49Know Thyself

Stephen M. Fleming, Know Thyself: The Science of Self-Awareness, Basic Books, 2021

Linked to 40 chapters: supports 7 · contradicts 6 · mechanism 33 · same claim 4 · pairing 4

Fleming runs a laboratory at University College London devoted to one question: how the brain knows what it knows. He opens with two versions of a medical appointment. In the first, a doctor recommends a heart bypass and, asked why she is confident, walks through her reasoning and what being wrong would entail. In the second an algorithm makes the same recommendation, the doctor cannot say why, and the only argument for surgery is that the system has been accurate before. What separates them is metacognition, the capacity to think about one's own thinking, which Linnaeus was taken enough with in 1735 to annotate his entry for the genus Homo with a single Latin line meaning those that know themselves. Fleming's argument is that this capacity is not a mystery standing outside science but a set of algorithms with a biology, a developmental timetable, a catalogue of characteristic failures, and a ruler by which it can be measured. The book has two parts. The first builds self-awareness out of machinery the brain needed for other purposes: its tracking of uncertainty, its monitoring of its own actions, the recursion it borrowed from thinking about other minds, and the stretch of prefrontal cortex where these meet. The second asks what reflection buys, in learning, in deciding, in working with others, in explaining ourselves, and in dealing with machines that do none of it. Two claims carry the whole. Confidence is a construction rather than a readout of accuracy, which is why it can be moved in the laboratory without touching performance. And the ability to judge one's own performance is partly independent of the performance itself: two people can be equally good at a task and differ widely in whether they know when they have got it right.

How to be uncertain

The book begins with Stanislav Petrov, the Soviet lieutenant colonel watching the early-warning satellites one morning in September 1983 when the alarms reported five American missiles inbound, twenty-five minutes from detonation. Petrov judged a satellite fault more probable than a surprise first strike and reported a malfunction; the satellites had mistaken sunlight off cloud tops for missiles. Holding a shade of grey where the system offered black and white is, for Fleming, the ordinary business of the brain. Sealed in a dark skull, it receives noisy, low-resolution samples and must run the causal chain backwards to guess what produced them. This is an inverse problem, and it is what perception is.

Fleming teaches the mathematics with three dice. Two are ordinary; the third is a trick die showing either 0 or 3 on every face. He rolls all three behind a curtain, gives you the total, and asks what the trick die was showing. Roll a thousand times and the counts form two hills peaking around 7 and 10, and Bayes's rule says answer 3 for totals of 9 and above and 0 for 8 and below. The interesting part is the middle. Where the hills are level, at 7 or 8, the evidence supports neither answer, and the machinery that works this out has already worked out how likely the answer is to be wrong. Estimates of uncertainty are needed to perceive at all; a grading of one's own answers arrives with them at no extra cost, and the distinction between decisions about the world and decisions about the accuracy of those decisions is the one signal-detection theory calls type 1 and type 2.

Perception works the same way, combining prior expectations and incoming data in inverse proportion to how uncertain each is — batter poured into a mould, in Fleming's picture, thick batter keeping its own shape and runny batter taking the mould's. Because uncertainty tracking is this basic, animals have it. J. David Smith trained a bottlenose dolphin, Natua, to classify tones as high or low, then added a third lever that skipped the trial; Natua pressed it when the tone fell in the ambiguous range, and when certain swam to his answer so fast that his bow wave soaked the switches. Macaques take a risky bet over a safe one more often when they are right, and transfer the behaviour to a memory test without retraining, which rules out the objection that they have learned a stimulus-response rule. Louise Goupil and Sid Kouider showed eighteen-month-olds a toy hidden in one of two boxes and then removed it; the infants persisted in searching when they had chosen the right box and asked their mothers for help when they had not.

The second building block comes from the other end of the system. Correcting an action requires a prediction of what should have happened. The cerebellum, which holds some sixty-nine of the brain's eighty-five billion neurons, is thought to receive a copy of each motor command, generate the sensory consequences it expects, and correct the discrepancy. Engineers call this a forward model, and it explains why expected sensation is damped: you cannot tickle yourself, and the recipient of a punch feels it harder than the thrower.

