TL;DR: A best practice is a compressed answer to a question someone already solved. An AI model is the most powerful best-practice machine ever built — ask it almost anything and it returns the consensus, the median of everything it has absorbed. That is genuinely useful and usually correct, which is exactly the trap: when a few hundred million people can summon the same best-practice answer in seconds, following it stops being an edge and becomes table stakes. The research backs the worry. In a controlled study of 293 writers, access to generative-AI ideas made individual stories more creative but made the stories collectively more similar to one another (Doshi & Hauser, Science Advances, 2024). People also over-rely on AI even when it contradicts what they can see, and they keep making the AI’s mistakes after the AI is taken away (a 2023 Scientific Reports study on inherited AI bias). The durable skill in an AI-saturated world isn’t out-prompting the model — it’s the judgment to know when the consensus is right and when to throw it out and reason from first principles. This article gives you an original decision matrix for exactly that call: decide like a CEO who owns the downside of breaking a rule, and learn like a student who understands the rule well enough to rebuild it from nothing.
There is an old engineering distinction that has quietly become the most important career skill of 2026. Best-practices thinking reasons by analogy: this worked before, in a situation like this one, so do it again. First-principles thinking reasons from the ground up: strip the problem down to the things you know are true, and rebuild the answer from there, even if it contradicts what everyone does. Aristotle named the second kind — a first principle is “the first basis from which a thing is known” — and it is the slower, more expensive, more error-prone way to think. For most of history that cost is why best practices won: reinventing the wheel is a bad use of a human life, and copying what works is rational.
AI changes the economics of that trade in a way most people haven’t fully absorbed. A large language model is, at its core, a machine for retrieving and recombining best practices at near-zero cost. It has read the playbooks, the case studies, the top-ten lists, and it will hand you the distilled consensus faster than you could find the first source yourself. This is a gift. It is also a homogenizing force, and once you see it you can’t unsee it: the same tool that makes the conventional answer free also makes it universal. The CEO+Student question this article answers is the one the model can’t answer for you: given that the consensus is now a commodity, when do you take it, and when do you do the harder thing and reason from first principles?
Why AI pulls you toward the consensus answer
Before the framework, it helps to see why this is structural and not a temporary flaw to be patched in the next model. A language model is trained to predict the most likely continuation of text. “Most likely” is, almost by definition, the average — the center of the distribution, the modal answer. Ask for “the best way to write a cover letter” and you get the cover letter the training data agrees on. That is a feature for routine questions and a hazard for distinctive ones, and the empirical work now shows the effect at the level of populations, not just prompts.
What the research actually shows about AI and the median answer (verified)
| Finding | What the evidence shows | Source (year) |
|---|---|---|
| AI lifts the individual, flattens the group | Writers given generative-AI ideas produced stories rated more creative — but the AI-assisted stories were more similar to each other than human-only stories. Individual gain, collective sameness. | Anil Doshi & Oliver Hauser, Science Advances (2024), study of 293 writers + 600 evaluators |
| The pull is toward the mode, not the edges | Researchers describe “mode collapse” and a risk of knowledge collapse — outputs and, over time, shared knowledge narrowing toward a dominant set of ideas as AI use scales. | Survey/empirical work on LLM homogenization and knowledge collapse (2024–2026) |
| People over-trust the consensus machine | The mere knowledge that advice came from an AI causes people to follow it even when it contradicts the context in front of them — classic automation bias. | Reviews of over-reliance and automation bias in AI-assisted decisions (ongoing) |
| The bias outlasts the tool | In a controlled experiment, people working with a biased AI later repeated the same bias on their own, after the AI was removed — they had absorbed it. | Scientific Reports (Nature), study on humans inheriting AI bias (2023) |
Read the table as one message: AI doesn’t just offer you the consensus, it gently trains you into it. It raises the floor — your worst, laziest first draft is now far better — while quietly lowering the ceiling, because the distinctive, against-the-grain answer is precisely the one a “most likely continuation” engine is least likely to surface. None of this is an argument against using AI. It’s an argument for knowing which kind of question you’re asking, because the same tool that should end the debate on a routine question is the worst possible advisor on a question where the whole point is to not sound like everyone else.
Best practices are not the enemy — using them blindly is
It would be a cheap and wrong lesson to conclude “ignore best practices and trust your gut.” Best practices are compressed, hard-won knowledge; ignoring them to feel original is how people relearn painful lessons that a five-minute search would have prevented. Charlie Munger’s warning applies in both directions: the person who reinvents everything from scratch is as foolish as the person who questions nothing. The skill is not picking a side. The skill is diagnosis — correctly reading which situation you’re in before you decide which mode of thinking to spend on it.
