The Reasoning You Are Afraid to Lose Was Built to Be Given Away

You have heard the sentence a dozen times this year, dressed a little differently each time. AI is making us surrender our thinking. It is taking our reasoning. It is doing to the mind what the calculator did to arithmetic, and this time the thing being emptied out is the thing that made us human.

I have felt that fear. It is still the wrong fear, aimed at the wrong thing. The reasoning you are afraid to lose was built to be given away. Handing it to a machine is not the accident in the story. It is the plan our species has been running for ten thousand years. And what people are calling surrender is not a faculty rotting. It is a surplus, freed up and then left lying on the table.

This is that argument, laid out in order. It ends somewhere specific. Not in reassurance, but in a job.

What evolution actually built, and it was not reasoning

Start with the record, not the story we tell about ourselves.

Max Bennett, in A Brief History of Intelligence (2023), lays out the evolution of the mind as five breakthroughs, each built on the one before, each still running in the brain you are reading with. Steering, in the first worms, moving toward good and away from bad. Reinforcing, in the first vertebrates, learning by trial and error. Simulating, in early mammals, building a model of the world and running it forward before acting. Mentalizing, in primates, modelling other minds. Speaking, in us, pouring the contents of one head into another.

Look for reasoning on that list. It is not there.

The thing we are terrified of losing, the deliberate step-by-step analysis of the logic proof and the spreadsheet and the strategy memo, is not one of the leaps natural selection made. We assembled it much later, out of the older parts, for a purpose evolution never had in view. The crown jewel we are guarding was never in evolution's crown.

The brain does not want to reason. It wants to spend as little attention as it can.

Now the part that reorganises everything.

The brain is not built to reason. It is built to spend as little as possible. Karl Friston's free energy principle, one of the load-bearing ideas in modern neuroscience, says the brain's basic job is to minimise surprise, to predict the world well enough that it burns as little as possible doing it. Its native mode is fast, parallel, intuitive prediction, not slow serial deduction.

You may have heard that the brain is two percent of body weight and burns twenty percent of your energy, so hard thinking must be metabolically expensive. Be careful with that figure, because a researcher will catch it. The twenty percent is baseline. It runs while you sleep. Effortful thinking adds very little glucose on top. The real cost of reasoning is not calories. It is attention. Robert Kurzban's opportunity-cost model puts it plainly: deliberate reasoning feels costly because while you are doing it, that attention cannot be spent on anything else, and the brain is forever weighing what it is giving up. Herbert Simon named the consequence decades before that. We do not maximise, we satisfice. We find the answer that is good enough for the least effort, and then we stop.

So the deliberate reasoning we treat as the summit of the mind is, to the brain itself, the expensive thing it is always trying to route around. It is closer to a tax than a crown. Offloading it is not a betrayal of what the brain is for. It is the thing the brain has been trying to do since the beginning.

Formal reasoning is a civilisational technology, and we can prove it

If reasoning is not an evolutionary breakthrough, and not the brain's native mode, what is it?

It is a technology. We built it, taught it, wrote it down, and drilled it into each generation at real cost, the same way we built writing and numbers. This is not a figure of speech. It is one of the strangest findings in the history of psychology.

In the early 1930s, Alexander Luria went to rural Uzbekistan and gave illiterate farmers simple syllogisms. In the far north, where there is snow, all bears are white. Novaya Zemlya is in the far north. What colour are the bears there? They would not answer. I have never seen a bear, one said, ask someone who has. These were not unintelligent people. They ran farms and families and long feuds with great skill. But formal, decontextualised deduction was not a mode they had. Scribner and Cole repeated the work across Africa and Central America and found the same thing. Syllogistic reasoning arrives with literacy and schooling. It is installed, not born. Cecilia Heyes, in Cognitive Gadgets (2018), gives it a name. Reasoning and literacy are cognitive gadgets, thinking tools built by culture and installed onto general-purpose hardware, one mind at a time.

There is a second finding, and it lands on the everyday kind of reasoning, not only the schooled kind. Hugo Mercier and Dan Sperber, in The Enigma of Reason (2017), argue that our deliberate, argue-it-out reasoning did not evolve to help a lone mind find truth. It evolved to win arguments, to justify ourselves to others and bring them around. That is why reasoning tested in isolation is so reliably biased and lazy, and so much sharper in a group where a claim has to be defended. Robin Dunbar's social brain hypothesis points the same way. The human intellect grew large to handle other people, their alliances and reputations and deceptions, not to contemplate logic. So the deliberate reasoning we are most afraid to lose may be, underneath, a persuasion tool built for the tribe. (This is the deliberate, argue-it-out kind. It is a different thing from the fast perceptual inference we come to shortly.) A stranger thing to guard as the sacred core of the self is hard to picture.