When a deviation is too large for quiet correction, it surfaces. Patrick Rabbitt, in the 1960s, gave people a button to press when they caught themselves erring and found they pressed it about forty milliseconds faster than their fastest responses to an external signal, proving the detection was internal. William Gehring's 1992 thesis found the signature: a wave arriving less than a hundred milliseconds after an error, generated in the dorsal anterior cingulate cortex, now called the error-related negativity. Twelve-month-olds produce it when they fail to look at a face that would have won them music and lights. Fleming reads it as a prediction error of the kind Wolfram Schultz found in dopamine neurons, which stop responding to juice once a light reliably predicts it: an internal reward signal that grades performance when no one else is marking it. These two blocks make up what Fleming calls implicit metacognition, and he separates it from the explicit kind with an aircraft — the autopilot monitors the plane, the pilots monitor the autopilot — except that here both sit in one head.

Knowing me, knowing you

Why should any of it become conscious? Fleming's answer follows Gilbert Ryle, who held that we learn about ourselves by the same methods and in the same terms that we learn about other people. With Nathaniel Daw he built this into a second-order model: self-reflection runs the machinery of social cognition on different inputs, so self-awareness and mindreading should appear together, depend on each other, and share tissue.

The developmental record supports it. Children under four fail the false-belief test, saying Maxi will look for his chocolate where it actually is rather than where he last saw it. In Simona Ghetti's laboratory, three-, four- and five-year-olds remembered the same number of pictures, but only the older two groups gave higher confidence to the answers that were right. Three-year-olds shown a sponge painted as a rock claim they knew it was a sponge all along; five-year-olds recognise that their first impression was wrong. Mindreading at four predicts self-awareness later even with language controlled, and thinking about what someone else feels interferes with judging one's own performance while leaving the performance alone. Both are shaped by the company children keep: Samoan children, raised where discussing one's thoughts is impolite, pass mindreading tests at around eight. Fleming and Sarah-Jayne Blakemore's group tested twenty-eight volunteers aged eleven to seventeen and found that age did not predict how well they could pick the brighter of two patches but did predict how well their confidence tracked whether they had picked it.

To measure any of this, Fleming turns Donald Rumsfeld's taxonomy inward, from known knowns to the things we do not know we do not know. Cross confidence with accuracy and four cells appear. High confidence when right and low confidence when wrong are the good cells, and the proportion falling into them is metacognitive sensitivity. It is distinct from metacognitive bias, the overall level of confidence, and the distinction does real work: a person can be overconfident on average and still register every individual error. Statistical models fit to confidence ratings put numbers on it, under names such as meta-d′ and Φ. Surveys find people rating themselves above average at driving, at work and at intelligence — of the American university professors in one study, ninety-four per cent judged their teaching above average — while remaining able to notice the particular occasion on which they got it wrong.

The brain work grew out of a lunchtime conversation in Queen Square in July 2008 between Fleming and the neurologist Rimona Weil. Volunteers judged which of two flashed images held a brighter patch, with difficulty adjusted trial by trial so that everyone performed at the same level, rated their confidence after each decision, and later returned to be scanned. People with better metacognition had more grey matter in the frontal pole, at the very front of the prefrontal cortex, and greater white-matter integrity in the tracts running to and from it. Katsuki Nakamura's laboratory in Japan reported a functional counterpart the same year, and Hakwan Lau replicated the structural result at Columbia. Fleming is candid that the sample was probably too small for the effect size, a caution he extends to imaging of individual differences generally. Convergent evidence comes from the cortical midline, where medial prefrontal cortex and the precuneus activate both when people judge whether adjectives apply to them and when they think about others, and where the meta-analysis Fleming ran with Anthony Vaccaro found the two overlapping.

The strongest evidence that self-evaluation is a faculty of its own comes from damage. Arthur Shimamura, working with amnesic patients in the 1980s, found a subgroup who could not rate their confidence in answers they could not recall, and they were the ones with Korsakoff's syndrome, whose damage reaches the frontal lobe. In 1989 he closed the argument with a double dissociation: temporal-lobe damage impairing memory while sparing metacognition, frontal damage impairing metacognition while relatively sparing memory. The separation from intelligence is of the same kind: in a sample of nearly a thousand, IQ predicted task performance and not metacognitive sensitivity, and patients whose tumours had been removed from anterior prefrontal cortex had normal IQ and impaired sensitivity.

