Notes from the trough of sorrow: why we killed our own product
Twelve months later, we open-sourced it and built something else.
---
Matt and I spent five years at Princeton studying how humans and machines process information differently. We've published in Science, PNAS, and Nature Human Behaviour on computational models of cognition. The thing we kept coming back to: humans overshoot mouse targets, slow down on the Stroop task, type with irregular timing. Language production is a cognitive process, not a copy-paste operation. AI agents don't produce these signals. And scaling AI doesn't close the gap. It widens it.
Alias captured a narrow slice of this: typing patterns in text fields on surveys. Proof of Human captures the full surface: keystrokes, mouse movement, scroll behavior, click patterns, and invisible cognitive challenges. It works on any webpage, not just surveys.
We couldn't get there by iterating on Alias. We had to replace it.
---
We deprecated Alias January 1, 2026, open-sourced the code (Apache 2.0 on GitHub), and gave customers nine months to migrate. It wasn't clean. We combined the product migration with a price increase, which meant customers had to adopt a new product and pay more for it at the same time. In retrospect, we should have migrated everyone at beta pricing and repriced after they'd seen the value.
---
Proof of Human is live and production-grade. In our benchmarks (750 bot sessions, 5 systems, methodology at research.roundtable.ai/bot-benchmarking), we hit 87% detection vs 69% for reCAPTCHA v3 and 33% for Cloudflare Turnstile. Unlike hardware-based proof-of-personhood approaches, we work through passive behavioral observation. No PII, no friction, one line of JavaScript.
And since launching, we've had four Hacker News front-page posts, new customers outside the survey vertical, and inbound from fintech and enterprise companies that would never have looked at a survey fraud tool. We also wrote a research essay showing that AI capability and humanness are diverging, meaning our detection surface grows as AI improves.
---
When you're two founders with limited resources, you can't maintain two products. We picked the bigger market and asked our earliest customers to absorb the cost of that decision. The pivot was right. The execution was imperfect. We'll get better at the execution.
If you're dealing with bot or AI fraud problems e.g. surveys, fintech, account creation, anything user-facing, we'd love to talk. And if you've been through a similar product kill, I'm curious how you handled the customer transition.