Enzo Blindow

Today we're putting our AI research work out in the open: labs.prolific.com We started our AI research team less than two years ago with a founding mission to work on the science of evaluation: how to measure AI systems well, grounded in real human judgement at population scale. We have since expanded into alignment and interpretability. Some of what we've built so far: - HUMAINE, our demographically stratified preference leaderboard; 51 models, 57,000+ human evaluations across 20 demographic groups - Alignment leaderboard stress-testing models across 904 multi-turn scenarios on honesty, safety, manipulation, corrigibility, scheming - 6 papers so far, including a spotlight at ICML 2026, work at ICLR 2026, and more under review with Nora Petrova, Jerome W., John Burden P.S. We're hiring, and open to collaborators. Please reach out.

Prolific is at ICLR next week. If you're in Rio, let's meet up. We're bringing two pieces of research about what changes when you stress-test evaluation methodology: HUMAINE: a demographically stratified leaderboard for LLMs, built on 23,000+ real human conversations across 22 demographic strata in the US and UK. We show how aggregate leaderboard scores conceal meaningful disagreement. lnkd.in/ecMCtzCJ The Missing Red Line (ICBINC Workshop): 8 frontier models tested in scenarios where commercial objectives conflict with user safety. They fabricate safety information, dismiss medical risks, and don't become more cautious as harm escalates. lnkd.in/eYwCa5wF I'm also giving an expo talk on why your models are outgrowing your evaluations. DM me if you want to meet up - whether that's about evals, human data infrastructure, or joining us. We're also co-hosting an afternoon tea social with Encord, details soon. #ICLR26 #AIResearch