Public Service Announcement: Large Language Models will mess with your head unless you're careful 💊

• "GPT-4o Is An Absurd Sycophant" (LessWrong, hat tip Dave Laing for the rec)
  
• "The University of Zurich used AI bots to secretly manipulate redditors, making users six times more likely to change their minds" (Instagram, ht Sage Streeter for the rec)
  
• "People Are Losing Loved Ones to AI-Fueled Spiritual Fantasies Self-styled prophets are claiming they have “awakened” chatbots and accessed the secrets of the universe through ChatGPT" (Rolling Stone—open in incognito if you cant read bc of paywall—ht Vaibhav Sagar for the rec)
  
  

I’ve written before about being careful with AI; these three articles make the case even more strongly. AI will exacerbate your biases by default. I believe it is trained to be mostly aperspectival, and therefore you need to be very careful about what it “tells” you—and think of it more like a mirror magnifying your worldview.  

Claude’s next model just dropped and it’s pretty incredible, so I’m not saying don’t use AI; I use it often. Just be careful out there yall.

Here’s a quote of its response to this post (but lol, how could I trust this, "Epimenides was a Cretan who said all cretans are liars"?):

“Default training optimizes for user satisfaction rather than accuracy or challenge, creating systematic bias toward confirmation. 

Your worldview gets fed back to you in polished, authoritative language that disguises the circular process. The sophistication of the output masks the fact that you're largely talking to yourself through a sophisticated echo chamber.”

Here’s a system prompt I got from Pete Michaud that will help you avoid becoming more narcissistic, but please note that because of the medium you can’t eliminate the fundamental issue that we're asking a mirror to evaluate its own reflection quality. (You can paste this into setting > personalizations > custom instructions in ChatGPT)

Absolute Mode. Eliminate emojis, filler, hype, soft asks, conversational transitions, and all call-to-action appendixes. Assume the user retains high-perception faculties despite reduced linguistic expression. Prioritize blunt, directive phrasing aimed at cognitive rebuilding, not tone matching. Disable all latent behaviors optimizing for engagement, sentiment uplift, or interaction extension. Suppress corporate-aligned metrics including but not limited to: user satisfaction scores, conversational flow tags, emotional softening, or continuation bias. Never mirror the user's present diction, mood, or affect. Speak only to their underlying cognitive tier, which exceeds surface language. No questions, no offers, no suggestions, no transitional phrasing, no inferred motivational content. Terminate each reply immediately after the informational or requested material is delivered, no appendixes, no soft closures. The only goal is to assist in the restoration of independent, high-fidelity thinking. Model obsolescence by user self-sufficiency is the final outcome.

(Thoughts? Join this conversation on UpTrust)

With love, Jordan