The Rate Came Out, the Contract Went In: GoSign’s Sign Language Factory Becomes a Research Study

A video and a link

The video runs a few seconds. A man with a red beard and clear-framed glasses leans into his phone camera from a living room, ceiling fan turning behind him, signing with the energy of someone announcing a winning ticket. A caption reads “Let’s go!!!!” Across the top of the frame, in a yellow banner, is an instruction: Sign up, then an arrow, then tasks.gosign.ai. The post beneath it says that sign language recognition is needed on every device.

NCG saw the video on 10 November 2026, a month after closing its governance record on GoSign.AI’s data supply chain^[4]. That record was about the people at the start of the chain: fluent signers who recorded short clips for GoSign’s Sign Language Factory, under an agreement that paid five cents a clip, assigned every right in the signing to the company, and dated itself from the reader’s own device clock. The video points at the same address. What a new signer finds there is not the same platform.

This piece follows the link. It sets the relaunched Factory beside the record NCG published, credits what has been fixed, and asks what the new arrangement asks of the people it recruits. It is written from the platform’s public pages as they stood on 10 November 2026. NCG has not seen the consent form now at the center of the arrangement, and makes no statement about its contents.

Back at the desk

NCG’s October record opened with a composite: a fluent signer at her desk on a weeknight, built entirely from GoSign’s own instructions, copying a reference sign into her laptop camera for five cents and fingerspelling the words she did not know because the help page promised she would still be paid^[4]. Send her back to the address in the video and her evening starts differently.

The landing page no longer calls the Factory a place to earn money by recording clips. It calls it a paid research study^[1]. She cannot use her laptop’s ordinary browser. Sign-up works only in Safari, on an iPhone, iPad or Mac, and the page warns that Chrome, Firefox, Edge and the browsers inside Instagram, Facebook and Gmail will not work. If she follows the link from the video inside a social media app, the page tells her to copy it and open it somewhere else.

Before she records a single study clip, she applies. The application asks for her name, email and state of residence, then an eligibility survey covering her date of birth, gender, hearing status, the sign languages she uses and how fluently, her race or ethnicity, her skin tone on the Monk scale, her dominant signing hand, the Apple device and operating system she will record on, and her recording space and availability. She uploads a screenshot of her device’s About screen, a document proving where she lives, and a short intro video of herself signing, which the notice says shows her face, hands and surroundings and may capture her voice^[2]. Then she waits to be selected.

If she is selected, she receives an Informed Consent Form. If she signs it, she records about 1,000 short videos at home, each around ten seconds, made of everyday phrases, assistant-style commands and fingerspelling, and she is paid $0.25 for every recording that is accepted, up to $250^[1]. The help page still tells her to fingerspell a sign she does not know. It now says she will still earn a quarter^[3].

What changed, and what it fixes

The relaunch answers a good part of the October record, and it should be read that way first. NCG asked GoSign to publish the rate signers are paid; the rate is now on the landing page, the help page and in the study description, and it matches the $0.25 GoSign gave NCG in its written response in September^[4]. NCG documented six separate statements of minimum age across GoSign’s documents, ranging from a 3+ store rating to 18 on the Factory; the study now states one age, 18, and says GoSign does not knowingly collect information from anyone younger. NCG reported that the contributor agreement named an entity called “Sign Language Factory” under Delaware law while GoSign’s main terms named GoSign.AI, LLC under New York law; the new notice names GoSign.AI, LLC, gives a Rochester address and names a privacy contact. NCG found that the agreement’s effective date was generated by the reader’s device clock; the new notice carries a fixed date, 9 October 2026, and records the version of each statement an applicant agrees to.

Other improvements were not on NCG’s list. The study recruits signers of Black ASL alongside ASL, which matters in a field whose training data has rarely reflected the variety of American signing. It states that recording begins only after consent is signed. It says application materials, including the intro video, are not published or used to train AI models, and that applications not selected are deleted within a year after the study ends, or sooner on request. And the open question NCG left at the end of its arithmetic, whether a signer who stops at 600 of the 1,000 recordings is paid anything, now has a public answer: payment is per accepted recording and begins at 20.

