PSYCHOLOGY · BEHAVIORAL ECONOMICS · SYSTEMS · DECISIONS
The Sandbox Behind Your Eyes: Why No Two People See the Same Reality
Your brain does not record the world. It simulates one. And the smartest people are not the ones running the truest simulation.
Your brain does not record the world. It simulates one. And the smartest people are not the ones running the truest simulation.
In 1973 the psychologists William Chase and Herbert Simon sat chess grandmasters and ordinary players in front of a board. Five seconds of looking. Then they took the board away and asked each person to rebuild the position from memory.
The grandmasters rebuilt it almost perfectly. The novices managed a handful of pieces.
That is the part everyone remembers. Here is the part that matters more.
When Chase and Simon scattered the same pieces into random arrangements, positions that could never occur in a real game, the grandmasters collapsed to novice level. Same pieces. Same five seconds. The advantage evaporated.
The grandmaster was never seeing pieces. He was seeing meaning. Strip the meaning out and the master disappears with it.
Same board. Two different worlds.
The Gap Is Not What You Think
This is the rule, not the exception. Put any two people in front of the same fact, the same chart, the same crashing market, and they will build different internal pictures of it. We like to explain that gap with intelligence. One person is sharper, so they see more clearly.
That explanation is wrong, or at least it is shallow.
The gap is not horsepower.
It is the parameters of the simulation each brain is quietly running behind the eyes. Different size. Different resolution. Different rules for when to stop. Change those parameters and you change the reality the person lives in, even when the facts in front of them are identical.
What a Mental Model Actually Is
Start with the machinery, because the field spent eighty years confusing itself about it.
In 1943 a Cambridge psychologist named Kenneth Craik wrote that the brain carries a small-scale model of reality and runs it to predict what happens next. Decades later Philip Johnson-Laird narrowed the idea to reasoning. We do not solve logic with hidden rules, he argued. We build a little simulation of the possibilities and check it for counterexamples. Donald Norman used the same two words for the unstable, half-wrong pictures people build of their devices. Peter Senge used them for the slow, deep assumptions that filter how we read an organization.
For a long time these camps talked past each other. Modern cognitive science reconciles them with one frame.
The brain is sealed inside the skull. It never touches the world directly. All it gets is a stream of noisy signals, and to make sense of them it builds a guess. It generates a model from memory, projects that model downward as a prediction, and compares the prediction to what arrives. When they match, the model stands. When they clash, the mismatch travels back up and edits the model.
Perceiving is predicting.
A mental model is the simulation you run to keep that mismatch low. There is one equation worth holding onto.
In plain words: what you perceive is your prior, the stored expectation of how the world tends to go, multiplied by the evidence currently hitting your senses. Evidence alone does not produce perception. It is always run through a prior you did not choose and usually cannot see.
Two people with different priors, fed identical evidence, compute different perceptions. Not because one is lying. Because the math forces it.
Call the result a sandbox. A small simulated world, built from priors, run on a tiny workbench, trusted because it is cheaper than reality. The differences between people are not vague vibes about personality. They are specific settings on that sandbox.
The Parameters That Actually Differ
Five settings do most of the work. Each one has survived the kind of scrutiny most psychology has not, which I will get to. For now, watch how each one quietly forks two people’s view of the same scene.
The Size of the Sandbox
Working memory is the workbench where the simulation runs, and it is small. Most people hold somewhere between four and seven items at once before the structure buckles. That ceiling decides how many moving parts, how many branches, a model can carry at the same time.
Give a complicated conditional contract to someone with a small workbench and they build one likely outcome and stop. Give it to someone with a larger one and they hold the main outcome plus three edge cases in parallel, refusing to commit early. Same document. One person sees a path. The other sees a decision tree.
Expertise cheats this limit. The grandmaster from the opening does not have extra memory slots. He has bigger chunks. Years of exposure let him compress twenty pieces into two familiar shapes, so two slots carry what would cost a novice twenty. A radiologist does the same with a chest film. The novice processes every shadow one at a time and runs out of room. The expert sees one pattern and has working memory left over to ask the next question.
