The people who label training data, rate model outputs and review harmful content so that systems behave acceptably. Much of this work is outsourced, low-paid and psychologically demanding, particularly content moderation, and it is largely invisible in how these systems are described.
Technical view
The labour is structurally necessary: preference data, safety classifiers and evaluation all depend on human judgement, and the quality of that judgement bounds model behaviour. Documented concerns include pay, precarity, lack of psychological support for exposure to harmful material, and opacity of supply chains. Practical responses include supply chain disclosure, wage floors and welfare provisions in procurement, and shifting the most harmful review categories to automated pre-filtering.
Related
Part of AI Terminology Index, section: Society and Impact.