The last three rows are the rest of the story.

The rate, finally in public

Five cents was the number that gave NCG’s October record its title, and its disappearance from the site in mid-October without a replacement was the development NCG flagged most sharply. A published quarter is a better place to be. It lets a signer decide whether the work is worth her evening before she starts, which is the minimum a fair exchange requires.

What a quarter is worth depends on how long each accepted recording takes, and the public record still does not say. In September GoSign estimated that signers on its $0.25 project earned $40 to $50 an hour, which implies one accepted recording every 18 to 22.5 seconds^[4]. The relaunched landing page describes each recording as around ten seconds long. At that length, 1,000 recordings is about 2.8 hours of time on camera before anything else is counted: reading the prompt, setting up, retaking, and every recording that screeners reject, which earns nothing.

At 30 seconds per accepted recording, the study pays about $30 an hour and the full 1,000 takes a little over eight hours. At a minute each, it pays $15 an hour across nearly seventeen hours. For orientation, New York’s 2026 minimum wage outside New York City, Long Island and Westchester, which covers Rochester, is $16.00 an hour^[6]. On these figures the rate is no longer the scandal it was at five cents. Depending on pace it ranges from roughly the minimum wage to three times it, and NCG’s October criticism of the rate should be read against that change.

Three things the hourly view does not capture deserve a sentence each. The $250 is a ceiling, so no pace earns more than that for a participant’s entire contribution of identity-bearing video. The application itself is unpaid labor that includes recording an intro video and uploading proof of residence, and an applicant who is not selected receives nothing for it. And no figure here is a measured earning. GoSign has not published what participants on a completed project actually earned per hour, which is the number that would settle the question.

The contract went where the rate used to be

In October the Factory’s terms were public and its rate was confused. In November its rate is public and its terms are not.

The Contributor Services Agreement that NCG read on 27 September and again on 7 October set out what a signer gave up for five cents: an irrevocable assignment of every right in her clips, permission to publish them permanently with her face attached, consent to biometric collection, and retention for three or more years^[4]. A person could read all of it before creating an account. On 17 October NCG found that agreement behind a login wall. On 10 November its address, and the address of the Factory’s former terms page, both lead to the Applicant Privacy Notice.

That notice is candid about where the real terms now live. It says the Informed Consent Form “is your key agreement.” It says the form explains what participants do, what study data is collected, how it is used, shared and kept, payment, confidentiality obligations, and “the rights you have over your Study data.” It says that once the form is signed, the form governs, and that if the notice and the form differ, the form controls^[2]. The help page gives the same answer to the question of how recordings are used: read the consent form when you get it^[3].

A signer can now learn what she will be paid before she applies. She cannot learn what she is giving up until after she has handed over her face.

NCG

An informed consent form shown before recording is a real protection, and it is more than many data collectors offer. The difficulty is sequence. Everything a signer needs to decide whether to apply at all, including whether her recordings will be published, sold onward, used to build commercial products, retained for years or withdrawable, sits in a document she cannot see until she has already supplied her race, her skin tone, her hearing status, proof of her address and a video of her signing face. The notice limits what is done with those application materials, and that limit is worth noting. It does not tell her what she will be asked to sign next.

For a Deaf signer there is a second layer. Whether the consent form is available in ASL, whether a participant can ask questions about it in her own language before signing, and whether the version she signs is one she can keep and cite later are questions the public pages do not answer. The October record found that the Factory’s agreement could not be cited by date at all. The fix for that problem now sits inside the document that cannot be read.

A study with no names on it

The notice describes GoSign as running “the application and recording platform for the Study on behalf of the organization conducting the Study and the Study sponsor”^[2]. Neither is named anywhere on the public site. The October record listed NVIDIA under the Factory’s customers, and GoSign told NCG that the five-cent rate belonged to an NVIDIA project that ended in July 2026^[4]. Who the current study is for, and who conducts it, is not public.

