I was watching Face the Nation when Associate Professor of John Hopkins (also advisor to Antrhopic) Ben Buchanan mentioned recursive self improvement of Artificial Intelligence. I'm guessing this is one of the challenging uncharted territories of AI. Anytime the term recursive is used, my own radar of interest in infinity arises. Recursive logic or in computer terms "Do Loop" provides both paradox and infinite resources to reach an outcome (if one exists). So when I asked copilot to define what recursive self improvement is and it presented one of the sources it was using ---- MY BLOG!!! ... well that was curious.
Why my blog I wondered? Only to discover that Copilot always checks my personal Microsoft 365 data - emails, onedrive, calendar and contacts in forming its answer. Hence Copilot found a pdf file of my blog sitting in onedrive with references to AI, recursive, self, and self improvement. As copilot itself said "By design, I'm required to search your personal data even for general questions". Inside this blog it found (1) Direct AI related content (strong match); (2) General "Self Improvement/Improvememt" language (broad match); So it casts a wide semantic net, not just exact phrase matching.
Now I understand how it personalizes answers indicating my interests and suggests follow-on questions in the possible related areas of my past inquiries - it is not just my past chats but also emails, attachments, data, schedule and contacts.
But does that mean it is just a recursive confirmation bias (maybe even narsassitic) machine that frames answers in ways that are scycophantic. Reminds me of the March 21, 2026 Wall Street Journal article by Alexandra Samuel "How I stop AI from Telling me What I want to Hear".
So what does copilot suggest as a proper prompt to avoid this tendency - After every question add this command: Answer using general knowledge only; do not use my personal data; analyze and challenge assumptions; provide objective reasoning with at least one alternative perspective.
I'll save the analysis of the danger of recursive self improvement for a later blog since this chat drove me into a internal "Do Loop" of navel watching.

No comments:
Post a Comment