In 1932, the Cambridge psychologist Frederic Bartlett handed English students a Native American folk tale called “The War of the Ghosts” and asked them to read it twice. Later he asked them to write it out from memory. Then he handed those reproductions to other students, who read them and wrote their own versions, and so on down a chain. The story shrank with every pass, falling from roughly three hundred and thirty words to under two hundred. It also got more English. Canoes became boats. The supernatural elements, which did not fit what the students expected a story to do, were smoothed into something familiar or dropped entirely. None of them were lying. Each person reported what they remembered, honestly and with confidence, and what they remembered was the story bent toward what they already knew.

Bartlett’s students had days between the reading and the retelling, which makes his study easy to file under memory, though the same compression runs in real time inside a single conversation with no delay at all.

Boaz Keysar, a cognitive scientist at the University of Chicago who spends his career on how people actually use language rather than how grammar says they should, built a deceptively simple test of this. Two people sit on opposite sides of a grid of cubbyholes holding various objects. Some cubbies are blocked from the speaker’s view, so the listener can see objects the speaker cannot. When the speaker says “move the small candle,” and the listener can see a candle even smaller than the one the speaker means, the listener’s eyes go to the smaller one, and often their hand does too. They know the speaker cannot see it. They know it, and they reach for it anyway. Adults asked about it afterward can describe the speaker’s limited view perfectly well. In the moment, they interpret from their own position first and correct for the speaker’s position late, if they correct at all.

Multiply that by a conference table and a fiscal year, and you have most of what people call politics.

Why does everyone leave the same meeting with a different version of it?

Over the last two decades I have worked across departments, cultures, stakeholders, boards, crews, and every gradation of thought variability a person can encounter, and what recurs is that most conflict is a product of comprehension asymmetry. I use that word deliberately, since asymmetry says nothing about who is correct. It describes two understandings sitting in different positions relative to the same object, each one complete and coherent to the person holding it, each one missing precisely what the other one has.

The oldest version of this story is the Tower of Babel, where the project fails not because anyone opposed it but because the workers stopped being able to transmit to each other. We treat “lost in translation” as a light phrase, something you say about subtitles, when it actually describes the mechanics of nearly all corporate, project, and institutional friction. Some people do play badly on purpose, and always will, but that population is small, and blaming it for the daily grinding of an organization gives it far more credit than it has earned. The larger culprit is fractional agreement and comprehension that diffuses as it travels.

I have run meetings this way for as long as I have been running meetings. Somebody finishes explaining something and I ask whether everyone understood what they said, and I get a yes, because you always get a yes. Then I pick one person at random and ask them to explain it back. Then another. Then I turn to whoever originally said it and ask whether that was what they meant. We all sat in the same room and heard the same words in the same order, and almost always the versions do not line up. Sometimes the drift is minor. Sometimes it is severe enough that two people were about to go spend a month building opposite things. I do this internally and I do it with clients, and it can be annoying for everyone involved, mine included. The more I sat with what kept coming back, the more I wanted to understand why something so consistent goes so unnoticed, which is what produced this article. A bit of behavioral psychology explains why it happens, and a bit of probability math explains why it gets so much worse than anyone expects.

Going to introduce some math here. It might look scary for those who are not familiar with equations, but I will walk you through it, and it matters because the numbers show something the words alone cannot.

Think about what has to happen for a group to genuinely agree on something. Not for one person to get it, but for all of them to get it, and to get the same version of it. Say the odds of one person understanding a point correctly are eight in ten. For two people to both get it, you do not average the two eights. You multiply, because both things have to happen together, and eight tenths times eight tenths gives you sixty-four hundredths. Two people at eighty percent each land at sixty-four percent together. Add a third and it drops to fifty-one. Every person added multiplies the odds down again, the same way a chain holds only if every link holds.

So write c for the chance that any one person understands a given point correctly, and n for the number of people in the room. What you get is:

A ∝ cⁿ

You read this as: real agreement is proportional to one person’s accuracy multiplied by itself once for every person present. Six people at eighty-five percent each, which is a generous estimate for most meetings anywhere, is 0.85 multiplied by itself six times. That comes to about thirty-eight percent. Two out of three points do not survive the walk to the elevator, and nobody notices, because the failure is silent and the calendar has already moved on to the next thing.