Loss of this faculty has its own names: anosognosia to neurologists, lack of insight to psychiatrists, long assumed to be denial and now understood as damage to the machinery itself. Fleming describes patient LM, sixty-seven and recently retired, paralysed down her left side by a stroke and cheerfully certain she could clap her hands, waving her good hand at an imaginary partner. Henry Marsh's version of the difficulty is that a person with frontal-lobe damage rarely has insight into it, because the self doing the comparing has nothing left to compare itself with. The same pattern makes Alzheimer's crueller than memory loss alone, since a patient unaware of the loss will not write the list, take the medication or stop driving — and metacognition is still not part of a standard neuropsychological assessment. LM's case ends better than it begins. Katerina Fotopoulou filmed an examination and played it back. Watching herself, LM said, "I have not been very realistic," corrected her account of her own body within minutes, and still held the corrected view six months later.

Avoiding self-awareness failure

Before turning to uses, Fleming extracts three lessons, and they are the load-bearing ones.

The first is that the capacity to reflect and the accuracy of the resulting picture are different things, and accuracy is harder won here than in perception. Perception stays tethered because the world corrects it constantly: misjudge the size of a cup and you knock it over. Self-assessment has only a loose and slow feedback loop, so its illusions survive; the person who talks over everyone in a meeting collects raised eyebrows, not a spilled drink. Fluency, the ease with which anything is processed, is the main source of those illusions. Companies with pronounceable names are valued more highly on the stock market than companies with awkward ones. People feel more confident about decisions made quickly, and about words printed in a large font, in neither case for any good reason.

The second is that self-awareness is absent more often than we suppose, because it is engaged on a need-to-know basis. If the lower levels are coping, no signal propagates upward, exactly as a problem a junior solves never reaches the corner office. Laboratory estimates of how much of the time attention has drifted off the task run from fifteen to fifty per cent, and the drift is commoner when the task is well practised. A study at the Weizmann Institute found that making a task harder and faster raised activity in parietal, premotor and visual regions while lowering it in the prefrontal network that introspection engages. Stress does the same by another route: people who released more cortisol in a social stress test had the most impaired metacognition, and a dose of hydrocortisone lowered metacognitive sensitivity against placebo. The consequence is uncomfortable — reflection fails at the moments when it would pay most.

The third is that metacognition is a cause and not a commentary. Repeat a pair of words during study and people feel they have learned it, and decline to study it again. Raise the brightness of both the signal and the noise in a perceptual decision and the difficulty is unchanged while confidence rises; people then seek less information and change their minds less often. Fleming's parallel is the Coriolis illusion, in which a pilot in cloud feels banked while the instruments correctly read level, and correcting the felt bank is what puts the aircraft into trouble. Student pilots are taught to trust the instruments, and his advice is the same: distrust a self-assessment that disagrees with what colleagues report.

Learning to learn

The classroom holds most of the applied evidence, and Fleming's target is the assumption, running from John Leslie's 1817 arithmetic primer through Dickens's Gradgrind, that education is the loading of facts. Metacognition enters learning at three points: in judgments about what one knows and how to study it, in the confidence attached to that knowledge in use, and in the decision after answering about whether to revise.

At the first point the illusions are well documented. More than nine in ten teachers believe in matching teaching to a pupil's visual, auditory or kinaesthetic style, and controlled studies find no link between stated style and performance. A study of fifty-two students explains why the belief persists: preferred style did not predict how well they learned from pictures or words, but it did predict how confident they felt. Seventy undergraduates at Haifa read leaflets on screen and on paper, scored alike when study time was fixed, and felt more confident about the screen; when they could stop when they liked, the screen readers stopped earlier and scored 63 per cent against 72. Nate Kornell found ninety per cent of students performing better after spaced practice than after cramming while seventy-two per cent reported cramming the better method, an illusion he likens to leaving the gym pleased after choosing weights too light to do anything. Testing oneself beats rereading, and feels worse. What works is intervening on the judgment rather than the material: Patricia Chen's group at Stanford gave one set of students the usual reminder that an exam was a week away and another the same reminder plus an exercise asking which resources they would use and how; the second group scored about a third of a letter grade higher, 3.6 and 4.2 per cent in two experiments, and reported less anxiety.

