#325 Phelim Brady: Why AI's Future Depends on Human Judgement
Episode
47 min
Read time
2 min
Topics
Health & Wellness, Investing, Startups
AI-Generated Summary
Key Takeaways
- ✓Model Evaluation Over Benchmarks: Academic benchmarks like math Olympiad scores have become unreliable because frontier models train directly against them, saturating their usefulness. Enterprises and labs now need real-world human evaluation instead. Prolific runs structured head-to-head model comparisons using demographically controlled participant pools to generate trustworthy, context-specific performance rankings across different use cases.
- ✓Demographic Representation Changes Model Rankings: Prolific's "Humane" benchmark replicates chatbot arena-style model comparisons but adds census-matched sampling controlling for age, ethnicity, and political affiliation. Results show model preference rankings shift measurably depending on audience demographics, meaning enterprises should evaluate models against their specific target user population rather than relying on aggregate public leaderboards.
- ✓Agentic Fraud Is an Emerging Threat to Human Data: AI agents can now replicate human behavior accurately enough to infiltrate online data collection platforms. Prolific counters this through layered identity verification, periodic re-verification, behavioral analysis during task completion, and KYC-style checks — making participant authenticity a core infrastructure investment rather than a one-time onboarding step.
- ✓Expert Participant Segmentation Across Three Tiers: Prolific structures its AI evaluation workforce into three roughly equal segments: general consumer samples for representativeness testing, trained taskers qualified for standard AI evaluation workflows, and domain experts whose prior professional credentials are the primary selection criterion. Enterprises building regulated-domain applications in healthcare, finance, or law should target that third expert tier specifically.
- ✓Human Judgment Remains Necessary at the Capability Frontier: Automated AI judges cannot evaluate capabilities that do not yet exist in current models — assessing a gap requires a benchmark that exceeds the model being tested. Human evaluators remain essential wherever tasks involve ambiguity, subjective preference, or low-confidence model outputs requiring escalation, meaning evaluation budgets should preserve human review for edge cases and novel capability assessments.
What It Covers
Phelim Brady, cofounder and CEO of Prolific, explains how his human data platform connects verified global participants with AI labs and researchers for post-training evaluation. With roughly 2 million registered participants and a 50/50 split between academic research and AI work, Prolific addresses the growing demand for rigorous human judgment in model evaluation.
Key Questions Answered
- •Model Evaluation Over Benchmarks: Academic benchmarks like math Olympiad scores have become unreliable because frontier models train directly against them, saturating their usefulness. Enterprises and labs now need real-world human evaluation instead. Prolific runs structured head-to-head model comparisons using demographically controlled participant pools to generate trustworthy, context-specific performance rankings across different use cases.
- •Demographic Representation Changes Model Rankings: Prolific's "Humane" benchmark replicates chatbot arena-style model comparisons but adds census-matched sampling controlling for age, ethnicity, and political affiliation. Results show model preference rankings shift measurably depending on audience demographics, meaning enterprises should evaluate models against their specific target user population rather than relying on aggregate public leaderboards.
- •Agentic Fraud Is an Emerging Threat to Human Data: AI agents can now replicate human behavior accurately enough to infiltrate online data collection platforms. Prolific counters this through layered identity verification, periodic re-verification, behavioral analysis during task completion, and KYC-style checks — making participant authenticity a core infrastructure investment rather than a one-time onboarding step.
- •Expert Participant Segmentation Across Three Tiers: Prolific structures its AI evaluation workforce into three roughly equal segments: general consumer samples for representativeness testing, trained taskers qualified for standard AI evaluation workflows, and domain experts whose prior professional credentials are the primary selection criterion. Enterprises building regulated-domain applications in healthcare, finance, or law should target that third expert tier specifically.
- •Human Judgment Remains Necessary at the Capability Frontier: Automated AI judges cannot evaluate capabilities that do not yet exist in current models — assessing a gap requires a benchmark that exceeds the model being tested. Human evaluators remain essential wherever tasks involve ambiguity, subjective preference, or low-confidence model outputs requiring escalation, meaning evaluation budgets should preserve human review for edge cases and novel capability assessments.
Notable Moment
Brady describes a UK AI Security Institute study where participants held real political conversations with roughly 20 different AI models, each instructed to use specific rhetorical strategies. Researchers measured opinion change before and after, revealing measurable differences in how persuasive various models were with real people.