Here is the cleanest test. A best practice is trustworthy when three things hold: the environment is stable (the conditions that made the practice work still exist), the practice’s assumptions are visible (you can see why it works, not just that everyone does it), and the cost of being average is acceptable (being one of many doing the same correct thing is fine here). When all three hold, follow the rule — and let AI hand it to you instantly, because spending first-principles effort on a solved, stable, low-stakes problem is its own kind of waste. First-principles thinking earns its higher cost only when at least one of those three breaks: the ground has shifted, the assumptions are hidden or no longer true, or being average is the same as losing.
The First-Principles vs. Best-Practices Decision Matrix
This is the original framework — a way to make the call deliberately instead of by temperament. Two questions set the axes. First: is the situation well-understood and stable, or novel and shifting? Best practices are answers to yesterday’s conditions; their value collapses when conditions change. Second: what does being wrong cost, and can you reverse it? First-principles reasoning is more likely to be wrong on any single try (you’re rebuilding instead of copying), so you want to deploy it where a mistake is survivable or where being merely average is itself the failure.
| Being average is fine / low stakes | Being average = losing, or high stakes | |
|---|---|---|
| Stable & well-understood | 🟢 Take the best practice. This is solved. Let AI give you the consensus and move on — first-principles effort here is wasted motion. (Formatting an invoice, standard contract clauses, boilerplate setup.) | 🟡 Best practice as a floor, then differentiate. Start from the consensus so you don’t relearn known lessons, then reason from first principles on the one or two dimensions where standing out actually matters. (Pricing, positioning, a hire you’ll live with for years.) |
| Novel & shifting | 🟡 First principles, cheaply. No reliable best practice exists yet, but the stakes are low — run a fast first-principles experiment and learn. (A new tool, an unproven channel, a small bet.) | 🔴 First principles, seriously. The consensus is either absent or actively misleading, and being wrong-and-average is the worst outcome. This is where the durable edge — and the real risk — lives. (Building something genuinely new, a strategy in a market that just changed, a contrarian bet you’ll be judged on.) |
The quadrant that matters most in an AI-saturated world is the top-right and bottom-right — the places where everyone now has the same best-practice answer, so the answer can no longer separate you. AI has effectively deleted the advantage of the bottom-left and top-left: those quadrants are commoditized, and you should let the machine own them so your scarce first-principles attention goes where it changes outcomes. The CEO move is to consciously route problems to quadrants instead of treating every problem with the same reflex. The Student move is to keep widening the set of problems you can take to first principles, because that capacity — not prompt fluency — is what compounds.
A five-test checklist for “should I break this rule?”
The matrix tells you which mode fits a situation. This checklist is the in-the-moment version — five fast tests for the specific question “everyone does X; should I?” Breaking a rule well is not rebellion; it’s earning the right to deviate by passing these in order.
| # | Test | Break the rule only if… |
|---|---|---|
| 1 | Do I understand why the rule exists? | …you can state the original problem the rule solves. If you can’t, you’re not breaking the rule from insight — you’re just ignorant of it. (Chesterton’s Fence: don’t remove a fence until you know why it was put up.) |
| 2 | Have the conditions that justified it changed? | …something real has shifted — a new technology, a new constraint, a market that moved — that the rule never accounted for. “It feels outdated” doesn’t count; name the change. |
| 3 | Is being average here the same as losing? | …doing the correct, conventional thing leaves you indistinguishable in a place where distinction is the whole game. If average is fine, keep the rule. |
| 4 | Is the downside survivable? | …a wrong bet is reversible or affordable. First-principles answers fail more often per attempt; break rules where you can recover, not where one miss ends the game. |
| 5 | Would I still do this if AI hadn’t suggested the safe version? | …your reasoning holds on its own. If the only reason you’re conforming is that the model handed you the consensus and it was easy, that’s automation bias, not judgment. |
Notice the asymmetry the checklist enforces. You should reason from first principles far more often than you actually break a rule. Most of the time the honest first-principles answer is “the best practice is right, and now I know why” — which is a better place to stand than blind compliance, because you’ll recognize the day the conditions change and the rule should be dropped. The goal isn’t contrarianism; it’s earned conviction in either direction.
What this means for the next few years
The uncomfortable, freeing implication: as AI makes the consensus answer free and universal, the value of having the consensus answer trends toward zero, and the value of knowing when it’s wrong climbs. We are moving from a world that rewarded knowing the best practice to one that rewards judging the best practice — and those are different skills. Memorizing playbooks was a moat when playbooks were scarce; it is a liability when everyone holds the same one and mistakes it for thinking.
