What the AI Boom Owes to the People Labelling Its Data

The apparent intelligence of a modern model is partly the accumulated judgement of a very large number of people who rated, sorted, transcribed, corrected and compared text, images and audio according to instructions written by someone else.

That work is the least discussed part of the supply chain and one of the most determinative.

Raters decide what counts as helpful and true

When a model is trained to prefer one response over another, the preference comes from human raters. Their instructions encode assumptions about what is helpful, polite, safe and true. Those assumptions are not neutral, and the people applying them are often working quickly, for low pay, on tasks with no context.

• Rater demographics shape model behaviour in ways that are rarely disclosed.
• Guidelines written for consistency often flatten cultural nuance.
• Quality depends on pay and working conditions, which are frequently poor.

Nobody discloses who labels the data

Most companies disclose little about who labels their data, where, under what terms and with what support. Concerns about psychological harm from reviewing disturbing content have been raised by researchers and, in some jurisdictions, by regulators.

The asymmetry is stark. The output of this labour is the core asset of the most valuable companies in the world, and the labour itself is contracted through intermediaries with limited transparency.

A reporting duty that would cost little

Requiring companies to report who performs data work, in which jurisdictions, and under what conditions would cost little and change behaviour meaningfully. It would also give buyers — enterprises with procurement standards of their own — a lever that currently does not exist.

Hardware supply chains get audited; data work does not

Companies are now scrutinised for labour practices in their hardware supply chains, and reporting requirements exist in several jurisdictions. Data work has no equivalent, despite being more directly connected to the product than the mining of any component.

The reason is partly that data work is easier to contract out through layers of intermediaries, and partly that the physical distance makes it less visible than a factory. Neither is a principled distinction.

Tired workers annotate differently than rested ones

The commercial case for better conditions is that quality depends on the judgement of the people applying the guidelines. Workers operating under time pressure with inadequate pay make different decisions than workers with time to think, and the difference propagates into the model.

Several labs have said as much publicly while contracting through vendors that compete primarily on price. The gap between the stated standard and the procurement decision is the place where the issue actually gets resolved.