Episode Transcript
I have in past interviews with people, working with human resources for AI is that it's kind of the dirty little secret, the of AI that it's built with humans or or that a lot of the other these armies of of human evaluators or or labelers out there. Or computational biology. But while there realized that there was core pain points in running high quality, academically rigorous human subjects experiments online. My name is Fainim Bradley. I'm the cofounder and CEO of Prithvik, And Perifik is a human data platform. So we connect both the infrastructure methodology and high quality global pool of humans and connect them with data collectors for purposes of research and human data for AI, in particular, post training and evaluation use cases. Started pretty quick when I was doing a PhD at the University of Oxford. I actually did my PhD in bioinformatics or computational biology. But while there, realized that there was core pain points in running high quality, academically rigorous human subjects experiments online. So platforms that were popular at the time had problems of poor quality data and verification of the humans who are taking part in research. The tooling and infrastructure was extremely poor user experience for both sides of the platform, again, impacting both the ease of use and the quality of the resulting resulting data. And, ultimately, it was a created a kind of a a an unhealthy platform and dynamic. We set out to build Prolific in order to solve that problem. And, originally, it was a side project while I was completing my PhD and grew the company to the reasonable attraction before completing the my my PhD. And the story since then really has been rediscovering that same core pain point across a range of different industries. How do I recruit high quality data online from a an audience of highly verified trustworthy participants? This is a pain point across research applications, so re user research, polling, academic research, which was kind of our origin story. And now more recently also human data in the AI development life cycle. And I think we we try to apply some of this methodological and academic rigor of the behavioral sciences to some of the problems in in evaluation of generative AI applications. Yeah. And this is an interesting area, and I know you some of your other interviewers have, pointed this out as I have in past interviews with people working with human resources for AI is that it's kind of the dirty little secret of AI that it's built with humans or that a lot of the other disease armies of human evaluators or labelers out there. My experience or my knowledge of this goes back to interviewing Faylee, who I just had on the program. But the first time I interviewed her was about ImageNet, which was this massive database, really the first image database for training supervised learning programs. And it was her dataset …
Get the full transcript (7,190 words) + summary by email — free
One-time email with the complete transcript and AI summary of this episode. No account needed.
One email, no spam. We’ll also show you what SignalCast does.
You just read a 3-minute summary of a 44-minute episode.
Get Eye on AI summarized like this every Monday — plus up to 2 more podcasts, free.
Pick Your Podcasts — FreeKeep Reading
More from Eye on AI
86% of What Coding Agents Do Is Just Reading — Not Solving | Alexander Whedon of Subquadratic
Sep 8 · 54 min
Machine Learning Street Talk
Why Humans Are Still Powering AI [Sponsored]
Nov 3
More from Eye on AI
From 10 Drones a Month to Nearly 100,000 — Inside Ukraine's Largest Drone Manufacturer | Marko Kushnir, General Cherry
Sep 3 · 38 min
Latent Space
Simulation: the new Scaling Law — Joon Sung Park, Simile AI
Aug 21
Books, tools, and gear mentioned in this episode
SignalCast may earn commission on purchases via these links.
company
“Phelim Brady, cofounder and CEO of Prolific, explains how his human data platform connects verified global participants with AI labs and researchers for post-training evaluation.”
More from Eye on AI
We summarize every new episode. Want them in your inbox?
86% of What Coding Agents Do Is Just Reading — Not Solving | Alexander Whedon of Subquadratic
From 10 Drones a Month to Nearly 100,000 — Inside Ukraine's Largest Drone Manufacturer | Marko Kushnir, General Cherry
In 5 to 10 Years, Using Weapons Without AI Will Be Considered Unethical | Yaroslav Azhnyuk, The Fourth Law
Inside Ukraine's Azov Drone R&D: The Engineer Building AI Weapons 18 km From the Front Line | Alexander Palamarchuk
95% of AI Agent Projects Fail to Reach Production. Here's Why | Manoj Saxena, TrustWise
Similar Episodes
Related episodes from other podcasts
Machine Learning Street Talk
Nov 3
Why Humans Are Still Powering AI [Sponsored]
Latent Space
Aug 21
Simulation: the new Scaling Law — Joon Sung Park, Simile AI
Latent Space
May 20
Railway: The Agent-Native Cloud — Jake Cooper
Practical AI
Feb 13
AI incidents, audits, and the limits of benchmarks
Machine Learning Street Talk
Oct 18
The Secret Engine of AI - Prolific [Sponsored] (Sara Saab, Enzo Blindow)
Explore Related Topics
This podcast is featured in Best AI Podcasts (2026) — ranked and reviewed with AI summaries.
Read this week's Health & Longevity Podcast Insights — cross-podcast analysis updated weekly.
You're clearly into Eye on AI.
Every Monday, we deliver AI summaries of the latest episodes from Eye on AI and 192+ other podcasts. Free for one show.
Start My Monday DigestNo credit card · Unsubscribe anytime