At the second point, metacognition has been scored for decades without anyone saying so. Until 2016 the SAT offered five answers, awarded one point for a correct one and deducted a quarter for a wrong one, so that blind guessing averaged zero and skipping paid whenever confidence was low. A student with good metacognition dodged the questions she would have failed; one with poor metacognition and good ability collected the penalties. The British Civil Service does it deliberately, asking candidates to rate their own performance during its entrance examinations and counting the ratings. A study running from ages seven to twelve found weak links between metacognition and IQ at any one moment but found early metacognition predicting later intelligence, which is what one would expect if knowing what you do not know is what steers learning.

At the third point sit beliefs about the self. Albert Bandura's self-efficacy, close to what Fleming calls metacognitive bias, moves performance rather than reflecting it: children's beliefs about their mathematics at nine predicted their performance at twelve with ability controlled. Low confidence can be the useful setting, since knowing that memory will fail is what prompts the list, and in an experiment by Fleming and his student Xiao Hu people learning word pairs paid a small cost to file the ones they doubted, precisely when confidence was low. The remedy he favours throughout is the third-person view. Undergraduates told they would have to teach the material studied it better, and an eight-minute exercise in which pupils gave study advice to younger pupils raised their own grades over the rest of the year. Rakefet Ackerman and Asher Koriat found the reversal cleanly: judging themselves, students read short study time as mastery; judging another student over a video link, they correctly read long study time as the sign of it.

Decisions about decisions

Fleming opens with Mark Lynas, who had spent years destroying experimental crops and then stood up at the Oxford Farming Conference in 2013 to say he had been wrong about genetic modification, and who described the recantation as changing sides in a war. Changes of mind on committed positions are rare, and he wants to know what governs them.

The account is the cake mould again: the more confident the prior belief, the more rigid the mould and the less the new data reshapes it, which is correct behaviour for an observer whose confidence tracks accuracy. His laboratory had people judge the direction of a cloud of moving dots, showed them a second cloud moving the same way, and asked them to rate confidence in the original decision. The dorsal anterior cingulate cortex — the region that produced the error signal — showed the pattern a Bayesian model predicts, rising with evidence strength in one direction when the first decision was right and the other when it was wrong, which makes it a belief-updating signal rather than an alarm.

The trouble begins when confidence detaches from accuracy. Max Rollwage, raising confidence with the positive-evidence manipulation without raising accuracy, found people then changed their minds less often. Confirmation bias arrives at the same place from the other side: when people bet on house prices and then saw a partner agree or disagree, agreement raised their confidence substantially while disagreement lowered it only slightly. Rollwage's magnetoencephalography experiment sharpened this — when people were highly confident, evidence contradicting their choice was close to undecodable from the magnetic signals, as though the brain had stopped processing it. Fleming's synthesis is that confirmation bias is tolerable, even efficient, when metacognition is good, since it then discards contradiction mostly where one is in fact right, and corrosive when confidence and accuracy have come apart. Two online experiments found the link directly: the people whose confidence best tracked their accuracy were the ones most willing to reverse a decision after an error.

With Benedetto De Martino, Fleming extended the question to choices that have no right answer. The standard economic account makes confidence proportional to the difference in value between the options, which leaves no room for regret — you can never be more confident in the option you did not take. Their volunteers, fasted for four hours, chose between hundreds of pairs of snacks, one choice played out for real and eaten afterwards, and separately stated what they would pay for each. Confidence rose with the value difference, but at equal differences it varied, and the low-confidence choices were the ones where people had taken the item they would pay less for — and, when the pairs came round again, the ones they switched.