Capacity is not the whole story. It is the size of the table everything else has to fit on.
The Resolution
For decades we sorted people into visual and verbal thinkers. That split is dead. Visual cognition is not one thing. It is two systems fighting over the same scarce resource.
One stream handles object imagery: color, texture, the high-definition picture. The other handles spatial imagery: relationships, movement, transformation. Holding rich pictorial detail crowds out the room you need to rotate and manipulate, so the brain trades one against the other. People drift into being object visualizers or spatial visualizers, and the two profiles read the same image differently.
Show both a graph of velocity over time. The object visualizer sees a literal picture and reads the rising line as a physical slope, a hill going up. The spatial visualizer throws away the picture and treats the line as an abstract vector, which is exactly what lets them recover the math underneath.
An architect and a theoretical physicist are both “visual.” Their models are nearly alien to each other.
The Manager
A model is only as good as the system deciding whether to trust it. That job belongs to what Keith Stanovich calls the reflective mind, and people differ enormously in how readily it interrupts.
The cheapest test of it is one question. A bat and a ball cost a dollar and ten cents. The bat costs a dollar more than the ball. How much is the ball?
The answer that arrives first is ten cents.
It is also wrong. Ten cents plus a dollar ten is a dollar twenty. The real answer is five cents, and getting there means catching the intuitive answer, distrusting it, and spending the effort to check.
Some people’s manager almost never interrupts the cheap answer. Others reliably stop and audit. This disposition, captured by the Cognitive Reflection Test, is strikingly stable across years, with test-retest correlations around 0.75 to 0.80, and it predicts who falls for a long list of classic reasoning traps. It is not the same thing as being good at math. It is the habit of overriding your own first draft.
Watch the managers diverge.
A disease shows up in one person in ten thousand. The test for it wrongly flags healthy people five percent of the time. A result comes back positive. The doctor whose manager stays asleep matches “positive” to “sick” and models a ninety-five percent chance of disease. The doctor whose manager wakes up runs the count instead. Out of ten thousand people, one is truly sick and tests positive, but five hundred healthy people also test positive. The real chance this patient is sick is under one percent. Same result on the same test. Two completely different models of the patient in the room.
The Firmware
Capacity and disposition feel like personal hardware. The next setting is installed by culture.
Work by Richard Nisbett and colleagues mapped a deep split in how attention itself gets allocated. Analytic cognition, more common in Western samples, pulls the focal object out of its background and explains events through the object’s own traits. Holistic cognition, more common in East Asian samples, attends to the whole field and explains events through relationships and context.
Drop a market crash in front of both. The analytic observer isolates the failing company and builds a model around the chief executive: greed, incompetence, bad calls. Fix the person, fix the problem. The holistic observer widens out to interest rates, cyclical patterns, the ground shifting under the entire sector, and is comfortable holding the contradiction that the same executive was brilliant last decade and doomed this one.
Neither is hallucinating. Culture writes the firmware.
That firmware decided, before any conscious thought begins, what even counts as a cause.
The Rule for When to Stop
Every simulation needs a stop condition. When do you decide your model is good enough and quit updating? That threshold is set by what you believe knowledge even is.
Some people carry a naive picture: knowledge is certain, simple, and handed down by authority. Others carry a sophisticated one: knowledge is tentative, tangled, and built by reasoning. Marlene Schommer showed these beliefs act as hidden settings that govern how long you keep revising when contradictions arrive.
Hand both a pair of medical studies that disagree. The naive reader concludes one of them must be a lie, because real knowledge cannot contradict itself, and closes the case. The sophisticated reader builds a larger model where the disagreement is expected, a normal product of different methods and incomplete data, and keeps the question open.
One brain stopped early. The other refused to.
Half the Famous Parameters Are Not Real
This is where most writing about the mind goes soft, so I will be blunt. A lot of the differences people quote with confidence did not survive the last fifteen years.
