Support Personnel

In the AI Story

The global data annotation workforce numbers in the hundreds of thousands to low millions, concentrated in Kenya, India, the Philippines, Venezuela — countries where English-language literacy is high and wages are low. The work pays a few dollars per hour: enough to attract workers locally, far below what the same work would command where AI companies are headquartered. Their contribution is essential — the model cannot be trained without labeled data — but the conventions of the AI world assign them no credit, no visibility, and no voice.

Content moderators occupy a similar position. They review and filter outputs, flag harmful content, red-team systems for safety. The emotional and psychological toll has been documented: exposure to violent, disturbing, and traumatic content as routine work. Their labor keeps AI outputs within the bounds users expect, and the conventions acknowledge their contribution no more than the concert world acknowledges the custodian's mop.

The open-source community represents a different kind of support personnel. Developers contributed code under licenses permitting sharing and modification. That code was subsequently ingested by AI training processes. The contributions are essential to model capabilities, but the developers did not consent to this specific use and receive no compensation. This is a convention gap — a situation where existing shared understandings do not provide clear guidance for new circumstances.

Becker would not have called this exploitation exactly — his vocabulary was deliberately less loaded. He would have called it a convention of credit and compensation that distributes cooperative returns in a way that reflects power relations among participants rather than the value of their contributions. The convention persists because those with power to change it have no incentive, and those who would benefit from change have no leverage.

Origin

Becker developed the concept through ethnographic observation of orchestras, film productions, theaters, and other cooperative creative enterprises. The concept was formalized in Art Worlds (1982), Chapter 4, which analyzed how different art worlds handle the support function — some inclusively (film credits), some exclusively (the novelist's byline).

Mary Gray and Siddharth Suri's Ghost Work (2019) documented the vast distributed workforce performing micro-tasks on which AI systems depend, providing the empirical foundation for applying Becker's framework to the AI world.

Key Ideas

Invisibility is a convention, not a fact. Support personnel are invisible because art world conventions render them so. Different conventions would render them differently visible.

The AI world's support layer is geographically distributed and contractually obscured. Global labor arbitrage routes annotation work to wherever workers will accept the lowest rate, maintaining invisibility through distance.

The convention serves power. It persists because changing it would require those with power to accept reduced returns, and those who would benefit lack leverage to force the change.

Convention gaps arise when practices outpace shared understandings. Open-source code ingested into AI training is a new use that existing licenses did not anticipate.

The scale is different from traditional art worlds. The global support network for AI dwarfs the cooperative networks of any previous creative domain, making the invisibility more consequential.

Debates & Critiques

Defenders of current arrangements argue that annotation workers freely accept the wages offered — that labor markets reveal preferences and contracts reflect agreement. Becker's framework does not dispute the market analysis but asks what it leaves out: the conventions that determine which participants appear as negotiating parties and which appear as raw material; the social structures that constrain the options available to workers in particular geographies; the invisibility that prevents the full cooperative network from being recognized as such.