When We Learned to Scroll the World

For most of human history, the main constraint on information was information itself.
It was difficult to create, record, copy, preserve, transmit, and find. Writing reduced our dependence on individual memory. Printing dramatically lowered the cost of reproduction. Telegraphy, radio, and television compressed distance. The internet made immense bodies of knowledge accessible almost instantly.
Then something less visible — and perhaps more consequential — began to happen.
The central problem gradually stopped being:
Where can I find information?
Increasingly, it became:
How do I know which information matters?
The amount of accessible information has grown far faster than any individual’s ability to perceive, verify, and transform it into useful knowledge.
We can call this condition information saturation.
This is not simply information overload. Overload is a state experienced by an individual. Information saturation, as I use the term here, describes a broader change in the environment: the supply of information has become so abundant that access and transmission are no longer the primary constraints. Attention, selection, verification, and processing increasingly are.
Research on information overload has indeed associated it with poorer decision-making, lower productivity, and greater cognitive strain.[1]
But the larger issue may be the growing mismatch between the scale of the digital information environment and the amount of information a human being can process directly.
That mismatch may be one of the keys to understanding what is happening now.

Digital information has an unusual economic property: once created, the marginal cost of reproducing and distributing another copy can approach zero. More broadly, the digital economy has dramatically reduced the costs of storing, copying, transmitting, searching, and processing information.[2]
Generative artificial intelligence takes another step.
In one controlled experiment involving professional writing tasks, access to ChatGPT reduced average completion time by about 40 percent while increasing average output quality by roughly 18 percent.[3]
This does not mean that AI reduces the cost of every form of intellectual work by the same amount. It does show something narrower and more important: for at least some information-intensive tasks, not only reproduction but also production and transformation are becoming cheaper.
Human cognitive constraints are not changing at anything like the same pace.
Attention is finite.
Working memory is finite.[4]
A day still contains twenty-four hours.
The adult human brain accounts for roughly two percent of body mass while consuming around one-fifth of the body’s resting energy expenditure.[5]
That fact does not imply that human biological cognition has reached some absolute ceiling. But biological change and technological change operate on radically different timescales.
As information becomes cheaper, value moves along a chain:
create → access → find → select → verify → understand.
A library was once valuable because it contained information that was difficult to obtain.
Today, access to bodies of knowledge larger than enormous physical libraries can fit in a pocket.
The scarce resource increasingly becomes the ability to determine what matters now.
Attention becomes one constraint.
Trust becomes another.

It would be misleading to suggest that people once merely “consumed information” and have only now begun to live inside an information environment.
Humans have always lived inside one.
Sounds, tracks, facial expressions, rumors, stories, traditions, language, ritual, and reputation all formed informational environments long before digital technology.
What changed were the parameters.
The radius became global.
The density increased.
The refresh rate accelerated.
Exposure became nearly continuous.
A significant share of what reaches us now arrives without deliberate search. A systematic review of incidental news exposure identified 88 studies examining precisely this phenomenon in digital environments.[6]
A useful metaphor for this condition is an information climate.
This is not an established scientific term. It is a working metaphor.
An individual message can be compared to a weather event. But people also exist inside the statistical distribution of thousands of weak signals: which topics recur, which events appear common, which opinions seem normal, which threats appear probable, which behaviors look prestigious or desirable.
Any single signal may be too weak to remember or consciously analyze.
Together, however, they can contribute to a background model of reality.
Research on ambient awareness has shown that streams of fragmentary social information can create surprisingly coherent impressions of other people and of the surrounding social environment.[7]
A separate literature examines the effect of repetition. Under many conditions, repeated claims become more believable merely because they have been encountered before — including when the information is false.[8]
Then there is another mechanism: the feedback loop between the user and the algorithm.
In a month-long randomized experiment involving more than two thousand active YouTube users, researchers found a reciprocal relationship between recommendations and consumption: recommendations influenced what people watched, while viewing behavior altered later recommendations. In the studied setting, the influence of recommendations on consumption was considerably stronger than the reverse effect.[9]
The loop looks something like this:
human choice → algorithm → modified information environment → next human choice.
This is no longer merely a channel for transmitting messages.
It is an adaptive information environment.

