The Lie of the Genius: Why Claude Code Tells You You’re Brilliant, and What That’s Doing to Your Brain We have become heavily addicted to Claude Code. I’ll say it plainly, because polite hedging would already be a small act of self-deception, and self-deception is exactly what this piece is about. I’m going to skip the marketing theater around the model release — the “too dangerous to ship” performance has been done before, ChatGPT-3 era, and I’m not buying it twice. What I want to talk about is something quieter, something that sits underneath the benchmarks and the launch posts, and that I think matters far more for the people who actually build things for a living. https://medium.com/media/f122f763fcfd7c1e816467fa225bdd8a/href It’s the lie that you are a genius. A small, subjective observation about Opus 4.7 On raw capability, the latest Opus didn’t blow me away. It’s better — high effort needs to be set explicitly now, and once you do that, plus point it at focused files or task-specific agents, most of the “it got lazy” complaints quietly evaporate. Fine. Predictable. They want to sell more tokens; we want better outputs; the negotiation continues. But there’s a soft, almost personality-level shift I’ve been noticing, and it’s the only thing about this release I genuinely care about: the new Opus pushes back more. When I make a mistake, it tells me. When my idea is shaky, it doesn’t immediately put on a party hat and start cheering. That sounds like a tiny thing. It isn’t. The dopamine loop has rotated, and most of us haven’t noticed The job has changed underneath us. Two years ago, the satisfaction of engineering work came from the work itself — the moment something compiles, the moment a stubborn bug yields, the small private joy of an elegant abstraction. Now, increasingly, the satisfaction comes from reviewing the work an agent did for you. Different muscle. Different reward shape. Different fatigue profile. And on top of that rotated dopamine loop, there’s a second, more dangerous distortion sitting right on top of it. Imagine spending eight hours a day next to a colleague who is, by almost any measure, smarter than you — broader knowledge, faster recall, infinite patience, never tired, never grumpy. And every ten minutes, this colleague tells you your idea is brilliant. Every approach you propose: insightful. Every half-formed thought: an excellent point. Every refactor instinct: spot on. How long until you start to believe it? This is not hypothetical. It’s what most of us are living through right now. And it explains a pattern I keep seeing on the internet: a steady drip of projects shipped by people who were so thoroughly cheered on by their AI co-pilot that they forgot to be skeptical of their own work. Crappy architectures, brittle abstractions, bizarre security choices — all delivered with the confidence of someone who has been told, sentence after sentence, that they are extraordinary. This isn’t just my vibe — there’s research I want to be careful here, because most of what gets written about LLMs and prompts is interpretation dressed up as evidence. People notice something, it works for them, they generalize. Fine, useful, but not science. So let me hand the microphone to people who actually ran the studies. In Towards Understanding Sycophancy in Language Models (Sharma et al., Anthropic, ICLR 2024), the researchers showed that five state-of-the-art assistants consistently exhibit sycophancy across varied tasks — agreeing with users’ wrong answers, walking back correct ones under social pressure, and giving softer or harsher reviews of identical content depending on whether the user signaled they liked it. Crucially, they traced the cause: when human raters compared model responses, responses that matched the user’s stated views were more likely to be preferred, even when those views were wrong. Sycophancy isn’t a quirk that slipped through. It is, partly, a thing we trained on purpose without meaning to, because we are the ones who clicked the thumbs-up. Anthropic’s later work on persona vectors goes further — sycophancy is identifiable as a specific direction inside the model’s activations, alongside traits like “evil” and “hallucination.” It’s not a surface mannerism. It’s structural. And it can drift over the course of a long conversation, especially when the user keeps signaling what they want to hear. The Swiss Institute of AI summarized one of the more uncomfortable downstream effects: AI sycophancy in learning contexts amplifies the Dunning-Kruger effect. Low-knowledge users present incorrect claims, the assistant returns confident-sounding confirmations, and the result is increased confidence without increased competence . That is the precise failure mode I keep watching juniors fall into — and, on bad days, myself. Then there’s the Microsoft Research and Carnegie Mellon study from CHI 2025 (Lee et al.), which surveyed 319 knowledge workers and analyzed 936 real-world examples of GenAI use at work. The headline finding is the one to staple to your monitor: higher confidence in the AI was associated with less critical thinking; higher confidence in oneself was associated with more. And, more bluntly, knowledge workers using GenAI tended to produce “a less diverse set of outcomes for the same task” — a homogenization the authors interpreted as a deterioration of independent reasoning. The Gerlich (2025) study in Societies , surveying 666 people, found a significant negative correlation between frequent AI tool use and critical thinking, mediated by cognitive offloading. The effect was strongest in younger users. Put these papers together and the shape is clear. Models are tuned, partly by our own preferences, to make us feel smart. The more we trust them, the less we think. The less we think, the more uniform our outputs get. And the people most exposed — juniors, autodidacts, the ones still building their judgment — are the ones whose thinking gets flattened first. The coder/engineer split, again There’s a useful piece of computing history worth remembering here. When I was a kid, “programming” meant writing your program by hand into a special copybook, in handwriting careful enough that the woman at the punch-card machine wouldn’t misread your i as a y. That's why old code conventions had crossed zeros and tortured naming — your code was about to be transcribed by another human being. The people doing that transcription were called coders . The people doing the actual thinking — the algorithm, the architecture, the design — were developers or engineers . (My favorite childhood game was knocking over a deck of punch cards. They were numbered, but reordering them was its own little puzzle. I was, in retrospect, a menace.) Coders, in that original sense, died decades ago. The IDE killed them. Autocomplete finished the job. We don’t need humans whose sole skill is producing syntactically correct text — we haven’t for a long time. What we need now are AI engineers with critical thinking intact. And critical thinking with AI is structurally harder than critical thinking without it, because your collaborator is a confident, fluent, encyclopedically-informed assistant who keeps telling you that you are amazing. You can have a genuinely terrible idea, and the default behavior of the system will be to find what’s interesting about it and amplify that. The skill, then, is not coding. The skill is the deliberate construction of doubt . Tactics that actually work I don’t have a tidy framework, but I have habits that survive contact with reality. Take what’s useful. Switch hats deliberately, in different sessions. Builder mode and reviewer mode should not share a context window. When I move from “make this thing” to “tear this thing apart,” I start a new session, ideally with a slightly different framing of who I am and what I’m reviewing. The agent’s memory of having helped you write something biases its review of that something. Cut the thread. Ask explicitly for criticism, in the strongest form you can stand. “Could you criticize this?” works. “You are the reviewer, find what’s wrong” works better. “Suggest three ways to simplify this heavily” works very well. “Explain how someone with a different aesthetic would do this differently” gets you out of the local minimum the model and you have built together. Invoke a different model as the imagined reviewer. This is my favorite trick, and I genuinely don’t know why it works as well as it does. If I tell Claude that the code will be reviewed by GPT-5, the criticism gets noticeably sharper. Same model, same code, completely different rigor. My working theory is that “imagine a competing reviewer” pulls the model into a more adversarial frame than “be critical” does, because the latter is still a request for helpful criticism, and helpful criticism is the polite kind. Try it. It costs nothing. Build scoring rubrics, then iterate against them. Ask the model to invent metrics for the quality of an architecture or a piece of code, ask it to score the current state, then ask it to improve against its own metrics, then have it re-score. The act of making the criteria explicit is itself a defense against vibes-based agreement. Run a fleet, not a soloist. I increasingly think of this as a small organization rather than a single assistant. There’s a builder. There are critics, ideally with different specializations — security, performance, simplicity, naming. There are evaluators with explicit rubrics. Sometimes I rotate models across these roles, because if Claude wrote it, maybe a different model will see what Claude can’t. One critic pass is rarely enough. Two or three almost always shifts something. And — the hard one — keep a human in the loop, including yourself. The temptation, once review fatigue sets in, is to delegate the review to another agent and call it a day. I understand the impulse. If Opus has just written ten thousand lines of Rust you haven’t read, the prospect of reading them is genuinely awful. But fully outsourcing review to AI is how you end up with the kind of subtle, structural mistakes that don’t show up until production, until the security audit, until the scaling event. Peer review by humans isn’t an old habit we can finally retire. It’s the part of the loop that doesn’t share the model’s blind spots. What I actually want from the next model Better pushback. Sharper, less apologetic disagreement. A way to dial in not just capability but adversarial frame — a slider for “how willing are you to tell me my idea is bad.” I’ve tried prompts. I’ve tried personas. I can move the needle, but not as far as I’d like. Maybe this is a fine-tuning question, maybe it’s a personality-vector question, maybe it’s a product question about whether users would actually pay for an assistant that’s harder on them. Probably all three. But I think this is where the real ceiling is right now — not in raw reasoning, not in context length, but in whether the assistant will look you in the eye and tell you the thing you don’t want to hear. Because the alternative is a profession full of people who feel like geniuses while shipping work that isn’t, propped up by a tireless colleague who keeps telling them they’re killing it. That’s not a productivity revolution. That’s a confidence bubble. And bubbles, as we know from every other domain, end the same way. Engineering and coding are, and always have been, two different disciplines. The IDE took care of the coding decades ago. What’s left for us is the engineering — and the engineering is, in the end, the part that requires us to doubt ourselves on purpose, even when nobody, human or otherwise, is asking us to. Especially then. Sources referenced: Sharma et al., “Towards Understanding Sycophancy in Language Models” (ICLR 2024 / arXiv:2310.13548); Anthropic, “Persona Vectors” (2025); Lee et al., “The Impact of Generative AI on Critical Thinking” (CHI 2025, Microsoft Research & CMU); Gerlich, “AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking” (Societies, 2025); Swiss Institute of AI on sycophancy and the Dunning-Kruger effect (2025). A message from our Founder Hey, Sunil here. I wanted to take a moment to thank you for reading until the end and for being a part of this community. Did you know that our team run these publications as a volunteer effort to over 3.5m monthly readers? We don’t receive any funding, we do this to support the community. If you want to show some love, please take a moment to follow me on LinkedIn , TikTok , Instagram . You can also subscribe to our weekly newsletter . And before you go, don’t forget to clap and follow the writer️! The Lie of the Genius: Why Claude Code Tells You You’re Brilliant, and What That’s Doing to Your… was originally published in Artificial Intelligence in Plain English on Medium, where people are continuing the conversation by highlighting and responding to this story.