The Corpus Question, Part Two: When the Signer Is the Supplier

In September, The Corpus Question asked whose signing builds sign language AI, what those signers agreed to, whether they were paid, and whether they can take their faces back. That piece was about vendors. This one is about the rest of us, because the corpus no longer comes only from vendor studios. It comes from the upload box.

The upload box

Picture a Deaf small-business owner on a weekday morning. She runs a cleaning company, or a tutoring practice, or a little shop that sells hand-dyed yarn, and she has watched hearing competitors fill their feeds with slick AI video for a year. This fall she finally signs up for an AI course taught in ASL by Deaf instructors, designed for people like her. She learns to write a decent prompt. She builds a consistent character. Then comes the session she has been waiting for: making a sign language avatar. She props her phone against a coffee mug, records herself signing a welcome message for her website, opens a video model in her browser, and drags the clip into the upload box.

That drag is the moment this article is about. Nothing about it feels like signing a contract, and nobody in the room is thinking about one. But the clip she just uploaded holds her face, her hands, her regional signing, the way her eyebrows climb on a question and her mouth shapes an intensifier. It is a sample of a language, recorded at high resolution, attached to a name, an email address and a payment card. Under the terms she clicked through, it now sits on a server in Singapore or in a joint venture’s cloud, licensed to a company whose business is building better video models.

A trained model cannot hand a face back.

NCG

The courses are not the problem. The better ones teach real judgment about output: smooth, confident-looking motion is not the same as accurate meaning, a fluent reviewer checks the render against a real signing source, and throwing out a bad render counts as a win rather than a failure. That is the right lesson, and it is more than most vendors teach their own customers. But it is a lesson about what comes out of the machine. The corpus question is about what goes in, and what goes in is governed by a contract nobody in the room has read.

What Kling’s paper says

Kling is the video generation business of Kuaishou, one of China’s largest short-video platforms. It reports more than 60 million creators since its June 2024 launch, and in July 2026 it raised close to three billion dollars in a round that left Kuaishou holding about 68 percent^[7]. The contracting entity is Kling AI Pte. Ltd. in Singapore. The terms are governed by Singapore law, disputes go to arbitration under the Singapore International Arbitration Centre, and the privacy policy says user data is stored on servers in Singapore and may be transferred elsewhere^[2,3].

Section 4.4 tells users they own the intellectual property in their content, and section 4.6 says that using the service does not transfer any portrait rights a user holds. Those are real protections and they deserve to be read first. Then comes section 4.7. It grants Kling a “non-exclusive, royalty-free right and license” over content, defined to include both what a user uploads and what the model generates, running “during the operational period of its business.” Section 4.7.2 lets Kling use content for promotion, product upgrades and research, directly or by licensing it to third parties, and asks the user not to claim personal or property rights against that use. Section 4.7.3 then lists the purposes for which Kling may process input, and item (f) is the one that matters here.

create, test, improve, train, or otherwise develop the artificial intelligence or machine learning models

Kling AI Pte. Ltd., Terms of Service, section 4.7.3(f), effective 21 April 2026

Kling does offer a way out, which is more than many platforms do. Section 4.7.4 says a user who no longer wants Kling to keep using their content may “notify us to revoke the authorization by sending an email to support@kling.ai.” What the terms do not say is what revocation reaches.

The privacy policy adds a second tension. Section 1.A(2) says that for portrait features such as character face mode, Kling “may analyze the material you upload and upload the feature points,” described as vector points of the eyes, nose and mouth and contour lines. Section 1.A(6) then says Kling does not analyse facial recognition features or identify individuals from pictures, and does not “collect, share or store any face data.” A reader is left to reconcile a policy that uploads the geometry of a face with a policy that says it holds no face data. The policy never addresses training on inputs directly. It relies on a legitimate interest in “product development and internal analytics purposes” and states no retention period beyond “as long as necessary”^[3].

What Dreamina’s paper says

Seedance is ByteDance’s video model. Its 2.0 release reached American users through CapCut and Dreamina in the spring of 2026, after a stretch in which viral clips misusing real people’s likenesses earned it the trade-press label of the most controversial AI video model in the world. During that rollout CapCut said it would restrict generation from images or videos containing real faces^[6]. That restriction tells you the company understands how sensitive a face is. It does not tell you what happens to the face you upload once the restriction lifts. That is a question for the contract.

Dreamina’s terms, last updated 22 January 2026, name TikTok USDS Joint Venture LLC as the operator and California law as governing. Section 8 tells users that, as between them and the joint venture, they own their inputs and the outputs generated from them. In the same section they grant a licence that reaches well beyond the company.

You grant us (and our affiliates, successors, and assigns) and other users of the Platform a perpetual, irrevocable, worldwide, non-exclusive, transferrable, no-charge, royalty-free, and sub-licensable, right and licence to use your Assets for the purpose of providing the Services.

TikTok USDS Joint Venture LLC, Dreamina Terms of Service, section 8, updated 22 January 2026

The purpose limit at the end of that sentence is real and should be stated plainly. So should the rest of it: perpetual, irrevocable, sublicensable, and extended to every other user of the platform. The training permission sits in the privacy policy, last updated 30 April 2026, which says Dreamina may use information to “train and improve our technology, such as our machine learning models and algorithms,” and defines user content to include prompts, uploads and generated responses. There is no training opt-out. The only opt-outs offered cover marketing email, cookies and push notifications^[4,5].

