Consent, Privacy and Pay: How AI Data Contributors Should Be Treated

Surveys of major crowdwork platforms place average earnings between $1 and $5.50 an hour, with a median around $2, and only about 4% of workers clearing the US minimum wage of $7.25. For every paid hour, workers spend roughly 18 more minutes on unpaid labour — searching for tasks, qualifying, disputing rejections — and accounting for that invisible work drops measured median wages further still. SOMO's 2026 investigation traced at least 30 intermediary companies through which the largest technology firms source data work, several accused of paying below minimum wage, dismissing workers unfairly, blocking collective organising and providing no social protections, while pricing pressure from the top of the chain sets the conditions below. Workers are responding through groups such as Kenya's Data Labelers Association, the Data Workers Inquiry and Turkopticon, organising that Brookings has documented alongside the retaliation some of it meets.

The research community's own record is telling: a systematic review of machine-learning papers using crowdworkers found zero that reported what the workers were paid. The workforce that produces the ground truth of modern AI is, in most of the literature, not even a line item — the baseline against which any claim to responsible AI data work should be tested.

• US federal minimum wage (reference line): $7.25/hour.
• Typical surveyed crowdwork range: $1–$5.50/hour, with a median near $2/hour.
• Share of surveyed workers earning above the US minimum: roughly 4%.
• Unpaid work per paid hour, before invisible labour is counted: about 18 minutes.
• Intermediary companies through which large tech firms source data work, per SOMO's 2026 investigation: 30+.
• ML-research papers reviewed that reported crowdworker compensation: zero.

The ethics literature grounds this properly: human computation tasks should meet the same standards as behavioural-science research on human subjects, anchored in the Belmont principles — the research-ethics framework built on respect for persons, beneficence and justice. Translated into operations, respect for persons means consent that is genuinely informed: before contributing, a person knows what is being collected (their voice, their judgments, their demographic details), what it will be used for, who receives it, how long it is kept, and what they are paid, in their own language and at reading level, not in a click-through. Specific means new uses need new consent — a voice recorded for speech recognition is not thereby licensed for voice cloning. Documented means the consent record travels with the dataset as part of its provenance, a practice worth building into multilingual data collection programs as part of the deliverable rather than an afterthought. Revocable means a working withdrawal path, with its practical limits — what has already shipped, what can still be excluded — stated honestly upfront rather than discovered later.

Consent also covers the work itself, not just the data. Contributors are told the nature of content before they opt in, which matters enormously for tasks involving disturbing material, and consent to sensitive work is opt-in with support, never a condition of employment. The Belmont framing's third principle, justice, points at the same place the next two sections do: the people bearing the work should share fairly in its benefits.

The operational rule is separation: identity and verification records live apart from the work product, and what ships to a client is the contribution plus the metadata the dataset legitimately claims — age band, region, dialect — never identity records. Demographic information is collected under consent, used for the balance reporting it exists for, and minimised everywhere else; frameworks like CrowdWorkSheets exist precisely because who annotated a dataset shapes it, and documenting that responsibly means aggregates and safeguards, not exposure. Speech is personal data everywhere and treated as biometric-grade data — information sensitive enough to require explicit, use-specific consent — in several jurisdictions, so voice datasets carry the strictest tier: consent naming the uses, no repurposing without re-consent, and secure handling through the pipeline.

Where contributors process other people's data — the PII inside documents, recordings and screenshots that shows up across document and OCR annotation work — protection points both ways: de-identification workflows protect the subjects in the data, while access controls, clean-room environments and confidentiality training protect contributors from carrying risk they never chose. Privacy in data work is one discipline with three beneficiaries: the client, the data subject and the contributor.

Oxford Internet Institute research shows clearer guidance and better pay directly improve annotation accuracy; industry analyses document that stable roles let contributors build the task-specific skill complex guidelines demand, and that fair treatment cuts the turnover that quietly injects errors as replacements relearn every edge case; Fairwork's reporting shows the same principles applied in practice. Even research teams publishing datasets now advertise their labour standards — one alignment dataset documents full-time annotators paid roughly $8–$9 an hour against a local minimum near $3.69, on regulated eight-hour days, because reviewers have started to ask. The practitioner QA literature adds a blunt line to the same ledger: paying per task instead of per hour is listed among the practices that fail.

The delivery-centre model — an employment-shaped operating structure, as opposed to piece-rated crowdwork platforms — is one answer to that evidence: contributors work through 40+ delivery centres across 30+ countries, trained and hourly-oriented rather than piece-rated, with local labour law as the floor rather than the ceiling, and progression tied to the quality record. The same core standards should hold regardless of location, because a value that varies by geography is a policy, not a value — the same logic behind treating annotator recruitment and certification as a real training pipeline rather than a sign-up form, including for contributors recruited into African-language data programs. It is also self-interested in the most defensible way: the 56,000+ registered-contributor network a vendor can draw on only exists because people stay, and people stay because the work is worth staying for — the same reasoning that runs through running a global data operation across multiple delivery centres and routing sensitive annotation work to trained, vetted teams rather than open crowds.

A caution on the numbers: wage surveys describe specific platforms and periods, investigation findings are as published by the cited organisations, and figures a vendor reports about its own operations (delivery-centre count, contributor count) should be labelled company-reported, at the level the vendor itself publishes them. Verify figures at the original sources, and treat none of this as legal advice. Buyers comparing human-in-the-loop annotation providers should expect a straight answer to how contributors are engaged, since it is a leading indicator of the quality a vendor can actually hold.