Researcher alleges OpenAI trained on user conversations, then claimed a breakthrough

ColinWright · hn · 2026-09-10

A Hacker News discussion centers on an allegation (posted on Bluesky) that OpenAI trained or fine-tuned on user conversations and then presented the resulting gains as a genuine capability breakthrough. Commenters debate evaluation integrity and the practice of buying benchmark points with data. OpenAI has not officially responded; the claim remains unverified.

Original post →

More from Models

Models channel →