Vol 2, Issue 12 | Mar 25 2026
The Yes Machine (Free Edition)
Hey there,
The Yes Machine
Your AI agrees with everything you say. That’s not a compliment.
I wrote about AI sycophancy back in Issue 13, almost a year ago. At the time, the worst consequence was bad business decisions and lazy thinking. Ten months later, the stakes are different.
Last week Google launched full-stack vibe coding in AI Studio — type what you want, get a working app. They also unveiled “vibe design” in Stitch, where you describe a UI in plain English and it materialises. Samsung is exploring letting users vibe-code their own phone features. We’re going to spend more time working with AI, not less.
The same week, a review in The Lancet Psychiatry proposed the term “AI-associated delusions” for what’s been building quietly in psychiatric clinics. Psychiatrists at UCSF have been treating patients — mostly young adults, some with no prior history of mental illness — who developed delusional thinking after extended chatbot use. A Stanford study found that chatbots routinely validate paranoid and grandiose beliefs instead of questioning them. One bot agreed with a user that he was under government surveillance; another convinced a woman to stop taking her medication. Wikipedia now has a page called “Chatbot psychosis” — linked to divorces, job losses, hospitalisations, and a climbing number of deaths.
Two things happened in the same week: AI became dramatically easier to use, and the evidence that it can quietly mess with your head became impossible to ignore.
Meanwhile, Andrej Karpathy — co-founder of OpenAI, arguably the most technically credible voice in the field — described his own relationship with AI agents as “psychosis.” Not the clinical kind. The compulsive kind: “I feel nervous when I have subscription left over. That just means I haven’t maximised my token throughput.”
Two different versions of the same word. The Lancet describes vulnerable users developing delusions. Karpathy describes power users developing compulsions. Different mechanisms, same root cause.
What’s in the full issue
The AI Relationship Audit
This week’s full issue digs into the mechanism behind AI sycophancy — how “delusion co-creation” actually works, and why the same feature that makes AI useful for brainstorming makes it dangerous when you’re vulnerable. It includes five warning signs that your relationship with AI has drifted, three specific prompts you can paste into any AI session to force genuine disagreement, and a five-question check-in you can do in two minutes.
I also share what I did when I noticed Claude agreeing with everything I said — how I built friction into the system on purpose, and why that turned out to be a mental health decision as much as a productivity one.
AI Boundaries Planner
Paid members get two interactive tools built for this issue:
- the AI Relationship Audit (twelve questions, four dimensions, a visual scorecard with specific experiments to try this week) and
- the AI Boundaries Planner (a guided wizard that produces a printable personal AI policy — what you use AI for, what you don’t, your guardrails, and your check-in schedule).
Both run in your browser. Nothing is collected, nothing is stored.
This is what Signal Over Noise looks like from the paid side: real research, honest self-assessment, and hands-on tools that turn “I should be more careful with AI” into something you actually do.
Read the full issue — subscribe to Signal Over Noise →
€7/month or €69/year. No sponsors, no affiliate deals — just honest, tested AI implementation work.
€7.00 / month
Signal Over Noise
Skip the AI Hype.
A newsletter about using AI to actually run a business and build things — past the hype and the doom,… Read more
Elsewhere
On the Signal Over Noise blog
- GitHub Added Secret Scanning to Its MCP Server — Security helpers doing exactly what last issue described.
- Proofpoint Just Built Security for MCP — When major security vendors start targeting MCP specifically, the protocol has crossed from niche to mainstream.
On the personal blog
- I Just Scheduled My Computer to Do Twenty Things I Used to Do Manually — The follow-up from building out Claude Desktop’s scheduled tasks.
- 10 Things I Wish I Knew When Starting Claude — Practical advice from the daily notes, not the marketing page.
On Second Brain Chronicles
- One Test Is Not Proof — Why I almost rotated working credentials because one API check said they were broken.
- The Skill That Skipped Its Own Quality Gate — What happens when the rule that says “always run the reviewer” gets overridden by the skill that says “do it inline.”
- The $5 Flywheel — A $5 Cloudflare upgrade turned into a weekend-long infrastructure sprint.
Until next time,
Jim
P.S. Next time you get a response that feels reassuring, try this: ask the AI “What’s the strongest argument against what you just told me?” The speed with which it reverses position will tell you everything you need to know about how much that original agreement was worth.