AI Claude’s Bias Bias 🤖

Guys, the LLM claude.ai let’s you pop in a ton of content, so last week I pasted the entire Three Things Thursday corpus and asked it to help identify the author’s blindspots. At first I was like, “holy shit, it nailed me!” and I felt guilty for all the obvious biases it pointed out that I hadn’t spoken to before—unchecked privilege, lack of representation, insufficient criticality, etc. There were a couple like “attachment to nonduality” that really shook me, but when I asked it to explain more and give examples it recanted “upon closer review” as being “incorrect inferences rather than based on evidence in the text.” This was a little shocking, but also great modeling of taking feedback and updating, for which I thanked and applauded Claude (the AI).

Then I was like, wait a second, even though there’s a lot of truth here, these all sound eerily reminiscent of the green meme / a particular egregore. I asked Claude what it looks for in identifying an author’s blindspots, and sure enough, the top ten list pretty much lined up with the feedback I’d gotten. It doesn’t mean it’s wrong—maybe it’s a good collection of the most common blindspots!—but it sure did remind me to be critical of the bias bias. And it helped me remember that at the end of the day the whole thing was just showing me a mirror of my own fears and projections! Of course I’m biased, in tons of ways that I can’t see, and I only got worried and felt “nailed” because I held some weird belief for a moment that I was supposed to see all of my blindspots or be perfect or something! Instead, I can trust you readers to do what you’re already doing, which is correct for biases in yourselves, or buy into them, or whatever you want to do, because it’s your life and you’re going to be doing it anyway. <3

With love, Jordan