Against this sits a social pressure in the opposite direction. Dominic Johnson and James Fowler's evolutionary simulations found overconfident agents doing better where resources were contested, because they entered contests they would otherwise have avoided. In a study where pairs located American cities on a map, the overconfident members were judged more competent by their partners and won more respect, and video ratings showed them talking more, in a more assertive tone, looking relaxed; the authors concluded that "overconfident individuals more convincingly displayed competence cues than did individuals who were actually competent." Fleming's way out is to split private confidence from the confidence one signals. Dan Bang had volunteers judge moving dots alongside partners engineered to be systematically over- or underconfident, under a rule that made the more confident judgment the group's; they spontaneously matched their partners, damping their reports with timid ones and raising them with brash ones. Ventromedial prefrontal cortex tracked private confidence and ignored the partner; the lateral frontal pole tracked the strategic adjustment. Bluffing, he notes, requires knowing that one is bluffing.

Knowing this about others is what makes collaboration work. Bahador Bahrami's experiments, published in Science in 2010, had pairs judge which of two flashes held a brighter target and discuss any disagreement; the joint decision was usually better than the better of the two alone, because people communicate confidence and a good combination rule weights each opinion by its reliability. The effect depends entirely on that confidence being informative, which is why Fleming turns to the places where it is not. Eyewitness misidentification contributed to about seventy per cent of the more than 375 American convictions overturned by DNA evidence, mock juries weigh a witness's confidence more heavily than the consistency of the testimony, and Thomas Busey and Elizabeth Loftus found that brightening the faces at test lowered accuracy while raising confidence. Since confidence at a first, uncontaminated identification does predict accuracy and decays afterwards, the repair is to record a number at the lineup and read it back at trial. When Fleming and the litigation funder Robert Rothkopf asked 250 lawyers and clients what a phrase like "significant likelihood" meant as a percentage, the answers ran from below 25 per cent to near 100, so Rothkopf's team now requires numbers. His own laboratory found that the best single predictor of dogmatic political views, on the left and the right equally, was poor metacognitive sensitivity on a dot-counting task, which also predicted an unwillingness to seek out new information.

Self-awareness in the age of machines

The last part of the argument is about explanation, and it starts with skill that runs without it. Sian Beilock tested forty-eight Michigan State undergraduates, some on the golf team, on 1.5-metre putts; the experts putted better and described what they had just done in less detail than the novices, an effect she called expertise-induced amnesia, which vanished when they were handed a deliberately awkward putter. Asked to monitor a series of beeps instead of their swing, they putted better still, which suggests that choking is the pilot interfering with the autopilot, and that the best players often make poor coaches. The extreme case is chick sexing, taught only by apprenticeship because no one could state the rule; one Japanese expert sorted 1,400 chicks in an hour at 98 per cent accuracy in 1935, and practitioners still cannot say how. Blindsight is the neurological version: Lawrence Weiskrantz's patient DB guessed the location of stimuli in his cortically blind field well above chance while insisting he saw nothing, with no confidence to attach to the guesses.

Where explanation is given, it is often manufactured. Lars Hall and Petter Johansson's supermarket experiment swapped the jam jars between the choice and the justification; only about a third of the 180 participants noticed, and the rest explained with enthusiasm why they preferred the jam they had rejected. Michael Gazzaniga's split-brain patients supply the mechanism: shown the instruction to walk in the right hemisphere, a patient stood and left the room and, asked why, said he wanted a drink. Gazzaniga called the speaking hemisphere the interpreter. The sense of agency is equally constructed; Daniel Wegner's comparison is that the feeling of willing is a speedometer reading rather than the engine. Fleming's answer is that the narratives are accurate enough for most purposes, and that responsibility, like money, is a collective agreement rather than a natural fact, which does not stop it working. The law already runs on it: every one of the Model Penal Code's four degrees of culpability is defined by a term about awareness, which is why an automatism defence dissolved the prosecution of Brian Thomas, who strangled his wife during a night terror in Wales in 2008.

This is what machines lack. Fleming's example is Air France 447, whose crew stalled a famously crash-resistant Airbus into the Atlantic in June 2009 after an iced sensor disconnected the autopilot, having spent most of their hours monitoring automation rather than flying. Socrates has Thamus complain that writing will make people seem omniscient while knowing nothing, and Fleming concedes the complaint was wrong about writing, the press and the computer, because those are transparent systems that answer the same way tomorrow. Machine learning is not. Artificial networks classify images superhumanly and AlphaGo beat Lee Sedol in March 2016 by playing itself, but a trained network's knowledge sits in the weights, and Andy Clark and Annette Karmiloff-Smith argued in 1993 that this is why a network trained to predict British loan defaults could teach an Australian network nothing: it holds knowledge in the network rather than for the network. Metacognition, Fleming argues, will not arrive as a by-product of intelligence, any more than it arrived for the chick sexers, and modern systems are characteristically overconfident outside the conditions they were trained on.

