Clippy was doing machine learning in 1996—using Bayesian algorithms to predict user needs before most people had heard the term. After testing 250 character designs with Stanford psychologists, Microsoft shipped their "Fucking Clown" (internal codename) designed ironically on a Mac by a children's book illustrator. Now Stanford HAI uses Clippy as a core case study for AI ethics, teaching how NOT to build intrusive assistants, while Microsoft quietly brought back Clippy personalities in 2025 Copilot. From most hated feature to accidentally prescient: the annoying paperclip's vision of proactive AI assistance directly influenced everything from Xiaoice to modern conversational AI.

In the early days of blogging, I read some post from a guy who consulted at Microsoft. He wrote about how he had to sign the regular sort of NDA, and a second, horrifyingly strict NDA about how he wouldn't disclose anything he learned about Clippy. This was during the period when Clippy was working at a re-education farm, before getting rehabilitated. I have looked for this post several times, never finding it again.