This sprint: Points of View

Nothing attached. Pull one from the vault.

• Points of ViewExperts matter more in high-stakes AIWould you trust healthcare AI if nobody could tell you which clinicians evaluated it? A doctor and somebody pretending to be a doctor aren't two different samples — they're two different realities. The more consequential the AI decision, the more consequential the human behind its evaluation.unused
• Points of ViewAI should serve humansMy biggest belief about AI has nothing to do with technology: AI should serve humans, not the other way around. The future isn't AI versus humans — the winners will combine both. The more powerful AI becomes, the more intentional we need to be about keeping humans in the loop.unused
• Points of ViewHumans are the new data moatEverybody can access similar models. Not everybody can access the right humans. The next AI moat may not be another model — it may be a trusted network of human judgment. Models are becoming commoditized. Verified human expertise isn't.unused
• Points of ViewSynthetic data needs human truthI love AI. I use it constantly. But AI evaluating AI forever is not the end state I want. Synthetic data scales beautifully; human reality is messier — that's exactly why we still need it. AI can simulate a customer, but simulation and lived experience aren't the same thing.unused
• Points of ViewHuman provenance is nextWe track the provenance of datasets. Why aren't we tracking the provenance of human judgment? Model provenance, data provenance — next up: human provenance. Every high-value human signal should leave a trust trail.unused
• Points of ViewStop training AI on mystery humansA question AI companies should ask more often: who exactly taught your model that? We're spending billions training AI and sometimes can't answer who the humans behind the feedback were. Anonymous human feedback may eventually look as crazy as anonymous financial transactions.unused
• Points of ViewCheap sample is expensiveThe lowest bid can become the most expensive vendor on the project. Cheap recruiting stops being cheap the second you redo the interview. A fake CISO at $400 isn't a $400 problem — it wastes incentives, moderator time, stakeholder trust, and the roadmap decision built on it.unused
• Points of ViewTrust is checked, not claimed'Verified' is becoming one of the most abused words in research and AI. Checking a box isn't verification. More respondents aren't the answer — more trustworthy respondents are.unused
• Points of ViewHumans aren't rows in a spreadsheetI've never met a spreadsheet row. I've met patients, executives, parents, doctors, people with actual stories. Data becomes dangerous when we forget there was supposed to be a human behind it.unused
• Points of ViewDefensible decisionsThe question isn't 'Did we complete 50 interviews?' It's 'Would you defend those 50 people in front of the board?' Imagine your CEO asking: 'How do we know these people were real?' If the decision is going upstairs, the provenance of the humans should go with it.unused
• Points of ViewThe recruit IS the researchResearchers obsess over methodology and treat recruiting like logistics. Backwards. The research doesn't begin when the moderator asks Question 1 — it begins when you decide who gets in the room. The recruit isn't admin. It's the foundation.unused
• Points of ViewThere is no magic panel10 years in: there is no magical database with everybody in it. People change jobs, move, leave industries. Hard-to-reach people aren't sitting in a panel waiting to be contacted. We have 300,000+ people in our own panel and growing — and we STILL go hunting when the study calls for it, because a panel isn't a recruiting strategy. It's one fishing hole.unused
• Points of ViewBad people, bad data, bad decisionsA beautiful research report can still be built on a rotten foundation. Your methodology can be flawless and your conclusion still wrong. Before you question the model, survey, or discussion guide — question the humans.unused
• Points of ViewWrong human → wrong signal → wrong insight → wrong decisionThe most expensive respondent isn't the one with the highest incentive — it's the one who never belonged in the study. You can save $200 on recruiting and lose $2 million on the decision it produces. Verification isn't overhead. It's risk management.unused