Essay 3

What Actually Helps You Think Better

The honest answer about debiasing — what works, what doesn't, and what works only sometimes.

May 15, 2026

Somewhere, on a calendar near you, is a session called Cognitive Biases in Decision-Making. There is a slide deck with a Wikipedia screenshot, a quiz about whether you are a fast or slow thinker, a story about how the housing crash happened because everyone was anchored on Greenspan, and a takeaway invitation to be aware of your blind spots. People nod. People take notes. Two weeks later, the same people who attended will be back to overestimating how long their project takes, hiring the candidate who looks like them, and arguing for the conclusion they came in believing.

I am not making this up to be unkind. There is now a fairly large literature on what these workshops accomplish, and what they accomplish is most honestly described as: a little, in the lab, sometimes, on tasks similar to the ones used in training, with significant decay over weeks. In practice, out in the office, the hospital, the cockpit, the family dinner, the effect is roughly nil.

The popular framing of this finding has a name, and it’s the rare academic phrase that’s actually pretty good. It’s called the G.I. Joe Fallacy, after the old American cartoon whose tag line was “and knowing is half the battle.” The finding is that, for cognitive biases, knowing is almost none of the battle. Knowing that you have a bias and not having that bias turn out to be very different states. People who can correctly define confirmation bias on a multiple-choice exam are not measurably better at avoiding it the next afternoon.

This is the place where most popular writing on biases gets shy, because the next sentence (so what actually does help?) is harder to answer in a way that sells a book. But there is a real answer. It’s small in places, modest in others, and oddly satisfying overall. The condition of getting better at thinking isn’t trying harder. It’s redesigning the situation that’s doing your thinking for you.

The honest empirical truth

Let me give you the numbers first, because they are clearer than the prose around them usually is.

A major meta-analysis of educational debiasing interventions pooled 54 randomized controlled trials covering 10,941 participants. Across all those studies, the average effect of an educational intervention on subsequent biased judgment was a Hedges’ g of about 0.26, with a confidence interval running from roughly 0.14 to 0.39.

If those numbers don’t mean anything to you, that’s normal. The translation: an effect of g ≈ 0.26 is small, but real. It is not “negligible.” Calling it negligible, as some recent writing on this topic does, is sloppy. A small effect averaged over thousands of people is enough to move group-level outcomes in a measurable way. It is also not transformative. The same people, on the same kinds of tasks, will still get a lot of the same kinds of things wrong.

The bigger problem isn’t the size of the lab effect. It’s transfer. A more recent systematic review of debiasing studies asked a more practical question: when an educational intervention works on the training task, does it generalize to other tasks that involve the same bias? Mostly, it didn’t. What people seemed to be learning was a context-specific trick (on this kind of word problem, don’t reach for the obvious answer) rather than a change in how they reason more broadly. The trick fades within weeks. The reasoning underneath is intact.

This is the part to internalize: the failure isn’t of effort or attention. It’s of architecture. Biases run on the part of the mind that operates in milliseconds, without your permission, before any conscious effort can come online. Your understanding of biases lives in the slower, deliberate part of the mind. Knowledge sits in one room; bias does its work in another. Educational programs are like writing a sternly worded letter from the first room to the second. The second room doesn’t read mail.

This is not a counsel of despair. The same primary research that produces this honest finding also produces a list of the things that do work, and most of them work by skipping the writing-a-letter step entirely. They restructure the second room. Or they wait until both rooms are awake, which is rarer than people think, and intervene in that narrow window.

What actually does help, divided into three categories

1. Redesign the situation, not the person

The single most reliable way to reduce a bias is to remove the input that triggers it. This is unromantic. It feels like cheating. It works.

Examples that have stood up to repeated study:

Blind review. Strip identity information from a candidate’s application or a manuscript’s authorship before it’s evaluated. The reviewers’ similarity bias and prestige bias can no longer fire on data they can’t see. Major US orchestras introduced blind auditions in the 1970s and 80s, with candidates playing behind a screen, and the share of women advancing through early rounds rose substantially (a 2000 study by Goldin and Rouse remains the canonical analysis, though the precise magnitude has been re-examined since). The reviewers didn’t get less biased. The information that was activating the bias was hidden from them.
Default change. Most people, most of the time, take whatever option is pre-selected. Switching the default for organ donation, retirement savings, vaccination, and dozens of other choices has produced some of the largest behavioral effects in the empirical literature, on the order of 30 to 50 percentage points of change with no other intervention. The people aren’t different. The architecture is.
Pre-commitment to criteria. Decide before you look at any options what the evaluation criteria are. Write them down. Score against them. The criteria can’t be unconsciously bent to favor the option you already prefer, because by the time you meet the options, the criteria are already locked in.
Mandatory deliberation periods. A built-in pause between hearing a proposal and deciding on it interrupts the Expedience family of biases, which run on speed. The pause is structural, not motivational; no one has to remember to slow down.

These interventions have something in common. None of them ask the person to be smarter, more virtuous, or more vigilant. They change what the person sees and when. The bias still exists. It just has nothing to bite.

2. Targeted, repeated practice — not awareness lectures

The educational interventions that do show meaningful effects look almost nothing like the corporate workshop. They tend to be:

Specific to a single bias rather than a survey of many
Game-based or simulation-based rather than lecture-based
Repeated with feedback so people see, in real time, when their intuition was wrong
Practiced against a wide variety of cases so the lesson generalizes beyond a narrow context

In one well-documented line of research — Carey Morewedge and colleagues’ 2015 work on game-based debiasing — a serious-game training on confirmation bias and other reasoning errors produced effects that were larger than typical educational interventions and that persisted for at least two months. The hard part is the repetition with feedback. The brain learns from being wrong in ways it does not learn from being told it might be wrong.