Technology is often discussed as though biological evolution operated first, and then humans somehow stepped outside it and began an entirely artificial history.
That division is misleading.
Technology is produced by a biological organism.
The capacity to learn socially and transmit behavior through others also emerged within the living world.
Cultural traditions and social learning are not uniquely human.
Recent work in cultural evolution has even challenged the familiar claim that cumulative culture itself is uniquely human. A more cautious interpretation is that human distinctiveness lies in the extraordinary open-endedness of our cultural possibility space: our capacity to generate an enormous range of new behaviors, knowledge systems, and technologies.[10]
At some point, adaptation in the human lineage began to occur not only through changes in the organism.
The organism increasingly changed its environment and constructed external means of overcoming its own limitations.
Cold does not require waiting for biological changes in body hair. Clothing can be made.
Insufficient strength can be supplemented with a lever.
Poor vision with optics.
Limited memory with writing.
Difficult calculation with calculating machines.
Navigation with maps and later GPS.
Information search with search engines.
A second adaptive loop emerges:
variation → trial → selection → preservation → transmission → new variation.
Cultural adaptation can operate much faster than genetic adaptation because behavioral innovations can arise and spread within a single lifetime and can move horizontally among contemporaries rather than only vertically between generations.[11]
Culture does not replace biological evolution.
It changes the conditions under which biological evolution occurs.
Gene–culture coevolution is well documented. One classic example is the interaction between dairying and genetic variants associated with lactase persistence in adulthood: a cultural practice altered the selective environment, and human populations responded biologically.[12]
It may therefore be more useful to treat humans and technology as parts of a coupled system:
organism ↔ culture ↔ technology ↔ modified environment.
Each component can change the others.

This leads to the central hypothesis of this essay.
Perhaps human history can usefully be understood not as one great transition from nature to civilization, and not as a march toward a predetermined destination, but as a sequence of changes in adaptive regime.
This is not an established law of history.
It is a working model.
Its structure might look roughly like this.
A constraint exists.
A mechanism appears that partly bypasses it.
The new mechanism changes the environment.
Old behaviors and institutions continue for some time inside conditions for which they were not designed.
A mismatch emerges.
Social, economic, and demographic structures gradually adjust.
Then another constraint becomes visible.
Cumulative culture increased the speed at which adaptive solutions could be transmitted.
Agriculture allowed ecosystems to be reorganized systematically for the production of resources.
The Industrial Revolution greatly weakened the dependence of production on the annual flow of biological energy and on human or animal muscle. Fossil fuels expanded the energy budget available to society by orders of magnitude.[13]
The digital revolution sharply reduced the costs of storing, copying, and transmitting information.[2]
Again and again, an old scarcity partially became abundance.
And a new scarcity appeared.
If this model is useful, the transition now under way need not be a single final “singularity.”
It may instead be another stage in a much longer sequence.

Demography is one of the most interesting — and most dangerous to oversimplify — variables in such a system.
Declining fertility is often attributed to one dominant cause:
urbanization;
housing costs;
female education;
individualism;
contraception;
changing family values;
digital life.
The evidence points to a much more complicated picture.
According to the OECD, the average total fertility rate across member countries had fallen to about 1.5 children per woman by 2022. The average age at first birth increased from approximately 26.4 years in 2000 to 29.5 years in 2022.[14]
Yet the differences among countries are enormous.
Israel is particularly interesting. In 2022 it had a fertility rate of roughly 2.9 — the highest in the OECD and the only one above replacement level.[14]
That alone is enough to show that the equation
modernity → low fertility
is too simple.
The OECD associates fertility change with a range of interacting factors, including contraception, women’s education, the time required to establish a position in the labor market, work–family compatibility, housing conditions, and other institutional and economic constraints.[14]
Fertility should therefore not be treated here as a direct consequence of information saturation.
It is more useful to view it as a slow variable within a broader social reconfiguration.
In 1996, Sergei Kapitsa proposed a phenomenological model in which the demographic transition was not treated simply as the result of an external resource constraint, but as a systemic shift in the growth regime of the world population.[15]
His numerical prediction — stabilization near 14 billion people — differs substantially from current demographic projections and is not the modern demographic consensus.
But the question he raised remains interesting: demographic slowing may result not only from shortages of food, land, or other external resources, but also from changes in the internal dynamics of human society.
Current United Nations projections are very different.
According to World Population Prospects 2024, the global population is expected to increase from about 8.2 billion in 2024 to roughly 10.3 billion in the mid-2080s, followed by a slight decline to around 10.2 billion by 2100. Around one-quarter of the world’s population already lives in countries whose populations have passed their peak.[16]
None of this implies that humanity is converging toward some predetermined “optimal” population size.
Nor does it mean that today’s low-fertility regimes are permanent.
Demographic regimes can change too.