Takeaway: Agreement is not evidence of comprehension. It is evidence that the listener has assembled something coherent from what they heard. Coherence and accuracy are different achievements.

Why does half an understanding feel like a whole one?

Partial understanding survives in intelligent organizations full of careful people because it does not feel partial. It feels like understanding. No internal alarm fires when you have absorbed sixty percent of a plan, because that sixty percent forms a complete-seeming picture on its own. Your mind fills the missing pieces with defaults drawn from experience, and those defaults feel exactly like the parts that were actually said. A half-built bridge announces itself. A half-built understanding does not.

Elizabeth Newton, then a psychology doctoral student at Stanford, designed an experiment in 1990 to measure the distance between knowing something and conveying it. Participants tapped out the rhythm of well-known songs on a table while a listener tried to name the tune. The tappers predicted listeners would identify about half. Listeners got around two and a half percent. Inside the tapper’s head the melody plays in full, with words and arrangement and tempo, while the finger moves, and the listener receives a series of knocks on wood. The tappers were not showing off or overestimating themselves in any general way. They were unable to subtract a song they could hear from a performance they were producing, which is the condition of every executive explaining a strategy they have lived inside for two years to a group hearing it for eleven minutes.

Herbert Clark, a Stanford psycholinguist who spent decades arguing that conversation is a joint activity rather than a transfer of packets between two heads, documented how people actually establish shared understanding. They use the cheapest evidence available. Heads move. Someone says mhm. Nobody looks confused. Verifying comprehension properly costs time and mild social friction, so speakers accept the thinnest proof that lets the conversation continue. It works well enough in ordering coffee. It fails badly when the thing being transmitted is a plan that six departments will spend a quarter executing.

Does talking longer make two people understand each other better?

Two people alone in a room produce the same divergence, because a conversation is not a series of separate deliveries. Each sentence gets decoded using the understanding assembled from the sentences before it. Whatever you got slightly wrong in the first five minutes is not a single error sitting quietly in the corner. It becomes the lens through which the next forty minutes are read.

Picture a handover. Tom is giving Sara a client account, forty minutes at one desk, no politics involved, both of them competent and paying full attention. Early on Tom says the client has been difficult about scope, meaning the client keeps adding requests after sign-off. Sara hears difficult about scope and pictures a client who argues over what was promised, because that is the version she has met four times before in her career. Small gap. Nobody would catch it, including her.

Twelve minutes later Tom mentions that the last project ran long. Sara now files this under the client disputing deliverables, which is the frame she is holding, rather than under the client expanding them, which is what happened. The second fact did not arrive neutral. It arrived through the first one. Around minute twenty Tom says the relationship needs firmer boundaries, and Sara understands this as being firmer about defending the original agreement, while Tom meant being firmer about charging for additions. Both are boundaries. They are opposite behaviors.

By minute thirty-five Sara is building a plan on top of three interpretations that each felt obvious to her, and Tom is answering her questions with increasing confidence because her questions sound exactly like the questions of someone who understands. She leaves the room fully briefed. Six weeks later she has a client who feels accused of something he did not do and a colleague wondering why she ignored the handover.

Neither of them was careless, and the mechanics are the same multiplication as before, except what gets counted is no longer people but moments. For Tom and Sara to still be talking about the same account at minute forty, every meaningful exchange between minute one and minute forty has to have landed, because each one shapes how the next is heard. Every exchange is another link in the chain, another coin that has to come up heads.

So let m be the number of meaningful statements passed between two people, with c again standing for how accurately any one of them is received:

U ∝ cᵐ

You read this as: what two people still share is one exchange’s accuracy multiplied by itself once for every exchange they have had. Since c is always a fraction below one, more exchanges push the product down rather than up. Twenty exchanges at ninety-five percent accuracy each leave about thirty-six percent held in common. An hour of talking can produce less alignment than fifteen minutes of talking, unless somebody stops to check, which nobody does because the conversation is going so well.

Imagine this. A pilot leaves Addis Ababa for Washington, one degree off the direction he intended to fly. Nothing looks wrong from the cockpit, since he is still pointed at America. Eleven thousand kilometers later he is about two hundred kilometers off course, hunting for a runway over the wrong stretch of the Atlantic, which is what a fixed error does. It grows in a straight line, in proportion to how far you fly.

