Consent & fair pay
Informed consent, at scale, at low literacy
Consent that a court would recognise, taken from people with widely varying literacy, in six languages. This is the operationally hardest thing we do, and the thing we are least willing to cut corners on.
The standard
What every contributor is told
Not a translated legal document. A plain-language explanation, tested with the population it is written for.
- Written in the contributor's own language at a reading level tested with the target population
- Read aloud and recorded where literacy is a barrier, with an independent witness present
- States explicitly that data will be used to train machine-learning models, including commercial and derivative models
- Names the retention period, the geographies of storage and the categories of recipient
- Heightened, separate consent for biometric and facial data, naming biometric processing directly
- Guardian consent plus child assent for contributors under 18, with additional safeguarding review
- A published withdrawal channel in-language, with a stated and honest limit on what withdrawal can undo
- Versioned: every record links to the exact consent text and version that governed it
Fair pay
Underpaid contributors produce bad data
This is an ethics commitment and a quality argument at the same time. Rushed, resentful, financially desperate work is low-variance, high-fraud and expensive to QA. Paying properly is the cheapest quality intervention available.
We disclose the rate before work begins, pay within 14 days, and support payment channels that work for people without bank accounts. What we pay is disclosed to clients under NDA — we think buyers should be asking every vendor for it.
At or above living wage
Benchmarked to Global Living Wage Coalition figures or the nearest regional equivalent, never below statutory minimum.
Disclosed before work
Rate, unit definition and expected time are stated in writing, in-language, before a contributor accepts.
Welfare for safety work
Capped exposure hours, briefing and debriefing, paid rest, access to support and unconditional opt-out for red-team and toxicity tasks.
The part most vendors leave out
Fraud is a constant pressure in this industry: one person operating five accounts, recycled audio, model-generated text submitted as human writing. We assume it and design against it — ID verification, device and location fingerprinting, duplicate fingerprinting across the whole corpus, and payment tied to QA-passed units.
Contributors caught committing fraud are removed and their entire history is re-collected at our cost, not yours.
Answers
Consent and pay questions
At or above the local living wage as benchmarked by the Global Living Wage Coalition or the closest available regional equivalent, never below Pakistan's statutory minimum, and disclosed in writing to the contributor before they begin. Payment is per QA-passed unit or hourly depending on task, made within 14 days, through JazzCash, EasyPaisa, Raast or bank transfer. We disclose what we pay, by role and province, to clients under NDA — buyers increasingly ask, and they should.
They can withdraw at any time through a channel published in their own language. On withdrawal we remove their records from our active corpus and stop further licensing of them. We are honest about the limit here: we cannot un-train a model that has already ingested the data, and we say that explicitly in the consent document rather than implying an impossible promise. Downstream clients are notified of withdrawals affecting delivered datasets so they can act on their own policy.
The consent is read aloud in the contributor's own language by a member of staff who is not the person recruiting them, with a witness present, and the reading itself is recorded. Comprehension is checked with three plain questions before a thumbprint or signature is taken. It is slower and more expensive than a checkbox, and it is the only version that would survive scrutiny.
Only where a client's spec genuinely requires child speech, and only with guardian consent plus the child's own assent, additional safeguarding review, a chaperone present throughout, and a shorter session cap. Many projects do not need it and we will say so.
Ask to see a consent form.
We will send the actual template, in Urdu and English, with the reading-level notes. Most vendors will not do this.
Average first response: under 6 business hours. NDAs signed same day.