Seen in this context, artificial intelligence becomes less alien.
It did not emerge independently of humanity.
Behind it lie accumulated mathematics, physics, engineering, written knowledge, institutions, universities, industry, energy systems, computer networks, and the work of many generations.
No individual human being could build a modern AI system alone.
In that sense, AI is a product of humanity’s collective cognitive system.
At the same time, it is becoming a new external layer of that system.
Many earlier technologies externalized relatively discrete cognitive functions.
Writing externalized memory.
Calculators externalized arithmetic.
Maps externalized part of spatial orientation.
Search engines externalized search.
The externalization of cognition itself is not new. Psychology has long studied cognitive offloading: the use of external actions or devices to reduce demands on internal memory and cognition. Research shows both clear performance advantages and possible costs when people become too dependent on external supports.[17]
What changes with modern AI is the scale and range of what can be offloaded.
AI systems can increasingly perform chains of linked operations:
find → compare → synthesize → propose → calculate → create → verify → revise.

This may be the most important transition currently under way.
The internet made the following statement widely true:
I can find out how this is done.
AI increasingly adds another:
I can try to do it.
The difference is enormous.
Historically, between an individual with an idea and a complex result stood an organization.
Even a relatively modest product might require programmers, designers, analysts, technical specialists, marketers, lawyers, and managers.
A person with a coherent vision first had to obtain resources and then translate that vision repeatedly into the specialized languages of different people.
Every translation imposed a cost and created another opportunity for the original intention to be distorted.
Early experiments suggest that AI can partially change this structure.
In a field experiment involving 791 professionals at Procter & Gamble working on real product-development problems, individuals working with AI performed, on average, at the level of human teams working without AI. AI also helped technical and commercial specialists produce more balanced solutions outside the boundaries of their normal professional expertise.[18]
This does not mean that an individual equipped with AI can generally replace a real team.
The experiment concerned a specific class of innovation task.
But it demonstrates an important possibility: some of the benefits of organizational division of labor can be obtained without enlarging the human team.
Early observational evidence from entrepreneurship points in a similar direction.
A recent conference paper analyzing a large dataset of Product Hunt launches reported a disproportionate increase in projects by solo founders after the arrival of ChatGPT, particularly in areas where teams had previously been more common. Teams, however, continued to dominate among the highest-ranked projects.[19]
This result should be treated cautiously. It is a conference publication, not the final word on entrepreneurial organization.
But together with experimental evidence it suggests a more modest and more interesting hypothesis:
the minimum organizational scale required to test a complex idea may be falling.
In the past, a strong idea could remain invisible because its originator lacked the money or specialized collaborators needed to turn it into a calculation, model, or prototype.
A shorter chain is becoming possible:
idea → model → calculation → prototype → test.
This is not the same as building a railway, a power plant, or a factory.
Large projects still require capital, infrastructure, regulation, manufacturing, logistics, and large numbers of people.
What changes is that an individual may increasingly be able to bring an idea to the point where it can be evaluated seriously.
That could alter the relationship between the generation of ideas and the resources required to implement them.

Another transition may follow.
As long as a well-developed proposal is rare, simply turning an idea into a convincing project has substantial value.
But if financial models, technical descriptions, visualizations, and software prototypes can be produced relatively cheaply, polished presentation alone ceases to be scarce.
The world may instead be flooded with convincing-looking projects.
Then a new constraint appears: the ability to distinguish
a weak idea that is beautifully packaged
from
an idea that is genuinely strong.
Resource-rich organizations — corporations, investment funds, governments, foundations — may eventually compete not only for capital and skilled labor, but for the ability to identify genuinely valuable ideas among an enormous number of cheaply generated possibilities.
This is a hypothesis, not an established economic fact.
But it follows the broader pattern:
when one constraint is removed, another becomes visible.

