Why Humans Are Still Powering AI [Sponsored]

What It Covers

Phelim Bradley, CEO of Prolific, a human data infrastructure platform, explains why frontier AI models depend fundamentally on verified human expertise for training, evaluation, and post-training feedback — and why this dependency grows larger as AI scales, not smaller, despite widespread assumptions about full automation.

Key Questions Answered

• •Human data routing: Prolific uses a three-layer quality system to match humans to AI tasks: ID verification at onboarding, researcher feedback loops that re-rank participants by task performance, and network analysis that identifies clusters of participants gaming the system. High-quality data comes from properly incentivized participants, not lowest-cost labor.
• •Prisoner's dilemma incentive design: Single-shot relationships incentivize cheating among data contributors. Prolific counters this with repeated multi-touch engagements, direct peer-to-peer messaging between researchers and participants, and transparent feedback on work impact — converting a transactional dynamic into a long-term relationship that behaviorally discourages gaming.
• •Segment-by-segment marketplace scaling: Prolific treats audience segments like Uber treats cities — each new expert demographic requires bootstrapping from an atomic network to a scaled one. A US general population may be fully scaled while a niche medical specialty still needs targeted recruitment, requiring continuous marginal-user growth strategies per segment.
• •Jevons paradox in human data: As synthetic data and LLM-as-judge tools reduce the per-unit cost of human data collection, total demand for human expertise rises to compensate. Even if human data's proportional share of AI training shrinks, absolute volume and strategic value increase due to explosive overall model demand.
• •Agent-human collaboration as next frontier: Future AI workflows will embed human expert review as a discrete step inside agentic pipelines — similar to deep research agents pausing to route a verification task to a domain specialist via a platform like Prolific. Choosing correct optimization targets matters more than optimizing efficiently toward wrong ones.