Summary

Human expertise remains the hidden input beneath frontier AI. Pádraig Ó Céileachair, Prolific’s co-founder and CEO, describes the company as human-data infrastructure and says data labeling, post-training and evaluation all depend on a “messy layer” of people, making artificial intelligence fundamentally “founded in human intelligence” despite the industry’s clean API narrative.
Prolific’s defensibility lies in verification, behavioral data and incentive design—not access to cheap labor. Participants are identity-checked, ranked through researcher feedback and screened through network analysis; repeated relationships then discourage gaming because “the highest data quality is produced by people who are properly incentivized.”
Expert matching is becoming a recommendation-and-liquidity problem. Prolific pairs task context with human context, analogous to “two towers” algorithms, while bootstrapping each specialist segment as its own atomic marketplace and targeting data collection “in a matter of hours rather than days or weeks.”
AI may expand demand for experts even as it automates parts of their work. Pádraig Ó Céileachair expects junior engineers to learn faster with AI coaches and active practitioners to review consequential outputs; even if humans supply a smaller proportion of training data, explosive model usage means its absolute “scale and importance” could rise.
Frontier-model centralization creates a geopolitical distribution problem. The labor impact may be global while efficiency gains flow mainly to US platform owners, making energy, data centers, locally produced models and training capability a “wake-up call” for the UK and Europe.
Human expertise could evolve into a licensable, recurring-yield asset. Pádraig Ó Céileachair imagines experts receiving ongoing payments when their data improves centralized models or personalized “digital twins”—closer to Spotify royalties or, in the host’s analogy, feeding solar power back into the grid.

Deep dive

1. AI’s clean interface conceals a messy human substrate

Pádraig Ó Céileachair is Prolific’s co-founder and CEO; he describes Prolific as a human-data infrastructure company. His opening claim: “Artificial intelligence is founded in human intelligence.” Across the stack of data, algorithms and compute, human data is the least glamorous input—and therefore the one most routinely glossed over.
Labeling is only the visible edge. People also provide RLHF and post-training data, then evaluate whether models actually perform; the supposedly automated pipeline repeatedly returns to human judgment.
The Mechanical Turk analogy captures the concealment: audiences saw an autonomous chess-playing machine, while a person operated it behind the scenes. Pádraig Ó Céileachair says Prolific wants to offer “human intelligence on demand via an API,” while building direct communication and feedback between researchers and participants rather than erasing the human layer.

2. Data quality comes from relationships, not commoditized labor

Drawing on his code-review startup, the host argues that automation is often really orchestration: a pull request needs an expert who understands the relevant code, not “superficial rubber stamping.” Matching expertise to problems is the underlying business challenge.
At the scale of hundreds of thousands of active participants or raters, Prolific begins with identity and location verification, then incorporates researcher QA into participant rankings. Network analysis identifies pockets of good behavior and participants trying to game the system, downranking them and ultimately removing them when required.
Pádraig Ó Céileachair’s incentive model treats data collection as a repeated game. In a single-shot prisoner’s-dilemma relationship, both sides have an incentive to steal; multiple touchpoints, communication and mutual feedback instead create the “win-win-win dynamic” among researcher, platform and participant.
His operating principle is explicit: “The highest data quality is produced by people who are properly incentivized,” who understand their work’s impact and have motivations beyond the immediate payment. Direct researcher-participant communication supplies that context and empathy.

3. Specification and liquidity are the marketplace bottlenecks

The host’s Upwork pushback—worth keeping—is that specifying an expert task can cost more than doing it yourself. Pádraig Ó Céileachair agrees task design matters as much as audience quality: asking for “PhDs in biology” is inadequate without distinguishing genetics, bioinformatics or healthcare. The platform spans self-contained tasks as well as projects involving weeks or months of training and repeated collection.
Each specialist audience resembles a new Uber city: even after general US or UK populations reach scale, every demanded skill segment must grow from an “atomic network.” Prolific then ranks candidates by task context and human context, analogous to recommendation systems surfacing the most relevant content.

4. The next workflow inserts experts inside agent loops

Prolific favors active healthcare workers evaluating a medical chatbot over people who left the field to become professional annotators. The objective is to reflect “real-world users” while training general participants into skilled, less-biased “high-taste evaluators.”
Pádraig Ó Céileachair rejects the categorical claim that senior engineers plus agents eliminate junior hiring. Prolific is still hiring junior software engineers aggressively because models can be “extremely powerful teachers and coaches,” accelerating their path toward senior-level competence amid elastic demand for more software.
The host imagines a button asking an expert, “Is this legit?” Pádraig Ó Céileachair extends that into deep-research agents whose workflows automatically route consequential outputs for human review. Because models optimize specified goals aggressively under Goodhart’s law, evaluation must approximate “real-world performance for real-world users” rather than rely on simplistic preference scores.

5. Intelligence becomes both geopolitical infrastructure and an asset

Pádraig Ó Céileachair sees frontier models centralizing under a small number of predominantly US players. Their staff may have global intentions, but if superintelligence produces widespread labor disruption, international societies bear the cost while platform owners capture the efficiency gains.
He calls for more UK and European “dynamism and accelerationism”: greater participation in training, locally produced models, data centers and abundant energy for power-hungry systems—even if the region is already late.
Synthetic and human data are complements, not an either-or choice. Cheaper AI-assisted expertise should stimulate demand; Pádraig Ó Céileachair therefore expects human data’s proportion potentially to decline while its absolute scale and importance increase, potentially supporting recurring licensing payments for expertise like Spotify royalties or energy returned to the grid.

Cohere's Chief AI Officer, Joelle Pineau: Why Scaling Laws Will Continue & Future of Synthetic Data Wang Yanpeng on China’s Compute Evolution and a 30,000-GPU Cluster