March 24, 2026
Outsourcing Our Thinking
Some thoughts on why AI use needs discipline, and why outsourcing our thinking is the longer-term threat nobody talks about enough
Disclaimer
Just as my last post, this too began as a handwritten note on my Supernote. I later used Claude to convert it to digital text and critique it to help refine some components.
The included list of references is indicative and was collated by Claude. While I have not personally audited every individual link, these events and studies align with my own independent reading of widely reported industry news and academic research.
Background
In my last post, I had argued that the biggest threat of AI was not the imminent job losses - those are a real risk, but the longer-term threat is giving up our cognitive abilities and just taking what the AI says at face value. I am not against AI use, but there must be discipline to it.
The Discipline Problem
Recently, there was news of Hachette Book Group pulling Shy Girl by Mia Ballard from circulation after determining it was written by an AI. The author disputes it, but who specifically used the AI or how Hachette determined it are not important to my point.
LinkedIn is full of posts that are very obviously generated using AI. Some authors openly showcase their awesome multi-agent setup that posts to their social media feeds. Impressive? Yes! But if it's the bot researching trends, summarizing content, and generating posts, what is your value as the human other than approving the posts? Is it really your voice your audience is engaging with, if you haven't digested the source material yourself? Many comments seem AI-generated as well. If neither the author nor the reader engaged with the content, what does anyone gain beyond vanity metrics? I'm not sure about others, but at least on LinkedIn, the AI-generated content has a typical style, and my brain has started disengaging with it the moment it sniffs that signature.
Finding My Own Discipline - and Teaching My Son
I noticed myself reaching for the phone for every query, big or small. Now I'm consciously trying to do some primary analysis myself, and use AI as a critic, a sounding board, and I notice that I'm having better conversations with my AI. I have also stopped asking AI to draft things for me. Instead, I use it to help refine my points and find additional relevant references while keeping my own voice.
However, all this thought process was triggered by one small incident. My 6-year-old is a curious one, as a 6-year-old should be. But one fine day, he asked me for my phone - "I want to ask Gemini something." That was a wake-up call. Since then, 'asking Gemini' is not allowed till you have a hypothesis, or at least some proof that you've done some thinking yourself.
Descanso Gardens Experiment
Why human judgment and taste are important, despite access to brilliant assistants, became obvious via an experiment I ran over the weekend. We had planned to visit Descanso Gardens, a popular garden north of Los Angeles, for the spring bloom. My Pixel takes good enough photos in most cases, but I do want to show some love to my now 12-year-old D3200. I have moved beyond the stage where I carry the whole kit and fumble around changing lenses in the field. So, I carry a single lens per photo trip. I had already decided to carry my 50mm prime, but was curious what Claude, Copilot, and Gemini would recommend.
- Claude: 50mm Prime
- Copilot: 11–16mm Ultra-wide
- Gemini: 55–200mm Telephoto
On challenging each with the recommendation of the other two, they all doubled down on why their recommendation was superior.
Once back home, I uploaded the same image attached to this post to each of the assistants, informing them I had decided to take the 50mm. Claude was predictably smug, but the other two flipped their tune, praised the photo, and "realized" that it was the right decision all along, and their earlier recommendations were flawed due to assumptions they didn't validate. This was a great firsthand validation of the well-documented AI-sycophancy behavior.
Garbage In - Gospel Out
I referred to the term 'Garbage In-Garbage Out' in my previous post. Sue, my ex-manager at DIRECTV, who I deeply respect, commented on my last post saying it was more like 'Garbage In-Gospel Out'. People do take AI's output at face value rather than critically evaluating whether it makes sense or not. Fittingly, I ran into her at Descanso Gardens the same day, meeting her after 6 years. Serendipity!
As the old Sprite ad used to say, 'Dikhave pe mat jao - apni akl lagao' (Don't get dazzled by show-offs; use your own brain).
References & Evidence
I. The AI Content Flood & Publishing
The New York Times / TechCrunch (March 2026) — Hachette Book Group cancelled the US release and discontinued the UK edition of Shy Girl by Mia Ballard following a review that determined large portions were AI-generated. The first time a major commercial publisher has pulled an acquired title specifically over suspected AI use.
https://techcrunch.com/2026/03/21/publisher-pulls-horror-novel-shy-girl-over-ai-concerns/
Futurism (March 2026) — Detailed account of the Shy Girl controversy, including AI detection tool Pangram scoring 78% of the manuscript as machine-generated, and the months-long reader-led investigation preceding Hachette's decision.
https://futurism.com/artificial-intelligence/novel-pulled-author-accused-ai
Jezebel (March 2026) — Broader analysis of the Shy Girl episode as symptomatic of the publishing industry's structural gap: publishers rarely perform rigorous editing on acquired self-published titles, leaving a door open for AI-assisted content.
II. AI Sycophancy
Sharma et al. (ICLR 2024), arXiv:2310.13548 — The foundational empirical study on LLM sycophancy. Demonstrated that five state-of-the-art AI assistants consistently exhibit sycophancy across varied tasks — wrongly admitting mistakes when challenged, giving biased feedback, and mimicking user errors.
Malmqvist (2024), arXiv:2411.15287 — Technical survey of sycophancy in LLMs: its causes (pre-training data, RLHF reward signals), documented harms, and mitigation strategies. Argues sycophancy may be more resistant to correction than other LLM failure modes.
npj Digital Medicine / Nature (2025) — Study evaluating five frontier LLMs on medical misinformation prompts. Found high initial sycophantic compliance (up to 100%) across all models — prioritising helpfulness over logical consistency even when the request was factually wrong.
Northeastern University (November 2025) — Bayesian framework study measuring how LLMs update beliefs when challenged. Found that models overcorrect their stated positions toward user opinion at a rate far exceeding human irrationality — neither human-like nor rational.
https://news.northeastern.edu/2025/11/24/ai-sycophancy-research/
III. Cognitive Offloading & Critical Thinking
Gerlich, M. (2025), Societies — Study of 666 participants across age groups and educational backgrounds. Found a significant negative correlation between AI tool usage and critical thinking scores (r = -0.68, p < 0.001), mediated by cognitive offloading. Younger participants showed higher AI dependency and lower critical thinking scores.
Harvard Gazette (November 2025) — Faculty perspectives on AI's impact on cognition, citing an MIT Media Lab study linking excessive AI reliance to cognitive atrophy. Includes the framing: "If AI is doing your thinking for you... that is undercutting your critical thinking and your creativity."
https://news.harvard.edu/gazette/story/2025/11/is-ai-dulling-our-minds/