What happens between Tom and Sara runs worse than that, because each misread statement re-aims the aircraft before the next one arrives, so Sara’s second error is measured from her first error rather than from Tom’s original direction, and the third is measured from the second. Small turns compound into large ones. You take off for Washington and land in Tokyo, and at no point did the plane tilt into a turn hard enough for anyone on board to feel it. A crack in a windscreen behaves the same way, since it does not simply lengthen, it branches, and every branch throws off branches of its own.

Both people are doing this at once, in different directions, on different points, which is where the asymmetry comes from. Tom is building a model of what Sara now knows, and that model is drifting too, so he keeps talking on the assumption that ground has been covered while she keeps listening on the assumption that the ground meant something else. Neither one is failing. Both are executing a perfectly reasonable process against material that has already deformed.

Takeaway: More talking does not produce more alignment. Past a certain length it produces more surface for divergence to attach to. Only correction produces alignment, and correction has to be asked for out loud.

Why do the best departments fight the hardest?

The expensive part in a group is not the average error but the spread of the errors, and every person’s version differs from every other person’s version, with the number of those relationships growing much faster than the number of people. Six people produce fifteen pairs who can disagree. Ten people produce forty-five. Adding four bodies to a meeting triples the surface area where two versions can grind against each other, which is why a decision that survived a small group intact starts falling apart the moment it is socialized more broadly.

Most managers misjudge their own teams by watching the average and ignoring the spread. A group where everyone understood seventy percent of the plan in the same way will outperform a group where everyone understood ninety percent of it differently. The second group looks better on any measure of individual comprehension and coordinates worse, because there is no shared subset to build on. Kahneman, Sibony and Sunstein found the same thing hiding inside insurance firms in Noise, where underwriters pricing the same file produced quotes that differed by roughly five times what their own executives predicted. Each individual number looked entirely reasonable on its own, and nobody had ever put them side by side.

Departmental friction lives here, which explains why it is so often worst between competent departments rather than incompetent ones. Finance heard “growth” and encoded margin. Marketing heard “growth” and encoded reach. Operations heard “growth” and encoded volume, and started hiring against it. All three are working hard, all three are working correctly against the version they hold, and every hour of that work carries them further from each other. The divergence compounds because each department then builds a second layer of decisions on top of its own first-layer interpretation, and a third on top of that.

What turns a misunderstanding into a grudge?

When two people discover their versions do not match, almost nobody concludes that the transmission was lossy. Lee Ross, the Stanford social psychologist who spent much of his career studying why negotiating parties fail to settle disputes that should be easy to settle, named the tendency behind this naive realism. I experience my perception as a direct view of things as they are, so when you see it differently, the explanation has to be located in you. You were not paying attention. You are protecting your budget. You are pushing your agenda. Ross found this reflex in people who were otherwise generous and fair-minded, appearing automatically in anyone whose own perspective is invisible to them as a perspective.

Once divergence gets attributed to intent, coordination becomes a status contest, and correcting the misunderstanding now requires somebody to concede in front of colleagues, which has a price. So people defend versions they only half hold, argue for positions they arrived at by accident, and build territory around interpretations that were never decisions to begin with. The ego flare-up everyone remembers from that meeting was not caused by ego. Ego arrived second, as a reasonable response to being accused of bad faith over a difference that started as a knock on wood.

Takeaway: Most conflict attributed to personality is comprehension divergence that nobody caught early enough to correct cheaply.

What survives the trip from the top of an organization to the bottom?

Understanding spreads out across people, thins over the length of a conversation, and also weakens with distance from the source. Engineers have a word for how much of an original survives being copied or moved: fidelity. A recording has high fidelity when it sounds like the performance that produced it. A translation has high fidelity when it means what the original sentence meant. Scan a photograph, then scan the scan, then scan that, and each generation carries less of the original than the one before.

Decisions lose it the same way, and for the third time the same multiplication does the work. Each level of an organization keeps most of what it was given and passes that along, so the loss at level four is taken out of what survived level three, not out of the original. Losses do not add up. They stack on top of each other.

With x standing for the number of levels a decision passes through:

Fidelity ∝ cˣ

You read this as: what arrives at the bottom is one level’s accuracy multiplied by itself once for every level it crossed. Four levels at ninety percent each deliver sixty-six percent. Four levels at eighty percent deliver forty-one percent, which means the person executing is working from less than half the original and has no way of knowing it. They are not being careless. They received a photocopy of a photocopy of a photocopy, and they are working faithfully from the smudge in front of them.