The general principle: the slower, deliberate part of the mind can train the faster, automatic part, but only the way you train any other physical or cognitive skill, through repetition with feedback over time. The one-day workshop tries to do this with one rep and no feedback. It is, mechanically, like trying to learn to ride a bike from a lecture.

3. Process scaffolds that compensate for what individual judgment can’t deliver

The third category is the one most practical for adult life, because you can install these into your own decisions without anyone else’s permission.

Reference-class forecasting. When estimating how long a project will take, how much it will cost, or how it’s likely to end, start from the historical record of comparable projects, not from the specifics of yours. The specifics make every project feel uniquely promising. The historical record corrects this in a way deliberate willpower can’t.
The pre-mortem. Before a plan is committed to, get the team to imagine, vividly, that the plan has already failed and explain why. The reframe is called prospective hindsight, and it works because it bypasses the social pressure that suppresses doubt during normal planning. The original research suggests it identifies meaningfully more risks than conventional risk assessment, though the specific size of that effect has been reported under several different numbers in popular writing and the underlying study is older than the figure suggests.
The decision journal. Before you know how a decision turned out, write down what you decided, why, and how confident you were. After the outcome is known, return to the record. Without this, hindsight bias silently rewrites your past beliefs to match what happened, and you “learn” nothing while feeling like you have. With it, the discrepancy between what you expected and what occurred becomes visible, and the visibility is the entire mechanism.
Consider-the-opposite. Before committing to a conclusion, explicitly list two or three reasons it might be wrong. This is the cheapest of all the interventions on this page; it can be done in thirty seconds; it has shown small but reliable effects across many studies. It works because the brain’s default is to gather support for whatever is already on the table, and the prompt redirects that effort.
Structured analytic techniques. Intelligence agencies, which have to make consequential judgments from incomplete information under deadline, have spent decades codifying methods that try to compensate for the analyst’s biases: analysis of competing hypotheses, key assumptions check, devil’s advocate. These are organizational versions of the individual techniques above, and the principle is the same. The structure does the work the lone analyst’s vigilance can’t.

None of these is large in effect on its own. The largest of them moves outcomes by a few percentage points in controlled studies. Put together, in a workflow that uses several at once, they move outcomes a lot more, but you have to put them together. The practitioner question, throughout this whole field, is one of portfolio assembly, not silver bullets.

A note on the Stoic ancestors of all this

The two interventions in the third category that feel most demanding (the pre-mortem and the decision journal) both have direct ancestors in Stoic practice, and the practice is more than two thousand years older than the modern research.

The Roman philosopher Seneca, writing letters to a friend in the 60s AD, described a practice he called premeditatio malorum, the premeditation of evils. The instruction was to mentally rehearse, in detail, the ways a planned course of action could go wrong: the trip canceled, the friend who fails you, the illness, the financial reversal. The reason for the exercise wasn’t pessimism. It was that the imagination, faced with future events, naturally tends toward the rosy version of them, and the exercise was a deliberate counterweight. The pre-mortem, in Gary Klein’s twenty-first-century formulation, is the same practice with the labels updated.

The Stoic emperor Marcus Aurelius kept a personal notebook for the last decade of his life. We have it under the title Meditations. It was never meant for publication. Read it from one angle and it’s reflective philosophy. Read it from another and it’s the most thoroughly documented decision journal in human history: a man writing down what he believes, why, and how confident he is, returning to the entries, and confronting himself when the world disagreed.

Neither Seneca nor Marcus had a meta-analysis. They had careful attention, sustained over years, by people who treated their own minds as something to be examined. Modern psychology has produced a more rigorous test of these practices, and the verdict, broadly, is that the careful people were right.

Browse the interventions

Below is the set of interventions discussed above, organized so you can filter by SEEDS family (pick the kind of bias you’re trying to interrupt and see which interventions target it) and see for each one its approximate effect size, how well it transfers out of the lab, what it costs to implement, and the situations where it earns its keep. The point of this tool is not to find the one big lever. There isn’t one. The point is to see, at a glance, which combination of small levers fits a specific decision you’re facing.

Effect sizes are from the primary empirical literature where available, with a note when the underlying citation is older or thinner than the popular version suggests. None of these tools is dramatic on its own; the skill is portfolio assembly. Open the microsite full-screen →

What this whole essay is and isn’t saying

It is not saying that knowledge of biases is useless. Specific, targeted training has small but real effects, and for some people the framework of which bias is at work here is genuinely useful as a diagnostic.

It is not saying that you should give up on your own judgment. Most of the time, in most environments, your judgment is fine. The previous essay was about how to tell when it isn’t.

It is saying that the popular story (learn about biases, become aware, override them with willpower) is roughly the wrong shape for how cognition actually changes. The faster mind that produces biases does not take instructions from the slower mind that knows about them. The leverage is somewhere else.

And it is saying that the question of what to do about all this is more practical and less dramatic than most of the writing on this subject suggests. There is no transformation. There are a few small, reliable, slightly boring practices, mostly ones that change the situation around you rather than the wiring inside you, that compound over years into measurably better judgment. The Stoics knew this. The behavioral scientists are finishing the homework.

The next essay zooms out from interventions specifically to the broader cognitive activity all of us are doing all the time, mostly without noticing: forecasting. Not the geopolitical kind. Yours. About how long your tasks will take, how your partner will feel about the news, whether the new job will make you happy. There’s a whole literature on this that almost nobody outside academia has heard about, and what it says is more useful and more humbling than the geopolitical-superforecaster genre.