How AI Training Jobs Became Precarious Gig Work

As generative AI matures, a hidden labor market has swelled: professionals across law, writing, design and science are being hired to create and grade the data that trains today’s models. Companies such as Mercor, Scale AI and Surge AI recruit thousands of contractors to write prompts, craft ideal responses, build evaluation rubrics and produce voice or visual recordings — often under strict nondisclosure rules and heavy monitoring.

Workers describe work that can pay well at first but is also intermittent, opaque and highly controlled. Projects are started and paused as client needs change, leaving freelancers waiting for new batches of tasks. Time-tracking software records activity to the second, and many said targets tightened while pay fell. Some reported working off the clock or disabling monitoring tools to keep up.

Those who once had steady careers — journalists, screenwriters, academics — find themselves competing for short-term gigs that, paradoxically, teach machines to replicate their own expertise. The jobs range from writing exhaustive rubrics defining what counts as a ‘good’ response, to generating ‘golden outputs’ and detailed reasoning traces for models to emulate.

The arrangement has raised legal and ethical questions. Several class-action suits allege misclassification of contractors and excessive control. Critics warn the model risks creating a precarious workforce that lacks benefits and collective bargaining power, while companies say project-based work necessarily fluctuates with client requirements.

For many, the emotional cost is significant: unpredictable schedules, surveillance, and the knowledge that their labor accelerates automation of their own fields. Some workers who started with hope and decent pay now say they prefer predictable, lower-paid roles with clearer protections.