Instructions survive the chain because they are short and concrete and easy to repeat, while reasoning does not, being long and contingent and feeling optional to whoever is passing it along. So the number arrives intact and the logic behind the number evaporates, which is precisely backwards. Reasoning is the part that lets a person adapt correctly when reality diverges from the plan, and reality always diverges from the plan. A team holding the instruction without the logic executes beautifully in the wrong direction.

This is the same fade rate that governs brand equity and skill atrophy in earlier pieces here. Understanding decays on a curve, and the curve is steeper when the original transmission was already compressed.

Why is repeating yourself the only thing that works?

Claude Shannon, a mathematician at Bell Labs ( also the dude Claude is named after), settled the engineering version of this in 1948 when he founded information theory more or less single-handedly. To move a message across a noisy channel without corruption, you spend part of your capacity on redundancy: repetition, checksums, error-correcting codes. Shannon showed that no encoding scheme escapes this tax regardless of how clever it is, and every reliable system built since, from deep space transmission to the phone in your pocket, pays it constantly and invisibly.

Organizations refuse this deal for a simple accounting reason. Redundancy is visible and immediate and costs eight minutes at the end of a meeting, while corruption is invisible and delayed, and when it surfaces three months later it wears the costume of two directors who cannot get along.

What should you actually check?

1. Run the readback on schedule, not when you suspect a problem. Pick someone at random, ask them to explain it back, then check against the person who said it. The moments you least suspect are the ones where nobody thought to ask.

2. Correct mid-conversation, not at the end. Since every minute of talking lands on the understanding built by the previous ones, a check at minute fifty arrives after fifty minutes of accumulated drift. Two short confirmations inside a long discussion are worth more than one thorough one at the close.

3. Attach a number and a date to every decision. Vague verbs are where fractional errors hide most comfortably. “Soon,” “prioritize,” and “scale up” mean different things to finance and to operations, and both will believe they were given the same instruction.

4. Send the reasoning with the instruction. Whoever sits four levels downstream needs the logic more than you do, because they are the one who will meet the situation you did not anticipate. If only one of the two can survive the trip, protect the reasoning.

5. When two versions collide, spend the first thirty seconds on what each person heard. Before anyone defends what they believe. Establishing that the mismatch came from transmission rather than motive is what keeps a coordination problem from converting into a grudge, and grudges have no expiry date.

Write it down once, in one place, and point everyone at it. A single canonical version does not stop misreading, but it collapses the number of competing originals from n to one, which is the cheapest reduction in variance available to any organization.

Can any of this actually be fixed?

Every one of those five improves the odds and not one of them closes the gap, which I should be honest about before anyone goes off and runs them expecting silence afterward.

Physics has a rule that energy is neither created nor destroyed, only moved from one form into another. Heat becomes motion, motion becomes sound, sound dissipates back into heat, and the ledger always balances no matter how the accounting is arranged. Misunderstanding behaves in a way worth naming the same: neither created nor destroyed, only relocated. Ask for the plan back and you have not eliminated the confusion, you have shifted it from a place where it would have surfaced in September to a place where it surfaces in the room, at a price you can afford. Write it down and the ambiguity leaves the conversation and enters the document, where a different reader will find a different gap in it. Add a number and a date and someone will now be certain about the wrong thing on schedule.

The accounting is still worth doing, since a misunderstanding caught in a meeting costs eight minutes and the same one caught in September costs a quarter, and the difference between a well-run organization and a badly run one is not the amount of confusion inside it but where the confusion gets paid for. Misunderstanding is so permanent a feature of human life because language was never a transmission medium to begin with. It is a set of instructions for reconstruction, handed to a listener whose materials are not yours. What they build will resemble what you meant. It will never be the same object.

So the goal is not comprehension but a shorter distance between the moment two versions diverge and the moment somebody notices. The most functional teams I have worked with were not the ones with the smartest people in them. They were the ones willing to look slightly slow on purpose, repeating things back to each other like air traffic controllers, absorbing a small tax at the front so they would not pay an enormous one later.

STRATEX by Naz publishes on the intersection of cognitive science, behavioral research, strategy, and market behavior.