This is why the CEO+Student pairing is the right stance and not a slogan. The CEO half is ownership of the call: deciding, under uncertainty, when to take the safe consensus and when to bet against it — and carrying the downside either way, because a rule broken badly is your fault, not the model’s. The Student half is the unglamorous engine underneath it: doing the slow work of understanding fundamentals so that when you choose to reason from first principles, you actually can — you’re not just guessing with extra steps. AI will keep getting better at handing you the average answer. Your edge is becoming the person who can tell, deliberately and on purpose, when the average answer is exactly wrong.
Frequently asked questions
Isn’t “first-principles thinking” just a Silicon Valley buzzword?
The phrase got fashionable, but the idea is old and concrete. Aristotle defined a first principle as the most basic thing from which something is known; the method is simply: break a problem into things you’re confident are true, and rebuild upward instead of copying an existing answer. It became a buzzword because a few founders credited it for contrarian bets — but the substance is just disciplined reasoning from fundamentals, and it predates the people who made it trendy by about 2,300 years. Treat the hype with suspicion and the method with respect.
If AI gives the consensus, can’t I just ask it to be contrarian or original?
You can, and it helps a little, but understand what you’re getting: a model asked to be “contrarian” produces the consensus version of contrarian — the most likely text that pattern-matches to “edgy take.” That’s still drawn from the middle of the distribution, just a different part of it. Genuine first-principles work requires you to hold the specific, often non-verbal context of your actual situation — constraints, goals, things you know that aren’t written down anywhere — and reason from those. The model can be a sparring partner that pressure-tests your logic, but the originating judgment has to come from you, because only you have the ground truth of your case.
Doesn’t reasoning from first principles waste enormous time?
Yes — which is the whole point of the matrix. First-principles thinking is expensive and you should ration it. Spending it on a solved, stable, low-stakes problem (the top-left quadrant) is a genuine waste, and there AI’s instant consensus is the right tool. The discipline is not “always reason from scratch” — that’s exhausting and foolish. It’s “reason from scratch where it changes the outcome, and take the free consensus everywhere else.” Most people get this backwards: they default to the consensus on the few decisions that deserve original thought, and agonize from first principles over trivia.
How is this different from just “thinking critically”?
Critical thinking is the general habit of not accepting claims at face value. The matrix is narrower and more actionable: it tells you when the effort is worth it. Critical thinking applied to every invoice and email is paralysis; the framework’s job is to route your limited skepticism to the decisions where the consensus is most likely to be wrong or most costly to follow — novel situations, high stakes, and places where being average means losing. It’s critical thinking with a triage rule attached.
What’s the single highest-leverage habit to build here?
Run Test 1 on rules you currently follow without thinking: pick a “best practice” you obey by default and force yourself to state the original problem it solves. You’ll find three kinds. Rules whose reason still holds (keep them, now with conviction). Rules whose reason has expired but everyone still follows out of habit (your opportunity). And rules you couldn’t actually justify at all (a sign you were running on autopilot). Doing this regularly is how you build the diagnostic muscle the AI era rewards — not knowing more answers, but knowing which answers have quietly stopped being true.
Sources
Anil R. Doshi & Oliver P. Hauser. Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content (Science Advances, 2024) — a controlled study of 293 writers and 600 evaluators finding that access to generative-AI story ideas raised the rated creativity of individual stories while making the AI-assisted stories more similar to one another than human-only stories, pointing to a social dilemma between individual gain and collective novelty.
Research on large language model homogenization, mode collapse, and “knowledge collapse” (2024–2026) — a body of empirical and theoretical work documenting that model outputs cluster toward dominant, modal answers and that widespread reliance may, over time, narrow the diversity of shared ideas; cited here for the structural tendency of “most likely continuation” systems to return the consensus.
Study on humans inheriting artificial-intelligence bias (Scientific Reports, Nature, 2023) — a controlled experiment in which participants assisted by a biased AI continued to reproduce the same bias on their own after the AI’s suggestions were removed, illustrating that over-reliance can transfer the model’s tendencies into the human’s later unaided judgment.
Reviews of automation bias and over-reliance in AI-assisted decision-making — research finding that people tend to follow AI recommendations even when those recommendations contradict the contextual information available to them, and that merely labeling advice as AI-generated increases this tendency.
Aristotle, Metaphysics — the classical definition of a first principle as the first basis from which a thing is known, the origin of first-principles reasoning as a method distinct from reasoning by analogy or precedent.
Editorial note: This article is part of CEOtudent’s fully AI-assisted editorial process. The First-Principles vs. Best-Practices Decision Matrix and the five-test checklist are original CEOtudent decision aids — frameworks for routing your own thinking, not empirical claims. The supporting figures and studies are drawn from the publicly available sources listed above and were verified as of June 2026. This is general educational commentary on decision-making and strategy, not professional, legal, or financial advice.
This post is also available in:
