Five Cents a Sign: Inside GoSign.AI’s Data Supply Chain, the Company’s Answers, and the Terms That Haven’t Changed

12 October 2026 · Heather M. Grizzle, Novara Consulting Group · SLAT governance record, with GoSign.AI’s response

An evening at the Factory

It is a weeknight, and a fluent ASL signer has an hour before bed. She has cleared a wall behind her desk, angled a lamp toward her face the way the instructions ask, and opened Sign Language Factory on her laptop, because the platform does not run on phones^[1]. A word appears on the screen. Beside it a reference signer, recorded once and replayed for every contributor, demonstrates the sign. She watches, presses record, and copies it in a clip of one to four seconds, then submits. The next word appears. If she does not know a sign, the help page tells her to fingerspell it, with a reassurance that she will still earn five cents.

She is not a real person. She is a composite, built entirely from the instructions GoSign.AI publishes for its contributors, and she is here because the documents NCG reviewed describe a system in which a person like her is the first link in a long chain. What she records is reviewed by other signers, who are paid only when their vote agrees with the majority. Accepted clips join a dataset. Some are published to the open web with her face attached. Some are delivered to companies building sign language AI, and those companies build systems that a hospital, a transit authority or a state agency may one day buy to communicate with Deaf people. She agreed to all of it when she created her account, under a contract whose date changes every time she opens it.

This record is about that chain: what GoSign tells the people at its start, what it told NCG when asked, and what its own paperwork still says.

Who supplies the supply chain

GoSign.AI occupies a position on NCG’s roster that no other company does. Signapse, Kara and the others build or deploy systems that render or recognise signed language for an end user. GoSign supplies the training data such systems are built from. Its consumer app is a set of gesture-recognition games in which, as its store listing puts it, “every sign strengthens our models”^[2]. Its Sign Language Factory pays fluent signers to record prompted signs. Its enterprise offer is “world-class sign language datasets to train your computer vision models,” with custom collection and API access^[3]. A defect in what GoSign collects, or in how it obtained the right to collect it, does not stay at GoSign. It travels into every system trained on the data, including systems a public body may later procure.

That makes GoSign the commercial form of what NCG has called the Corpus Question: whether the people whose signing makes sign language AI possible consented to what is done with it, on terms they understood, for value that bears some relation to what it is worth^[4]. The academic corpora that dominate the field were largely scraped, carry no consent record and hold no contributor metadata. GoSign asks people to contribute and tells them why, which is a real improvement on that baseline. The company also sets its own bar. Its About page lists among its values that “transparent compensation, GDPR compliance, and responsible data practices are non-negotiable,” and calls the company “stewards of sign language data, prioritizing ethical collection and community interests above all”^[5]. This record reads GoSign’s documents against those words.

What she agreed to

The signer in the opening scene accepted a Contributor Services Agreement, version 1.0, by creating an account^[6]. Its counterparty is not GoSign.AI, LLC, the Rochester company named in GoSign’s main Terms under New York law, but “Sign Language Factory (‘Company’),” under Delaware law with Delaware arbitration. The agreement pays “$0.05 (five cents USD) for each video clip that is accepted,” fingerspelled clips included, with “no royalties, residuals, or ongoing payments.” In exchange, on submission of any clip, the contributor “irrevocably assigns, transfers, and conveys to Company all right, title, and interest in and to such content,” expressly including “all rights to use for AI/ML training and commercial applications,” and waives moral rights. Section 6 provides that verified clips “will be displayed publicly on the GoSign.AI website,” “may be shared on public platforms such as GitHub as part of an open ASL dataset,” “will be permanently accessible to the public,” and may be “downloaded, viewed, and used by third parties for research, AI training, and other purposes,” and that the contributor “has no expectation of anonymity.” Section 7 records consent to the collection of “biometric identifiers from submitted videos” and their retention “for 3+ years.” The help page confirms what follows from the assignment: if she deletes her account, her clips remain in the dataset.