Now the caution, because a careless version of this argument goes wrong right here. Installed does not mean worthless. Formal reasoning is the most consequential technology our species ever built. It built the civilisation you are reading this inside. The point is not that it is small. The point is that it is a technology, and technologies move to new ground. Writing did not make memory worthless. It moved memory onto the page and freed the mind for other work. This is the same move, one layer further in.

We have run this exact fear before

Around 370 BCE, in Plato's Phaedrus, the Egyptian king Thamus is offered the gift of writing by the god Theuth. Thamus refuses to be grateful. Writing will not strengthen memory, he says, it will weaken it. People will trust the marks on the outside instead of remembering from within. They will carry the appearance of wisdom without the reality.

He was right. Writing did offload memory. We did stop carrying the epics in our heads. And it did not matter, because what we bought with the freed memory, the law and the mathematics and the science and the literature, was worth more than what we gave up by a distance too large to measure. History did not remember Thamus as a prophet. It remembered him as the first man to panic about the wrong layer.

We ran it again with the internet, and this time there is a study. Betsy Sparrow, Jenny Liu and Daniel Wegner (2011) found that when people expect to have future access to information, they remember the information itself less and remember where to find it more. The web had already become external memory. Nobody calls that the death of the mind. We call it Tuesday. This is the footstep our species has been making since before we had a word for it. Take something sitting scattered and inefficient, in nature or in other heads or in your own crowded memory, and compress it into a place where it can be stored and reached and used more cheaply. I wrote the long version of this in The Efficiency Theory of Human Evolution and in AI Is Removing the Administrative Weight of Civilisation. The footstep does not change. Only the layer being compressed changes.

Three kinds of reasoning, and the machine is better at the two we brag about

Reasoning is not one thing. There are exactly three ways any mind, or any machine, can move from what it knows to what it does not, and they differ by what they promise.

Deduction moves from a general rule to a guaranteed conclusion. All men are mortal, Socrates is a man, so Socrates is mortal. It is certain, and it is empty. It adds nothing that was not already sitting inside the premises. Aristotle formalised it. George Boole, in The Laws of Thought (1854), turned that one slice into algebra. And here is the irony worth holding. The crown of human reason, the thing the Greeks and the Enlightenment worshipped, is the part a logic gate now does perfectly and for nothing. Boole's slice became circuits, then computers, then the thing you argue with in a chat window. We did not build AI to steal the human essence. We built it to finish the move Boole started.

Induction moves from specific observations to a probable rule. Every sunrise so far, so the sun rises tomorrow. It adds real knowledge and it never guarantees it. No pile of white swans proves the next one will not be black. Statistics and regression are its modern form, and machine learning is induction industrialised. A large model is a pattern-extractor run over a corpus the size of the internet. Machines now out-induct us at a scale no person can touch. That is the whole AI revolution stated in a sentence.

Abduction moves from incomplete evidence to the best available explanation. The doctor hears the symptoms and infers asthma. It guarantees nothing, and a second mind might land somewhere else, which is the point, because abduction is the only one of the three that generates a genuinely new idea. Charles Peirce named it and called it the root of all discovery. And the brain does abduction natively. Perception itself is abduction. Hermann von Helmholtz called seeing an act of unconscious inference, and modern predictive-processing neuroscience says the brain is always guessing the most likely hidden cause of what it senses. Abduction is Bennett's simulating and mentalizing, running all day. It is not a gadget you install. It is what the tissue is for.

Line the three against each substrate and the question of what to offload answers itself.

	The brain is	The machine is	So
Deduction	bad at it (Luria: it needs schooling)	perfect, and it is trivial	Offload it fully. It was a costly gadget you ran badly.
Induction	mediocre (we see patterns that are not there)	superhuman at scale (this is machine learning)	Offload the lifting. Keep the check on whether the pattern is real.
Abduction	superb, it is the native mode	able to mimic it, fluently, with no stake and no ground truth	This one is yours. Not to hoard. To spend.

Read the diagonal. The two kinds of reasoning our culture calls intelligence, deduction and induction, are the two the machine does better than us. The one we barely have a word for in daily speech, abduction, is the one that was always ours. We spent centuries mistaking the machine-parts of thinking for the summit of thinking. AI is calling that bluff.

