Frontier AI systems are trained on data that people are paid to produce, annotate, or rate. The category spans several distinct arrangements: contractors labeling images and text for supervised training, raters producing the preference comparisons behind reinforcement learning from human feedback, content moderators reviewing harmful material, and, more recently, members of the public recording video of themselves performing physical tasks so that robots can be trained on them. The work is generally outsourced, generally paid per unit or per hour rather than salaried, and generally performed outside the countries where the models are built.
The labor is a governance subject on three grounds that are usually treated separately: the working conditions and pay of the people doing it, the effect of who they are on what the model learns, and the integrity problems that arise when the same payment structure that scales collection also rewards submitting data that only looks acceptable.
The contractor model
The dominant arrangement for language-model data has been a specialist vendor holding a global contractor workforce and selling annotation capacity to frontier labs. Scale AI is the largest example: data labeling and human feedback for frontier labs is its historical core business, delivered through a contractor workforce operated under the Remotasks and Outlier brands, alongside its later evaluation, enterprise-platform, and federal-government lines.
The model has produced sustained disputes. Scale has faced criticism over Remotasks and Outlier contractor pay, working conditions, and the psychological effect of content-moderation labeling, with class-action suits filed in California and in the Philippines. The complaints turn on the structural features of the arrangement — payment per completed unit, classification as contractors rather than employees, and exposure to harmful material as a condition of the work — rather than on any single employer's conduct.
A second structural question concerns the vendor's position between competitors. Recurring questions have been raised about whether training labels a vendor produces for one frontier lab remain adequately siloed from its work for others, and Meta's acquisition of a stake in Scale was treated by frontier-lab customers as a conflict-of-interest event that reshaped the company's revenue mix toward government work.
Labeler composition and model behavior
Who does the rating affects what the model becomes. In discussions of political bias in language models, the distribution of human labelers — commonly described as skewing academic or concentrated in particular regions — is one of the mechanisms through which RLHF is understood to introduce slant, alongside the instructions labelers are given and the composition of the pretraining corpus. The claim is contested in degree rather than in kind: that the raters' judgments propagate into the reward model is not disputed; how much of an observed slant is attributable to them, rather than to the underlying corpus or to explicit policy, is.
A related limitation is raised against the technique itself. Critics including Aschenbrenner argue that RLHF cannot scale to supervising systems substantially more capable than the human labelers rating them — an argument about the ceiling of human-generated supervision rather than about the labor arrangement.
Physical-data collection
Robotics has produced a different arrangement, driven by the claim that the data required does not exist to be scraped. Figure AI announced Index in August 2026, a crowdsourced consumer app through which members of the public record video of everyday tasks — household chores, restaurant and warehouse work, retail stocking — and are paid for accepted submissions, whom the company calls Creators (Introducing Index: Building The World's Largest and Most Diverse Physical Dataset (Figure AI, August 2026)). Figure reported 264,000 downloads across 108 countries, over 44,000 weekly active users, 16 million uploaded videos, an ingest rate of 30 minutes of video per second, and $15 million paid to contributors, against a stated forward commitment of over $1 billion on data and compute over twelve months. It states it first tried buying such data and found vendors could not meet its throughput, diversity, or quality requirements.
The arrangement runs in two directions: a contributor can record their own tasks, or a customer can book a Creator through the app to come to their home or business and perform tasks while recording. Figure gives a cumulative payout figure but publishes no rate, acceptance rate, or terms.
Skild AI sets out the opposite approach for the same problem, scaling four collection modes in house — robot teleoperation, hand-held collection, egocentric video, and simulation — traded off against hardware proximity, diversity, and scalability (Introducing S1: In-Context Learning for Robotics). Both are vendor accounts of their own methods; no published result compares the resulting datasets.
Quality control as a cost centre
Paying per submission creates an incentive to submit, and both robotics accounts treat screening as a first-order expense rather than an afterthought. Skild states that for every dollar it spends collecting data it spends three on quality control, screening each data point for low-level precision, task coherence, and annotation fidelity. Figure's five-stage pipeline — automated filtering, human fraud review, embedding-based deduplication, rebalancing by task quota and cluster, and hierarchical captioning — includes a dedicated stage in which human analysts audit samples at the user level specifically for "deliberate attempts to evade" the automated filters.
The deduplication stage is a second consequence of paying per unit: a contributor rewarded for volume has no reason to vary what they record, so Figure embeds each segment and discards those above a similarity threshold with data already accepted. The diversity Figure attributes to its contributor base is therefore a property the pipeline enforces as much as one the crowd supplies.
Open questions
- Whether the crowdsourced physical-data model attracts the same labor-classification and pay disputes as the contractor annotation model, and under which jurisdictions' rules.
- What rate per accepted submission the physical-data programmes pay, and what fraction of submissions is accepted; no developer has published either.
- Whether any published capability result isolates the effect of labeler or contributor composition on model behavior, as distinct from the effect of corpus composition.
- Whether vendor-held annotation workforces serving competing labs create the training-data leakage their customers have raised.
Relationships
- related: Scale AI — the largest annotation vendor and the subject of the labor disputes.
- related: Figure AI, Skild AI — the two published physical-data-collection strategies.
- related: Introducing Index: Building The World's Largest and Most Diverse Physical Dataset (Figure AI, August 2026), Introducing S1: In-Context Learning for Robotics — the primary accounts.
- depends-on: RLHF (Reinforcement Learning from Human Feedback) — human preference rating is the labor input the technique is built on.
- related: AI Political Bias — labeler composition is one proposed mechanism for model slant.
- related: Training Data Walls — the constraint that motivates paying for data rather than scraping it.
- related: AI Robotics, AI Labor Disruption.