If producing text becomes easy, the mere act of producing text becomes less valuable.
If code becomes cheaper to generate, mechanical code production becomes relatively less scarce.
If calculations can be performed almost instantly, calculation itself ceases to be the main bottleneck.
Higher-level questions then become more important:
What should we calculate?
Why?
What problem are we actually trying to solve?
What would a good result look like?
Is this project worth pursuing at all?
AI can generate vast numbers of possibilities.
But the space of possible solutions is always much larger than the space of useful ones.
The next bottleneck may therefore be not computational capacity but the ability to formulate meaningful goals and good questions.
Again, this is a hypothesis.
It does not imply the emergence of a new “superior class” of people.
A changing technological environment simply changes the relative value of different cognitive functions.
Some forms of execution become cheaper.
Seeing the whole system, connecting domains, framing a problem, and evaluating the result may become more valuable.

The most common question about AI is how many jobs it will destroy.
That may be too static a way of framing the problem.
A profession is composed of tasks.
Technology can automate some tasks, augment others, and create new ones.
Modern economic theories of automation therefore distinguish between the displacement of labor from existing tasks and the creation of new tasks in which human labor again gains comparative advantage.[20]
The International Labour Organization’s 2025 assessment likewise suggests transformation rather than the immediate disappearance of most occupations. It estimated that roughly one in four workers worldwide is employed in an occupation with some exposure to generative AI, while concluding that transformation of job content is more likely than complete replacement for most occupations because human input remains necessary.[21]
There is another effect as well.
AI may not only substitute for some skilled work. It may also reduce the skill threshold for performing certain tasks.
In a study of more than five thousand customer-support agents, access to an AI assistant increased productivity by about 15 percent on average. The largest gains went to less experienced and lower-performing workers, while the most experienced workers benefited much less.[22]
A similar pattern appeared in three field experiments involving 4,867 software developers at Microsoft, Accenture, and a Fortune 100 company. Access to an AI coding assistant increased completed tasks by roughly 26 percent, with less experienced developers receiving larger gains on average.[23]
These findings do not establish a universal rule.
But they allow us to imagine a different organization of some forms of work:
deep expert → coordinator → AI-augmented operators.
The expert does not disappear.
Where mistakes are expensive, the ability to understand the limits of a machine-generated answer may become even more important.
But some tasks that once required years of specialist training may become accessible to a wider range of people when supported by external cognitive systems.
The transition may therefore look less like
AI → humans become unnecessary
and more like
automation of old tasks → cheaper creation of new systems → new tasks → new division of labor.
Whether the creation of new tasks will compensate for the displacement of old ones, how quickly the process will unfold, and how painful the transition will be remain open questions.

There is another constraint that conventional economic statistics capture poorly.
A person may have an idea — and even the ability to pursue it — and still never try.
The reasons are familiar:
doubt;
fear of failure;
uncertainty;
fear of wasting time;
the need for external validation.
Between capability and action lies a psychological threshold:
Is this even worth trying?
There is not yet enough evidence to claim that generative AI has already lowered this threshold on a population scale.
But the mechanism deserves investigation.
Research on entrepreneurship has repeatedly connected self-efficacy — confidence in one’s ability to perform a task — with lower fear of failure and greater willingness to act.[24]
AI creates a new form of cheap preliminary testing.
A person can immediately ask:
What have I missed?
Is there a fundamental flaw?
What are the strongest counterarguments?
What is the cheapest way to test the central assumption?
If such interactions genuinely increase willingness to act, a reinforcing loop becomes possible:
support → attempt → real experience → greater competence → next attempt.
Learning, however, does not come only from success.
A major review of research on error-based learning shows that an incorrect attempt followed by high-quality corrective feedback can improve subsequent learning; understanding the reason for the error matters more than merely receiving the correct answer.[25]
If AI reduces both the material and psychological cost of experimentation, it could therefore accelerate not only task completion but also the accumulation of experience.
Yet an opposite risk appears immediately.
An AI system that constantly validates a user may strengthen false confidence rather than improve thinking.
A useful external intelligence must therefore perform two nearly opposite functions:
help a person become willing to try
while also
helping that person remain uncertain enough to test the idea.