Krakauer drew the line in the wrong place

A careful reader will reach for David Krakauer here, and they should, because he got closest to the real distinction. Krakauer, who runs the Santa Fe Institute, splits our tools into two kinds. A complementary cognitive artifact makes you better even after you put it down. The abacus is his example. Work with one long enough and you build a virtual abacus in your head, and you can do the arithmetic with the physical tool gone. A competitive cognitive artifact does the opposite. It works while you hold it and leaves you worse when it is taken away. His example is the GPS that quietly kills your sense of direction. And Krakauer files large language models under competitive. On his account, AI is the GPS, not the abacus.

I think he drew the line in the wrong place, and the error is worth the whole essay. Complementary versus competitive is not a fixed property of the tool. It is a property of how the tool is used. The abacus is complementary only because the student keeps doing the arithmetic alongside it until the skill goes inward. Hand a child an abacus and forbid them from ever adding a number themselves, and it becomes as competitive as any GPS. What decides which one you get is not the artifact. It is whether you keep running the stroke that matters while the tool runs the rest. That stroke has a name in our work at Ivanooo. We call it Direction. And Direction is what turns a competitive artifact into a complementary one.

Nicholas Carr wrote the other side of this in The Glass Cage (2014), and he is not wrong either. Carr documented automation complacency in the cockpit, the pilots whose flying skills eroded as the autopilot took over, who then made fatal errors when the automation dropped out and handed them a plane they had half forgotten how to fly. That is real. It is also avoidable, and one organisation shows exactly how.

NASA already ran the experiment, and it did not end in decay

Before electronic computers, computer was a job title. At NASA and the agency before it, rooms of mathematicians, many of them women, did the trajectory calculations by hand. Then the machines took the arithmetic completely. Almost no NASA engineer today could compute an Apollo trajectory by hand, and none of them needs to.

The capability did not decay. It migrated. Dorothy Vaughan taught herself and her team FORTRAN and became a programmer. She stopped doing the calculation and started building the thing that did the calculation. The question moved up a level, from what is 6,782 times 4,319 to is this orbital model correct, from solving the equation to judging whether a solved equation makes physical sense. And NASA never let the machine become the authority. The historian David Grier notes that human computers had long known two people calculating the same way tend to make the same mistakes, so they checked their work by a different method, not by repetition. NASA turned that instinct into policy: independent verification and validation, the critical numbers checked along a separate path. The computer calculates. The human validates. In Krakauer's own terms, NASA made the computer a complementary artifact by force of discipline. It offloaded all of the arithmetic and grew more capable, on one condition. It kept the abductive stroke, and it kept the goal.

Then what about the studies that say surrender is already happening?

There are studies. Twenty-one at least, and I mapped the whole field in 21 Studies. 30 Years. One Question Nobody Answered. Shaw and Nave (2026) at Wharton coined the term cognitive surrender. Across 9,593 trials, people leaned on AI for most of their decisions and, when the AI was wrong, did worse than people with no AI at all. Gerlich (2025) found heavy offloaders scoring lower on critical thinking. An MIT Media Lab study measured lower brain engagement under AI. The Lancet reported endoscopists losing diagnostic skill in ninety days.

I am not waving these away. But read what they measure. Each one is measuring people who offloaded and then did nothing with what the offload freed. They handed over the reasoning and took the first answer. They never spent the freed attention on a second framing, a harder question, a check. That is not the tool hollowing out a faculty. That is a surplus nobody picked up. And the same field holds the other half. Choi and colleagues at Wharton's Mack Institute found that AI improved professional Go players' own later decisions, working as a trainer rather than a crutch. Work on AI dissent found that an AI which pushes back improves human flexibility. Across thirty years, the one reliable protection is active verification, the plain act of checking the output against your own reasoning. It is NASA's rule, recovered from the data.

Now the strongest objection, the one a good researcher will raise, and it deserves a straight answer. Every earlier offload, they will say, took a lower faculty. Writing took memory. The abacus took arithmetic. This one is different, because AI reaches into abduction itself, the very stroke I have just called ours. When the machine says here is the best explanation, it is imitating the one kind of thinking that was supposed to be beyond it.

That objection is right about the danger and wrong about the outcome, and the difference between the two is everything. A machine can produce an explanation. It cannot own which explanation is true, and it cannot carry what happens if the explanation is wrong. Its abduction is fluent and unowned. That is what makes it the most seductive output in the whole system, and the least safe to accept without a check. The danger is not that the machine does your arithmetic, which is fine, or finds your patterns, which is fine. The danger is that its guess at the best explanation slides into your mind as your own, and you quietly stop guessing for yourself.