Building it in is possible and has begun. Nicholas Yeung, Jonathan Cohen and Matthew Botvinick added one extra unit to a network solving the Stroop task, computing how much its two candidate responses were competing, and reproduced a large body of human error-detection data. Axel Cleeremans's group trained a second network to bet on whether the first would be right, then cut the connection and produced a working simulation of blindsight, performance preserved and confidence uncoupled from it. Practical versions exist, among them a drone carrying a second network trained to predict its own crashes, which declined the manoeuvres it expected to fail. Fleming imagines self-driving cars glowing blue when confident and yellow when not, and lowering their confidence in response to each other at a junction: not theory of mind, but enough shared self-report to coordinate.

The alternative he prefers is to keep the human in the loop by interfacing machines with the metacognition we already have. Brain-computer interfaces, which since Matt Nagle moved a cursor by thought in 2002 have been built to control devices, could as well be built to monitor them. We do not need to understand a machine to supervise it, Fleming argues, any more than we need the optics of the eye to know an image is out of focus. His final chapter asks whether reflection can be trained and answers cautiously yes. Twenty minutes a day for two weeks of practice with feedback on confidence rather than on accuracy raised sensitivity, and the gain transferred to a memory task that had not been trained; two weeks of meditation training raised it in a study by Benjamin Baird and Jonathan Schooler. His caution is that the gains are small and the plasticity asymmetric — nothing training can add compares with what disease takes away, and more accurate self-appraisal is not always kinder. He ends on lucid dreaming, in which the frontal pole and precuneus light up and stimulating prefrontal cortex at the right frequency makes lucidity likelier, as the image for what improved self-awareness might feel like: noticing what was there all along.

Since 2021

The caution Fleming volunteered about his own frontal-pole result has been vindicated on a scale that reaches most of the imaging in the book. A 2022 analysis showed that correlations between individual differences in behaviour and MRI need samples in the thousands to stabilise, where the 2010 study rested on thirty-two volunteers. The regional claim now reads as a hypothesis rather than a settled localisation, while the broader involvement of prefrontal and midline regions in confidence, supported by lesion, stimulation and imaging work together, has held. The measure has been argued over too: meta-d′ assumes that confidence draws on the same evidence as the decision and is sensitive to where confidence criteria fall, and information-theoretic alternatives proposed since, among them Peter Dayan's in 2023, lean less on those assumptions.

Some applied findings have travelled better than others. Disfluent fonts did not survive: independent tests found that Sans Forgetica, the typeface designed in Melbourne to slow readers down, did not improve memory, which leaves desirable difficulty intact as a principle and removes one of its cheapest instruments. Spacing and self-testing, and the gap between how well they work and how well they feel, have held across replications, and the eyewitness position the book reports — confidence recorded at a first, uncontaminated identification is diagnostic, and confidence measured later is not — has become the mainstream recommendation rather than a minority one.

The largest change is in the chapter about machines, and it arrived in the shape Fleming predicted rather than the one most people expected. Large language models perform across domains, produce fluent accounts of their own reasoning on demand, and those accounts need not correspond to the computation that produced the answer — Gazzaniga's interpreter in silicon, and a sharpening of the book's distinction between an explanation a court would accept and a readout of the mechanism. Calibration has become an engineering problem: these systems can report confidence, the reports are sometimes informative and sometimes not, and training them to please human raters tends to make them worse. The deliberate construction of uncertainty estimates, which Fleming argued would have to be designed rather than awaited, is now a standard requirement; his second proposal, keeping people in metacognitive contact with systems they cannot inspect, remains the harder half. On the science's own terms the closing image has aged well: in 2021 several laboratories established two-way communication with lucid dreamers, who answered arithmetic questions with eye movements while asleep, which is as direct a demonstration as the field has that knowing one is conscious is a state that can be measured.