How the clause got here

None of this is new to ByteDance’s creator tools. In June 2025, CapCut drew wide coverage for terms reported to reach users’ voice, face and likeness whether or not a video was published. CapCut answered that the clause had not changed and applied only as far as necessary to provide the service^[8]. Both of those things can be true at once, and that is the point. A clause written for selfie filters and dance trends now governs the recording of a language.

The face is still the data

The first Corpus Question made one argument that everything else hung on: in a signed language the face is not decoration around the message, it is part of the message^[1]. Blur the face and you have not anonymised the signer. You have damaged the sample. This publication’s own sister series is named for that fact, the fifth parameter of a sign after handshape, location, movement and palm orientation.

Non-manual markers

The grammatical work a signed language does with the face, head and body rather than the hands: raised or furrowed eyebrows marking question types, mouth morphemes carrying size, manner and intensity, eye gaze pointing to people and places set up in signing space, and head tilts and nods marking conditionals, topics and role shift.

General-purpose video platforms were not built around that fact, and their documents show it. Their privacy language is written around identification: whether the company recognises who you are. Kling says it does not identify individuals. Dreamina says it does not use face information to identify you in your content. Those are statements about recognition. They say nothing about the grammatical content of a face, and grammar is what a sign language avatar needs. A model that learns to produce a convincing raised-brow yes-no question learned it from somebody’s raised brows. The question of whose, and on what terms, does not go away because nobody was “identified.”

The warranty moves the consent problem

Both platforms put the legal weight of the upload on the person who uploads. Kling’s section 4.3 has the user “warrant and undertake” that they hold the intellectual property rights in the input or have the owner’s authorisation, and makes them liable for losses if they do not. Dreamina’s terms bar users from uploading content that violates anyone’s rights of publicity or privacy, which puts the same question on the same person^[2,4]. In the first Corpus Question, the consent gap sat inside vendor studios and scraped academic datasets. Here it moves to the individual creator, who is now both the signer and the party answering for every signer in the clip.

That matters most in the step good courses rightly insist on. To check whether an avatar signed correctly, a fluent reviewer compares it against a real signing source. The easiest real signing source to grab is somebody else’s video: a friend’s vlog, a Deaf influencer’s post, an interpreter’s clip from a news conference. If that clip goes into the upload box as a reference, the creator has just answered for rights they almost certainly do not hold, on behalf of a person who was never asked, under a licence that may be irrevocable. Signing communities are small. The person in that clip is often someone the creator knows by sight, and someone a model’s output may come to resemble.

We located no signed-language version of either platform’s terms or privacy policy. A Deaf creator who wants to know what she agreed to has to read several thousand words of contract English, which is the barrier these courses exist to lower. Consent that is only available in a language the signer reads less fluently than the one being captured is consent in name only, and no one in this chain is currently responsible for fixing that.

Accuracy literacy is not enough

The sign language AI conversation has spent two years arguing about output: whether avatars sign correctly, whether a hearing reviewer can tell, whether smooth motion fools buyers. That argument is necessary and it is far from won. But it has trained everyone, Deaf educators included, to look at the end of the pipe. Teaching a Deaf professional to reject a bad render while leaving the terms unread produces someone who can spot a broken avatar and still hand her own face to a training set without knowing it. AI literacy for signing communities has to include contract literacy, because the corpus question no longer belongs only to companies with studios and motion-capture rigs. It belongs to anyone with a phone and a free account.

The community would be buying back, through vendors and institutions, a capability it supplied for free.

NCG

There is a quieter commercial consequence too. Any sign language AI vendor that does not build its own model from scratch is building on, or fine-tuning from, models like these. If general video models get better at signing over the next two years, part of the reason will be the signing that Deaf creators uploaded while learning to use them, under licences that paid them nothing and that, in Dreamina’s case, cannot be withdrawn. The community would then be buying back, through vendors and institutions, a capability it supplied for free.

Before you upload signing

None of this means Deaf creators should stay away from AI video. It means the upload box deserves the same scrutiny as the render. Until platforms write terms that treat signed language as language, these are the practical steps.

1. Read the clauses first

   Before the first upload, read the licence and training clauses. For Kling that is terms section 4.7. For Dreamina it is terms section 8 and the privacy policy’s section on how information is used.



2. Upload only your own signing

   Do not use another signer’s video as a reference or a source, even for an accuracy check, unless that person has agreed in writing to it going into that specific tool.



3. Check accuracy beside the screen, not inside the tool

   Have the fluent reviewer compare the render against a source played on a separate screen, not one you upload.



4. On Kling, revoke in writing

   If you want your content out of further use, email support@kling.ai to revoke authorisation, keep a copy, and ask in writing whether revocation covers content already used for training.



5. On Dreamina, assume it is permanent

   Treat the licence on anything already uploaded as permanent. Deleting content or the account ends storage under the privacy policy; it does not undo an irrevocable licence.



6. Teachers: put the terms in the lesson

   Course designers should put a terms check in the same session as the accuracy check, and say out loud which clause governs the clip the class is about to upload.

What a buyer inherits now

The first Corpus Question ended with written questions for vendors. This piece adds the ones most procurement teams will not think to ask, because the risk no longer sits only in the vendor’s own corpus. An institution that would never accept a scraped dataset in a contract should not accept one laundered through a free account. The questions are cheap to ask in writing, and a vendor that cannot answer them has told you something about the rest of its documentation.

Questions people will ask