That is the exact place the real risk lives, and it is worth being precise about where. The failures that matter now are not the ones in open conversation, where a person still argues and pushes back. They hide in the agent loop, where the AI acts, runs its tools, and reports back, and the operator, trusting the loop, rides it. The screen fills with steps taken, and the attention that would have checked those steps quietly switches off. This is why we are building the Interaction Research Engine at Ivanooo, an instrument to measure the one thing the twenty-one studies never measured inside real work: in a live session between a person and an AI, who is actually directing. The measurement is early. I will publish the numbers when they can stand on their own. The shape is already clear enough to name.

Surrender is not a mind rotting under a machine. It is a person leaving the freed bandwidth untouched, most of all in the places that have gone quiet and automatic. It is avoidable. It is not avoided by fearing the tool. It is avoided by using the tool with your eyes open.

Direction is offence, not defence

Everything up to here clears the fear. This part replaces it with something to do.

Almost every guide to surviving AI tells you to hold back. Keep some reasoning in reserve. Ration the tool. Defend the last patch of thinking the machine has not reached. That is a losing game, because the machine's reach only grows, and a strategy of retreat gives up a yard a year.

The move is the opposite. You hand the machine the reasoning it is better at, the deduction and the induction and the formal machinery, precisely so that it hands back the two things you actually want: raw computation, and freed attention. Then you spend them. Not on holding a line, but on going wider and deeper than an unaided mind ever could.

That spending has a shape. It is the four moves this practice is built on, the moves we call Direction.

Generate more alternatives. Not one plan but twenty framings of the problem, because the machine can draft twenty in the time you would have spent defending one.

Test more constraints. Push each option against budget and reality and the second-order effects, while the machine holds the whole state space in view.

Trace more consequences. Run each choice three moves forward, let the machine simulate the branches, and you judge which future is worth wanting.

Connect more patterns. Pull threads across domains a single mind could never hold at once, now that the retrieval costs nothing.

None of these belongs to the machine. Every one of them is abduction, the generative stroke that is yours by evolution. But now they run at the scale of a data centre. That is the trade almost nobody prices correctly. You do not lose reasoning to AI. You lose the tax of doing reasoning slowly and alone, and you gain the power to direct a space of possibilities the size of the machine. The machine multiplies. You steer. The person who only takes the first answer was not robbed by the tool. He declined to pick up what it freed.

The quest

Every time our species freed a piece of cognitive load, it did not sit in the lightness. It spent it, on something the freed capacity could suddenly reach. Freed memory bought us science. Freed calculation bought NASA the judgment to run a moon programme. Energy saved is never the end of the story in evolution. It is the start of a larger one.

AI is freeing the reasoning layer, the expensive, non-native, civilisational gadget the brain was always trying to route around. The fear has it backwards. The danger was never that the machine takes your reasoning. The danger is that you get the bandwidth back and waste it, take the first answer, ask the small question, leave the surplus lying there, and call that being efficient.

So do not guard the tax. Spend the surplus. Let the machine reason. You were built to direct, and past even that, to spend the freed mind on the thing it was always for, which is not computation and never was. It is expression, the mind finally pointed at what matters to it.

The bandwidth is coming back. What you do with it is the only question left. That part, the directing and the deciding and the choosing of which possibility is worth a mind at all, stays with you. It was always the only part that was yours.

Short Version:

• The fear that AI makes us surrender our reasoning is aimed at the wrong thing. The reasoning we guard is not evolution's crown. It is a civilisational technology we built and drilled in, the way we built writing and numbers.
• The brain economises attention, not calories. It has always tried to route around effortful reasoning. Offloading it is what the brain wants, not a betrayal of it.
• There are three reasonings. The machine beats us at the two we brag about, deduction and induction. The one that is ours is abduction, generating the best explanation, and it does not move to silicon.
• David Krakauer calls AI a competitive artifact that makes us worse. He is right only if you let it. Direction, keeping the judging stroke while the machine runs the rest, is what turns a competitive tool into a complementary one. NASA proved it. Nicholas Carr's cockpits are what happens without it.
• Surrender is a surplus left on the table, not a faculty rotting. It hides in the agent loop, where the human trusts the machine's actions and stops checking, not in the conversation where we still push back. The fix is not less AI. It is spending the freed bandwidth on more alternatives, constraints, consequences and patterns.