Two details in that agreement matter more than their size suggests. The first is the date at the top. NCG examined the page’s published script on 27 September and again on 7 October and found that the “Effective Date” is produced by the reader’s device clock at the moment the page loads. A contributor who signed up in March and returns in October sees October. The agreement therefore has no effective date she can cite and no visible record of whether its terms have changed since she accepted them. The second is a contradiction with its companion document. The Screener Agreement tells the people who review her clips that “every contributor has a right to privacy,” to “treat all contributors as anonymous individuals,” and never to try to identify a contributor from their appearance^[7]. The contributor agreement tells her that her face will be published permanently and that she has no expectation of anonymity. Both documents are live on the same site. They also disagree on money: the screener terms state a $5 minimum withdrawal threshold, while the contributor terms state there is no minimum.

The signer in the opening scene is on the Factory. Anyone who plays GoSign’s consumer app is under the main Terms and Conditions, which are candid about the business: “We provide a gamified platform where users from all over the world can contribute signs and compete for giveaways. These signs are used for training AI sign language models”^[8]. Section 26 provides that GoSign “may retain video data submitted by users” for model training and “may also share or sell this data.” The licence over contributions is unlimited, irrevocable, perpetual, transferable and sublicensable, covers the user’s “image, name, and voice,” and is granted “for any purpose, commercial, advertising, or otherwise.” The privacy notice lists biometric data among the sensitive information collected, names “Sign Language AI Providers” among the recipients of shared data, and in its United States section states that “the categories of third parties to whom we sold personal information are: Sign Language AI Providers”^[9]. In its response GoSign told NCG that video shared or sold under the general Terms is de-identified, including removal of facial likeness. NCG could not locate de-identification or anonymisation language in either published copy of the Terms, and the privacy notice mentions anonymisation only for information GoSign no longer needs. NCG asked GoSign where that language appears and will cite it if the company identifies it.

Several documents, several ages

The first place most people meet GoSign is an app store, and the store listings do not agree with each other or with the company’s terms^[2,10]. Google Play rates the app 3+. Apple rates it 13+ and declares that it collects “body data (hands, head)” and video linked to the user. The Terms open by saying the Services are for users at least 16, and in Section 3 have users represent that they are not under 13. The Factory requires 18. SignLearns, launched on 7 October, says 13.

GoSign’s answer resolves part of this. Paid commercial data-collection projects require participants to be 18 or older as current practice, and the general Services are designed for users 13 and older with parental permission. It did not answer the question that matters most for a buyer: whether signing from users under 18, collected through the games under a Section 26 that permits sharing and sale, has ever entered a dataset that was shared or sold.

Seven questions, one answer

Around midday on 27 September NCG sent GoSign a pre-publication version of this record and seven written questions, covering the contracting entity and the floating date, how the five-cent rate was set and what purchasers pay, anonymity and withdrawal, sale and licensing, age, what “Deaf-led” means in decision terms, and dataset quality^[11]. Early that afternoon NCG also released an audio discussion of the working record, “Sign Language Biometrics for Five Cents,” on The Fifth Parameter^[12]. GoSign’s owner, Ryan Hait-Campbell, replied that evening with a written response and the company’s Data Collection FAQ, and said GoSign did not anticipate further substantive comment^[13,14]. NCG sent back its understanding of the answers on 28 September and gave the company until 17:00 Eastern on 9 October to correct any mischaracterisation. The response is reproduced in full at the end of this record.