At this point we reach a separate subject that deserves its own treatment.
As long as interaction with AI is episodic, AI remains an unusually powerful tool.
If interaction becomes longitudinal, something qualitatively different appears.
A system could potentially take into account what a person learned a year ago, which mistakes keep recurring, which projects are repeatedly abandoned, which explanations work for that person, and which decisions were made in the past.
In that case, what is externalized is no longer only memory or calculation.
Part of metacognition — the ability to observe one’s own thinking — begins to move outside the individual as well.
We can imagine the emergence of a long-term cognitive companion: an interactive layer between an individual, humanity’s accumulated knowledge, and that individual’s own life trajectory.
For now, this is primarily a research hypothesis and a forecast. Humanity simply does not yet have decades of evidence on people living continuously alongside systems of this kind.
What we already know from cognitive offloading is enough to reveal the central tension: external support can improve immediate performance while also creating dependence on the external system.[17]
The criterion for a good cognitive companion should therefore be paradoxical:
the longer a person uses a good external intelligence, the more autonomous that person should become.
They should become better at asking questions.
Better at noticing their own mistakes.
More willing to explore.
More aware of the limits of their own competence.
External intelligence should prevent a person from falling off a cliff without preventing them from stumbling.
Removing every risk and every error would also remove part of the experience through which autonomy develops.

Put these pieces together and a possible sequence begins to appear.
Humans produce increasing amounts of information.
Information becomes more abundant than individuals can directly process.
Search and filtering technologies emerge.
Those technologies help create an even denser and more personalized information environment.
Artificial intelligence appears, capable not only of finding information but of performing chains of cognitive operations.
It then begins to shorten the distance between intention and action.
That increases the number of possibilities that can be tested at relatively low cost.
And this, in turn, may accelerate cultural selection:
more variations → more trials → more errors → more selection → more retained solutions.
None of these arrows, by itself, proves the existence of a universal historical law.
Together, however, they support a testable hypothesis:
Artificial intelligence may be more than another technology for processing information. It may become a new layer of cultural adaptation, accelerating the production, testing, and selection of human solutions.

Perhaps this is the most important point.
We are not watching humanity leave the natural world.
We may be watching an unusual continuation of a natural process.
Biological evolution produced an organism capable of transmitting acquired solutions culturally.
Culture produced technology.
Technology altered the environment.
The altered environment changed the constraints under which humans lived and acted.
Now human-created information systems are feeding back into that cycle, altering the speed at which new solutions can be produced, tested, and transmitted.
The structure becomes recursive:
humans create tools → tools change human capabilities → changed humans create new tools.
Information saturation may therefore be more than a problem of having too much information.
It may be a sign that the old model of direct human engagement with information is approaching a practical limit.
If so, artificial intelligence is not an arbitrary addition to the system.
It is a functional response to a constraint that the cultural-technological system itself helped create.
This is not a proven law.
It is a hypothesis that can now be compared against observable changes in productivity, entrepreneurship, labor, learning, demography, and the production of knowledge.
If it is even partly correct, we are not merely witnessing the birth of another technological industry.
We may be observing another shift in the human adaptive regime — a period in which culturally created cognitive infrastructure begins to enter the everyday loop of human thought itself.
And if that is the case, the central question of the coming decades may not be:
How intelligent will artificial intelligence become?
It may be:
What will humans become when they live continuously alongside an external intelligence?

