-instant-messaging-translator-for-sign-language
2. Signapse, SignStream pricing. https://www.signapse.ai/signstream-pricing
3. Signapse, SignStream API. https://www.signapse.ai/signstream-api
4. Signapse, AI Governance Framework (Ethos, Content Moderation Policy, Translation Quality Assurance Policy, AI and Human Translation Use Policy). https://www.signapse.ai/ai-governance-framework-2
5. Signapse, “How AI Transforms Accessibility: Automatic British Sign Language Translation”. https://www.signapse.ai/post/how-does-our-ai-technology-work
6. Signapse, Terms of Service, last updated 1 October 2023. https://www.signapse.ai/terms-of-service
7. “Real-Time Accessibility: Signapse Launches AI Digital Signer for German Sign Language,” press release, 2 March 2026. https://natlawreview.com/press-releases/real-time-accessibility-signapse-launches-ai-digital-signer-german-sign
8. G&L Systemhaus, “AI-powered sign language for video: IBC 2026 Innovation Awards finalist”. https://www.gl-systemhaus.de/en/blog/ibc-2026-innovation-awards-finalist-sign-language-for-video
9. European Commission, “Quick Facts: Transparency rules for AI systems”. https://digital-strategy.ec.europa.eu/en/factpages/quick-facts-transparency-rules-ai-systems
10. Morgan Lewis, “EU AI Act’s Transparency Rules: What Went Into Effect on 2 August?”, August 2026. https://www.morganlewis.com/blogs/sourcingatmorganlewis/2026/08/eu-ai-acts-transparency-rules-what-went-into-effect-on-2-august