Psychology ran a brutal audit on itself starting around 2011, re-running famous experiments at scale.
Some collapsed.
Ego depletion, the idea that willpower is a finite fuel tank that runs dry with use, was everywhere. Then a coordinated replication across twenty-three labs and more than two thousand people found an effect of essentially zero, around . Construal level theory, the tidy claim that psychological distance controls how abstractly you think, fails high-powered direct replications and shows heavy fingerprints of publication bias. The headline magnitudes of stereotype threat in real testing shrink toward nothing once you correct for which studies got published.
So I am not going to dress up dead findings as live parameters.
What survives the audit is the spine of this piece. The Cognitive Reflection Test holds up across years and labs. The object-spatial split is backed by both factor analysis and brain anatomy. And the strangest, most stubborn result of all still has to be dealt with.
One Machine, Five Dials
Step back and the five settings stop looking like a list. They are dials on one device.
Working memory is the size of the sandbox. The object-spatial profile is its resolution. Epistemic belief is the stop condition. The reflective mind is the manager deciding whether to spend energy on a better simulation or coast on the cheap one. Culture is the firmware that chose, before you could vote, what your model would treat as a cause.
One machine. Five dials.
None of this is you seeing reality. It is you running the cheapest simulation that keeps surprise low enough to act on.
The sandbox feels like a window because you have never once been outside it.
That last line is not there for comfort. It is the setup for the part that should bother you.
Intelligence Is Not a Defense
Here is the result that ends the comfortable story.
Almost every cognitive bias gets weaker as you get smarter. Working memory, reasoning power, reflective disposition, all of them help you catch your own errors. There is one exception, and it happens to be the bias that drives most real disagreement.
Myside bias is the tendency to test evidence in whatever way protects what you already believe. Keith Stanovich and his colleagues found that it is almost completely decoupled from intelligence. High-IQ people show as much of it as anyone. Among the highly numerate it can be worse, because a sharper mind is a better lawyer.
A sharper mind is a better lawyer.
Give two brilliant people the same neutral dataset on a politically charged question and they do not converge. They use their horsepower to build rival, internally airtight rationalizations and walk away more certain than before. More facts did not close the gap.
More facts widened it.
This is why “just educate people” keeps failing on the questions that matter most to identity. There, intelligence is not a brake. It is an accelerant.
And you cannot easily audit your own model from the inside, because of a second trap. Rozenblit and Keil named it the illusion of explanatory depth. Ask people how well they understand how a toilet, a zipper, or a market works and they rate themselves high. Ask them to actually explain the mechanism, step by step, and the rating crashes. They confused being able to look the thing up, or name its parts, with carrying a working model of it.
Labels, not blueprints.
So the people most certain they see clearly are often the ones running the shallowest simulation behind the most confident face, with a first-rate defense attorney on retainer.
The One-Line Version
You never looked through a window. Neither did anyone you ever disagreed with.
Sources
The model-of-reality idea traces to Craik (1943), with the reasoning version from Johnson-Laird (1983) and the interface and systems readings from Norman and from Senge. The chess work is Chase and Simon (1973). Working-memory capacity is Baddeley and Hitch (1974). The Cognitive Reflection Test is Frederick (2005), set inside the tripartite model of Stanovich (2011), whose work with West and Toplak established the decoupling of myside bias from intelligence. Analytic versus holistic cognition is Nisbett, Peng, and Norenzayan (2001). The object-spatial split is Kozhevnikov and Blazhenkova (2009). Epistemic beliefs are Schommer (1990). The illusion of explanatory depth is Rozenblit and Keil (2002), and the curse of knowledge is Camerer, Loewenstein, and Weber (1989). The predictive-processing frame draws on Clark and Friston. On replication, the ego-depletion null is the multi-lab report led by Hagger and colleagues (2016), and the corrected stereotype-threat estimates are from Flore and Wicherts (2014).