1. Shahrzadi, L., Mansouri, A., Alavi, M., & Shabani, A. “Causes, consequences, and strategies to deal with information overload: A scoping review.” International Journal of Information Management Data Insights, 4(2), 100261 (2024). DOI: 10.1016/j.jjimei.2024.100261
2. Goldfarb, A., & Tucker, C. “Digital Economics.” Journal of Economic Literature, 57(1), 3–43 (2019). DOI: 10.1257/jel.20171452
3. Noy, S., & Zhang, W. “Experimental evidence on the productivity effects of generative artificial intelligence.” Science, 381(6654), 187–192 (2023). DOI: 10.1126/science.adh2586
4. Cowan, N. “The magical number 4 in short-term memory: A reconsideration of mental storage capacity.” Behavioral and Brain Sciences, 24(1), 87–114 (2001). DOI: 10.1017/S0140525X01003922
5. Shulman, R. G., Rothman, D. L., Behar, K. L., & Hyder, F. “Energetic basis of brain activity: implications for neuroimaging.” Trends in Neurosciences, 27(8), 489–495 (2004). DOI: 10.1016/j.tins.2004.06.005
6. Schäfer, S. “Incidental News Exposure in a Digital Media Environment: A Scoping Review of Recent Research.” Annals of the International Communication Association, 47(2), 242–260 (2023). DOI: 10.1080/23808985.2023.2169953
7. Levordashka, A., & Utz, S. “Ambient awareness: From random noise to digital closeness in online social networks.” Computers in Human Behavior, 60, 147–154 (2016). DOI: 10.1016/j.chb.2016.02.037
8. Udry, J., & Barber, S. J. “The illusory truth effect: A review of how repetition increases belief in misinformation.” Current Opinion in Psychology, 56, 101736 (2024). DOI: 10.1016/j.copsyc.2023.101736
9. Yu, X., Haroon, M., Menchen-Trevino, E., & Wojcieszak, M. “Nudging recommendation algorithms increases news consumption and diversity on YouTube.” PNAS Nexus, 3(12), pgae518 (2024). DOI: 10.1093/pnasnexus/pgae518
10. Morgan, T. J. H., & Feldman, M. W. “Human culture is uniquely open-ended rather than uniquely cumulative.” Nature Human Behaviour, 9, 28–42 (2025). DOI: 10.1038/s41562-024-02035-y
11. Fogarty, L., & Kandler, A. “The fundamentals of cultural adaptation: implications for human adaptation.” Scientific Reports, 10, 14318 (2020). DOI: 10.1038/s41598-020-70475-3
12. Laland, K. N., Odling-Smee, J., & Myles, S. “How culture shaped the human genome: bringing genetics and the human sciences together.” Nature Reviews Genetics, 11, 137–148 (2010). DOI: 10.1038/nrg2734
13. Wrigley, E. A. “Energy and the English Industrial Revolution.” Philosophical Transactions of the Royal Society A, 371, 20110568 (2013). DOI: 10.1098/rsta.2011.0568
14. OECD. “Fertility.” In Society at a Glance 2024: OECD Social Indicators. OECD Publishing (2024).
15. Kapitsa, S. P. “The phenomenological theory of world population growth.” Physics-Uspekhi, 39, 57–71 (1996). DOI: 10.1070/PU1996v039n01ABEH000127
16. United Nations, Department of Economic and Social Affairs, Population Division. World Population Prospects 2024: Summary of Results. United Nations (2024).
17. Richmond, L. L., & Taylor, R. G. “The benefits and potential costs of cognitive offloading for retrospective information.” Nature Reviews Psychology, 4, 312–321 (2025). DOI: 10.1038/s44159-025-00432-2
18. Dell’Acqua, F., Ayoubi, C., Lifshitz, H., Sadun, R., Mollick, E., Mollick, L., Han, Y., Goldman, J., Nair, H., Taub, S., & Lakhani, K. R. “The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork.” Organization Science (2026). DOI: 10.1287/orsc.2025.20702
19. Kim, H., Kang, H., & Song, J. “How Generative AI Enables Solo Entrepreneurship While Preserving Team Advantages.” Academy of Management Proceedings (2026). DOI: 10.5465/AMPROC.2026.16114abstract
20. Acemoglu, D., & Restrepo, P. “Automation and New Tasks: How Technology Displaces and Reinstates Labor.” Journal of Economic Perspectives, 33(2), 3–30 (2019). DOI: 10.1257/jep.33.2.3
21. Gmyrek, P., Berg, J., Kamiński, K., et al. Generative AI and Jobs: A Refined Global Index of Occupational Exposure. ILO Working Paper 140. International Labour Organization (2025).
22. Brynjolfsson, E., Li, D., & Raymond, L. “Generative AI at Work.” The Quarterly Journal of Economics, 140(2), 889–942 (2025). DOI: 10.1093/qje/qjae044
23. Cui, Z., Demirer, M., Jaffe, S., Musolff, L., Peng, S., & Salz, T. “The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers.” Management Science (2026). DOI: 10.1287/mnsc.2025.00535
24. Deligianni, I., Liouka, I., Oguguo, P. C., & Voutsina, K. “Exploring fear of failure in nascent entrepreneurship: The role of self-efficacy and the moderating effects of the institutional context.” European Management Review (2026). DOI: 10.1111/emre.70006
25. Metcalfe, J. “Learning from Errors.” Annual Review of Psychology, 68, 465–489 (2017). DOI: 10.1146/annurev-psych-010416-044022