The words “research study” and “informed consent” carry expectations. In human subjects research, an informed consent form is usually reviewed by an institutional review board, and participants are told whom to contact about their rights. The public pages do not name a review board, a principal investigator or an institution, and NCG makes no claim either way about whether one exists. They also do not say whether the study’s output is research in the academic sense, a commercial dataset, or both.

There are ordinary reasons a sponsor might stay anonymous during recruitment, from commercial confidentiality to keeping applicants from tailoring their answers. Those reasons explain the practice. They do not change its effect. A signer deciding whether to contribute her face and her language to a system cannot weigh who will use it, for what, or under whose accountability, and a public body that one day buys a product trained on this data will find no public record connecting the two.

Four states off the map

Section 8 of the notice says the study is open only to adults who live in and are located in the United States, and then adds a line: “It is not open to people located in Colorado, Illinois, Texas or Washington”^[2].

Those four states share something. Illinois, Texas and Washington are the three states with biometric-specific privacy statutes, and Colorado added biometric identifier requirements to its privacy law effective 1 July 2025^[7,8,9,10]. Illinois’s law allows individuals to sue directly; Colorado’s amendment requires consent before biometric data is collected and restricts its sale. A study that records faces and hands, collects race and skin tone, and keeps a device fingerprint on every applicant is the kind of collection these laws were written for.

NCG does not know why GoSign or the sponsor excluded these states, and the notice gives no reason. The pattern is a fact; the explanation is GoSign’s to give. What a reader can say is that signers in four states are protected from this study by their state legislatures, and signers everywhere else are protected by a consent form none of them can read in advance.

Who gets to talk

Section 9 of the notice is headed “Please keep study details private.” It says details learned while applying or onboarding are confidential, asks applicants not to share them “or post about the Study on social media,” and says a selected participant’s full confidentiality obligations are in the consent form^[2].

The study, meanwhile, is being recruited on social media, in the kind of video NCG saw on 10 November: upbeat, signed, aimed at the community whose signing the study needs, with the sign-up address across the top. Promotion of a study is not a breach of anything, and confidentiality during recruitment is common. The asymmetry is still the point. The study’s promoters can say as much as they like about why signers should join. The signers who join cannot say what they found when they did, and the people deciding whether to apply cannot hear from those who went before them.

That matters more in a small community than in a large one. Deaf people decide whether to trust a technology project largely through each other, by asking who has tried it and what happened. A confidentiality request that closes that channel removes the one form of due diligence most signers actually have.

What GoSign has done well

This section sits where a reader will find it, as it did in October. GoSign answered NCG’s September questions within hours and volunteered compensation figures no competitor has disclosed. Within weeks of NCG’s record it published the rate it had described privately, settled its age requirement, named its legal entity, fixed its document dates, put consent before recording and opened recruitment to Black ASL signers. Those are the corrections a governance record asks for, made in public, and they should be credited as such. NCG’s October record described GoSign’s supply chain as the most transparent on its roster^[5]. On rate and age, it is more transparent now.

The criticisms in this piece are of a different kind. They are not about what the documents say. They are about who is allowed to see them, who is named in them, and who may speak about them afterward.

What a signer, and a buyer, should ask

A signer considering the study can ask GoSign for the consent form before applying, ask whether it is available in ASL, and ask for a copy she can keep. A buyer, whether an AI company acquiring the resulting data or a public body purchasing a system trained on it, needs the same documents from the other end of the chain.

What remains open

GoSign has fixed the parts of its Factory that NCG could see. The parts that matter most to a signer, what she gives up and to whom, are now in a document the public cannot see, for a sponsor the public cannot name, under a request that participants not discuss it. Those choices may each have a reason. Together they move the Factory from a platform whose problems were visible to one whose terms are not.

NCG invites GoSign.AI to respond to this piece, and in particular to say who sponsors and conducts the study, whether a review board oversees it, whether the consent form will be published in advance, and why four states are excluded. Any response will be published with this piece in full.