Learn it → Apply it → Save as Skill
Most people who inflict serious harm while “just following orders” are not sadists, and most do not believe what they're doing is right. They comply because of how the situation is built: a credible-seeming authority, a request that escalates in small steps, and enough distance from the person being harmed that responsibility starts to feel like someone else's. Stanley Milgram showed this could be produced reliably, in a single afternoon, in ordinary volunteers — and the full record, including a genuinely messy replication and criticism history, is covered honestly below rather than reduced to one famous number.
Dr. Wallace's Favorite Mental Models is designed to turn useful ways of thinking into reusable tools. Each model gives you a fast visual overview, deeper interactive learning, practical application exercises and an AI-ready Markdown skill you can save for future use.
Instead of merely reading about mental models, you can gradually build your own AI skills library: a collection of portable reasoning skills that can help ChatGPT, Claude, Grok and other AI assistants apply the mental models you value.
Obedience to authority is the tendency to carry out a harmful or ethically questionable instruction from someone perceived as legitimately in charge, even when it conflicts with your own judgment and even when refusing would cost you little. It depends on structure, not character: a credible-seeming authority, an escalation built from small steps, and distance from whoever is actually affected all move the rate of compliance far more than any trait of the person complying.
Important: this is not a claim that most people are secretly cruel, and it is not a legal or moral excuse for harmful compliance. Later replication and criticism show the exact rate depends heavily on setup — the honest version is covered in full below, not skipped.
The pattern shows up wherever a legitimate-seeming authority issues an instruction that escalates gradually, and the person carrying it out is distanced — physically or socially — from whoever is actually affected.
First demonstrated by psychologist Stanley Milgram at Yale between 1961 and 1963, and run in dozens of variations since — with a genuinely mixed replication and criticism record covered honestly further down this page.
Someone perceived as legitimately in charge — a title, a uniform, a lab coat, an official process — gives the instruction. Perceived legitimacy drives compliance, not actual expertise.
The request grows in increments too small to trigger a clear "I need to stop" moment. Each step looks like a small, consistent continuation of the one before it.
Physical or social distance from whoever is actually affected — a locked door, a phone call, a different department, a spreadsheet instead of a face — makes the harm feel abstract.
Responsibility feels displaced onto the authority ("I was just following orders, policy, or the process"), and having already complied with step N makes stopping at N+1 feel like admitting the earlier steps were wrong too.
Answer about recent, real situations, not how you'd like to answer. This stays in your browser; nothing is sent anywhere or scored by AI.
1. I've gone along with an instruction from someone senior that I privately thought was wrong, telling myself it was their call, not mine.
2. In my organization, a request that grows in small increments over time is harder to object to than the same request made all at once, up front.
3. I've seen someone comply with an instruction mainly because of a title, badge, or letterhead behind it, without checking whether the instruction itself actually made sense.
4. When something goes wrong after everyone "followed the process," it's hard to point to any single person who feels clearly responsible.
5. In meetings, a lone objection to a plan usually gets dropped quickly unless at least one other person backs it up out loud.
6. The further removed I am from who's actually affected by a decision — a different team, a different country, a spreadsheet instead of a face — the easier it is for me to approve it.
Open a use case to see what the model changes.
Watch for a chain of individually-defensible-seeming steps that, taken together, cross a line no one would have agreed to if it had been proposed in one piece at the start. That incremental structure, not a single evil decision, is usually how organizational scandals actually unfold.
In a 1966 field study, Charles Hofling had an unfamiliar caller claiming to be a doctor instruct hospital nurses by phone to give a patient twice the labeled maximum dose of a fictitious drug, in violation of explicit hospital policy. 21 of 22 nurses were prepared to comply — even though, beforehand, nearly every nurse and nursing student asked predicted they themselves would refuse.
Courts since the Nuremberg trials have treated this as a documented psychological pattern worth designing institutional safeguards around — clear channels to refuse unlawful orders, independent review — not as a legal or moral excuse that settles the question on its own.
A founder's instruction to skip a safety check, fudge a metric, or mislead a customer "just this once" becomes easier to normalize once one or two similar requests have already been granted, and dissent risks looking like you're not a "team player."
Aggressive targets set by distant management can produce compliance with scripts or practices employees would reject if asked about them directly, precisely because the instruction arrives as "the policy" or "the target," not as a personal request from a visible, answerable person.
The clearest sign is content, not comfort: are you complying mainly because of who's asking, rather than because you've actually thought it through and agree? That question is often enough to interrupt the pull of "I've already gone this far."
This is an illustrative example designed to teach the method; the company and figures are for demonstration only.
Expand for the original research, how the numbers actually varied, and an honest look at the replication and criticism record.