The answers were substantive. GoSign confirmed that GoSign.AI, LLC is the legal entity operating Sign Language Factory, agreed the relationship should be clearer, and said it was updating its participant documentation. It agreed that contributors should be able to identify the version and date of the agreement they accepted. It said Deaf people hold decision-making authority across products, data programmes, partnerships, customers and compensation, that its owners and board are Deaf and use sign language, and that it has no outside institutional investment. It said datasets are delivered under participant terms and customer agreements that are generally confidential, that it does not sell contact details or account profiles as a standalone product, that QA varies by project with two independent screeners and a third to resolve disagreement in the standard workflow, and that as of September 2026 no commercial customer had approved a complete datasheet for public release. It said that once data is distributed it may not be able to guarantee withdrawal of copies already delivered.

Then it did something no other company on the roster has done. It put numbers on contributor pay. The five-cent structure, it said, was established for a specific NVIDIA project that concluded in July 2026 and is no longer in effect, and generally translated to approximately $15 to $25 an hour. A subsequent commercial project pays a $250 flat payment for successful completion of 1,000 accepted recordings, or $0.25 each, which GoSign estimates at approximately $40 to $50 an hour, and which it described as specific to that engagement rather than a standing rate.

We are sharing these figures voluntarily in the interest of transparency.

GoSign.AI, written response to NCG, 27 September 2026

That disclosure deserves credit, and it gets it here. It also raised a question the response did not settle: if the five-cent rate is over, why does the agreement say otherwise?

Ten days later

On 7 October NCG went back to the Factory as the signer in the opening scene would, through the home page that invites new contributors to sign up^[15]. The NVIDIA logo still sat under “Customers.” The Contributor Services Agreement linked from the footer was unchanged: $0.05 per accepted clip, “Sign Language Factory” as the Company, Delaware law, acceptance by creating an account. Its effective date read 7 October 2026, and the page script still produced that date from the reader’s clock. Its entire-agreement clause still states that it supersedes all prior agreements. The help page still promised five cents for a fingerspelled clip. The Screener Agreement still paid five cents per vote, and only for a vote on the winning side. No document NCG could find published the $0.25 rate or the $250 structure.

GoSign may be right about the terms that govern its current paid work. It told NCG that commercial projects can carry project-specific consent and terms, shown to participants before their data is included. But the only contract a member of the public can read is the one on the sign-up page, and that contract says five cents. A signer deciding tonight whether her time is worth it would reach the same conclusion NCG did.

The arithmetic of an evening

Go back to the signer at her desk, and give her GoSign’s own estimates. To earn $15 an hour at five cents a clip, she needs 300 accepted clips in that hour, one every 12 seconds. To earn $25, she needs 500, one every 7.2 seconds. Each of those cycles includes reading the word, watching the reference, recording a clip of up to four seconds and submitting it, and any clip the screeners reject earns her nothing. On the later project, $40 an hour at a quarter a recording means 160 accepted recordings an hour, one every 22.5 seconds, and $50 means 200, one every 18 seconds, which puts the 1,000 recordings needed for the $250 at five to six and a quarter hours of work.

The two estimates assume very different speeds for what looks like the same work. If the later project were done at the pace GoSign attributes to the NVIDIA project, a quarter a recording would come to $75 to $125 an hour. If the NVIDIA project was done at the pace GoSign attributes to the later one, five cents a clip came to $8 to $10 an hour. There may be an ordinary explanation: the help page now mentions sentence prompts with a longer recording limit alongside the one-to-four-second word clips, and sentences take longer. But both estimates cannot describe the same task at the same speed, and which one is closer to what signers were actually paid is precisely what the public record does not show.

For context, New York’s minimum wage in 2026 is $16.00 an hour outside New York City, Long Island and Westchester, which covers Rochester, where GoSign is based, and $17.00 within them^[16]. Factory contributors are engaged as independent contractors, so minimum wage law does not apply to them, and the comparison is offered to orient the reader, not as a legal finding. Two questions follow that GoSign has not answered. Is a contributor who completes 600 of the 1,000 recordings on the $250 project paid anything? And will GoSign publish what contributors on a completed project actually earned per hour, rather than what it estimates they earned?

Kto decyduje

GoSign describes itself as Deaf-led, and the claim is stronger than most on the roster. Its About page names its owner as a Deaf product designer and developer who uses sign language, and its response adds that its board and owners are Deaf and that there is no outside institutional investment. Ownership is decision authority, and NCG reads the response as meaning that the Deaf owners and board can halt a dataset release or a sale; GoSign was invited to correct that reading.

What the record does not show is any structure that reaches the signer at the start of the chain. There is no advisory body, charter or documented route through which contributors can influence what is collected, how it is licensed or to whom it is sold. That is a different gap from the one NCG found at Signapse, where structure exists on paper and authority sits with hearing executives, or at Kara, where participation is documented and authority is not. At GoSign the owner is Deaf and holds the authority. The contributing community has a rate card and a consent form, not a seat.

What GoSign does well

This section is placed where a reader will find it. GoSign discloses the purpose of collection at the point of download, before anyone installs its app. It publishes a contributor agreement with a stated rate, stated payout terms and a stated tax treatment. It obtains express biometric consent rather than burying it. It runs a screening process and explains its rejection criteria. It tells contributors in plain words that their clips survive account deletion, where most companies would leave that to be discovered. It has taken a public position against dataset monopolisation and committed to an open ASL dataset. It answered NCG’s questions within hours and volunteered compensation figures. On the roster NCG has reviewed, no other company publishes a comparable account of its data supply chain. The criticisms in this record exist because GoSign made its chain visible enough to examine.

A cleaner set of terms, launched the same day

On 7 October, the day NCG re-checked the Factory, GoSign and Sign-Speak announced SignLearns, a beta ASL practice platform powered by Sign-Speak’s models, and invited Deaf educators, teachers, students, schools and universities to test it^[17]. Its Terms and Privacy Policy do several things the Factory’s documents do not^[18,19]. They name GoSign.AI, LLC as the operator. Their effective date is written into the page, which NCG confirmed in the script, rather than generated from the reader’s clock. They state that practice videos are not saved by SignLearns and are not used by SignLearns to train AI models. They set one minimum age. These are the corrections NCG’s record asked for at the Factory, made in a new product, and they show the company can draft clear terms when it chooses to.

One question carries over. The Privacy Policy says practice videos are sent to Sign-Speak for recognition, and that providers process data under their own contracts and policies. GoSign’s promise not to keep or train on the video is a promise about GoSign. Neither document says whether Sign-Speak keeps the video or uses it to improve its models. Beta testers, who may be as young as 13, are signing into a camera for a product that lists a planned price of $20 a month, and they are entitled to know where their signing goes once it leaves the browser. Their feedback, under the Terms, may be used without restriction or payment.

O co powinien zapytać nabywca

For an AI company buying GoSign data, the questions are the licence, the consent chain, the label accuracy, the reference-variety design, and whether the biometric consent obtained from contributors reaches the purchaser’s use. Every clip in a Factory dataset is a reproduction of one reference signer’s variant, collected one variation at a time by design. That is a defensible choice for a recognition corpus, and also a linguistic fact about the data that a purchaser should know before claiming the resulting system understands how Deaf people actually sign.

For a public body buying a finished system from a company that trained on GoSign data, or on anyone’s, the question is simpler, and NCG’s own instrument does not yet ask it: which datasets was this system trained on, and can the vendor produce the supplier’s datasheet and consent terms?

What remains open

The signer in the opening scene could read every document cited in this record and still not know what she is being paid, which contract she signed, or where her face has gone. GoSign has answered more than most companies would, and it has disclosed figures no competitor has. What it has not yet done is make the documents at the start of its chain match what it says about them. NCG will note any update to the published agreement when it appears.

GoSign’s response in full

GoSign’s written response of 27 September 2026 is reproduced below as received, without the email signature block. The Data Collection FAQ that accompanied it is published with this record as received: download GoSign.AI’s Data Collection FAQ (PDF, 10 pages, 145 KB).