Stanley Milgram recruited ordinary adults by newspaper ad to play a "teacher" who administered what they believed were real, increasingly painful electric shocks to a "learner" (an actor, unharmed throughout) for wrong answers on a word-pairing task, up to a labeled maximum of 450 volts. A lab-coated experimenter prompted hesitant participants with a standardized sequence of verbal prods. In the best-known baseline version, 26 of 40 participants (65%) continued all the way to 450 volts. 150 volts — the point where the "learner" first demanded to be released — was the most common place participants tried to stop; of those who pushed past it anyway, Milgram's own data show roughly 79% continued all the way to the maximum.
Milgram ran dozens of variations across more than 700 participants, and the obedience rate swung enormously by condition. Moving the "teacher" into the same room as the "learner" cut compliance to roughly 40%. Requiring the teacher to physically force the learner's hand onto the shock plate cut it further, to roughly 30%. Giving instructions by phone instead of in person cut it to about 20%. Having two confederate "teachers" refuse partway through cut compliance to roughly 10%. The common thread across all of these: physical and social distance, and the visible presence or absence of peer dissent, moved the numbers far more than any plausible measure of individual character.
Full replication to 450 volts is no longer considered ethical. Jerry Burger's 2009 partial replication stopped participants at 150 volts — the original's most common refusal point — for ethical reasons, and found an obedience rate "not significantly different" from Milgram's original rate at that same checkpoint, suggesting the core pattern still holds up within current ethical limits. A widely covered 2010 French television documentary restaged a similar setup as a fake game-show pilot in front of a live studio audience; roughly four in five of 80 contestants continued administering apparent shocks toward the apparent maximum. At the same time, psychologist Gina Perry's archival research (Behind the Shock Machine, 2013) found that some experimenters deviated from the standardized script, that debriefing was inconsistent — one disobedient participant said decades later he was never told the shocks were fake — and that the famous "65%" figure describes only one of many very different conditions: the overall rate across all 700-plus participants in all of Milgram's variations ranged from near zero to over 90% depending on setup. The responsible summary: the phenomenon is real and has held up reasonably well within ethical replication limits, but the single "65%" number is often repeated as if it were a fixed, universal human tendency rather than one data point from one condition among dozens that produced very different results.
The rate ranged from near zero to over 90% across Milgram's own variations depending on structural factors like distance and peer dissent, not fixed individual character.
It's a documented psychological pattern, not a legal or moral excuse; courts since the Nuremberg trials have explicitly rejected it as a complete defense.
The Hofling hospital study found 21 of 22 trained nurses were prepared to violate explicit policy and give a dangerous overdose on an unfamiliar caller's phone instruction.
Training that only appeals to values, without reducing distance, slowing escalation, or requiring a visible second sign-off, skips the levers the research shows actually move compliance.
Labeling someone's self-interested or independently motivated bad behavior as "just obedience" misapplies a model that was never meant to erase personal responsibility.
A single confederate modeling refusal dropped compliance from 65% toward roughly 10% in Milgram's own data — one visible objection changes outcomes more than most ethics posters do.
the person had clear independent benefit or motive — personal gain, fraud for its own sake — rather than reluctant compliance under pressure. That's self-interest, not obedience.
no one in a position of legitimate-seeming authority was actually involved. A lone actor pursuing their own initiative isn't this specific pattern, even if the outcome is similar.
citing a specific number, like "65% of people would...", as settled science. The honest range across conditions is enormous, and the real figure depends heavily on the exact setup.
This browser tool structures your thinking. It does not send your information anywhere or perform AI analysis.
Save as Skill converts the model into portable Markdown instructions that teach an AI when and how to apply Obedience to Authority, what questions to ask, what evidence to seek, what mistakes to avoid, and how to structure its analysis.
Save the skills you find useful and organize them into your own reusable library. The Markdown is intentionally model-neutral, so it can be used wherever an AI system accepts uploaded files, project knowledge, custom instructions or reusable text context. Platform capabilities differ, so this is a portable skill file rather than a claim of one-click installation.
No. That's the core misreading, and an important one to correct. Milgram's participants were an ordinary cross-section of adults who answered a newspaper ad, and available follow-up assessment found no unusual sadistic traits among them. The variable that changed behavior was the structure of the situation — distance, escalation, legitimacy, diffused responsibility — not who happened to walk through the door that day.
Groupthink is about a group of peers converging on a decision and suppressing their own dissent, with no one giving a direct order. The Bystander Effect is about failing to act when responsibility is diffused among onlookers, with no instruction involved at all. Obedience to Authority is specifically about actively complying with a direct instruction from someone perceived as legitimately in charge.
Cite it as a well-documented, partially-replicated pattern whose exact size depends heavily on setup, not as a single fixed percentage that applies to all people in all situations. Favor structural fixes — reducing distance from consequences, slowing escalation, requiring an independent second check, and making dissent visible and easy — which the data support regardless of exactly how the number moves from study to study.