Physical Labor and Cognitive Labor

In Capital, Marx compared human labor with the work of animals:

The same mix runs through modern roles. In O*NET's profile for transit bus drivers, the most important work activity is operating the vehicle, and most of the activities ranked just below it involve the public, information, or decisions.^[2] Over a shift, an operator decides whether a rider needs help, how to handle an incident, and what to write in the incident report.

An insurance claims analyst has the reverse mix: nearly every step produces information, though gathering evidence can mean a site visit.

Two Kinds of Output

The two kinds of labor differ in what a step produces:

Earlier in this series, knowledge work was defined as "the transformation of information into more information." Both kinds of labor are labor in the sense of "What Is Work?" when they are paid. "Physical" and "cognitive" describe the content of a step, and a role's mix of the two can be measured step by step. Part 9 looks at what happens inside a cognitive step.

How the Split Became a Job Structure

In craft production, the same person usually planned the work and carried it out. A cabinetmaker worked out the design, chose the materials, and built the piece. Frederick Taylor's scientific management moved planning into a separate department, and workers received written instructions for each task.^[3]

Harry Braverman argued that this separation of conception from execution produced the modern divide between white-collar and blue-collar jobs.^[4] His further claim, that the separation deskilled work, has been disputed.^[5]

The division between physical jobs and cognitive jobs is partly a choice about how to bundle steps. A role labeled physical, such as the bus operator's, can contain cognitive steps that software can now reach, while some of a claims analyst's evidence gathering still has to be done in person.

Measuring Each Kind of Labor

Physical labor can be measured in energy.

Mechanical work is force times distance, and a person can sustain only about 60 to 100 watts of useful mechanical output over a working day.^[6] James Watt rated his engines in horsepower, about 746 watts. When engines replaced muscle in industrial Britain, the change showed up as coal-fired energy replacing human and animal effort.^[7]

Cognitive labor does not show up in energy accounts.

The human brain is about 2 percent of body weight and uses about 20 percent of the body's energy, and that rate stays nearly constant across very different mental and physical activity.^[8] The extra energy used during demanding mental tasks is small compared with the brain's baseline, often less than 5 percent.^[9]

So cognitive labor is measured by time, attention, and output. Daniel Kahneman described mental effort as a limited capacity that people allocate across tasks.^[10] Wouter Kool and Matthew Botvinick apply the labor-supply model from economics to mental effort, treating it as a cost people weigh against its rewards.^[11]

Markets count cognitive labor in hours and in outputs such as claims closed or reports filed. The quality of those outputs varies widely, and a correct diagnosis can take the same time as a wrong one. That makes productivity in cognitive work harder to measure than productivity in physical work.

How Each Kind Gets Automated

Physical steps were automated first by supplying power, through water wheels, steam engines, and electric motors, and then by supplying control, through numerical control and industrial robots. Control works best in a structured environment. Autor notes that production robots are bolted to the floor and caged because they cannot sense what is around them.^[12]

Computers took over cognitive steps first by following rules, executing procedures that people could write down.^[13] Machine learning then let computers learn some tasks from examples. LLMs now generate language across many domains.

Hans Moravec observed that:

"it is comparatively easy to make computers exhibit adult-level performance in solving problems on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility."^[14]

The current wave fits his observation. Earlier computerization hit routine middle-wage work hardest. In the 2023 working-paper version of their study, Tyna Eloundou and colleagues estimated that about 80 percent of US workers have at least a tenth of their tasks exposed to LLMs, with exposure higher in higher-wage occupations.^[15] Exposure means LLMs can perform or assist those steps — not that those steps have been automated. Even within the newly reached territory, many cognitive steps retain hard-to-automate properties: the output requires judgment no procedure captures, someone must answer for it, or the knowledge behind it cannot be fully stated. Plumbing, elder care, and skilled repair remain hard to automate.

The Shift Toward Cognitive Labor

US employment has shifted toward jobs built around abstract cognitive tasks. Professional, technical, and managerial occupations grew by about 3 percentage points of total employment per decade from 1940 to 1980, and faster from 1980 to 2010.^[16]

To understand why, it helps to classify occupations along two dimensions: whether their tasks are routine or non-routine, and whether those tasks are primarily cognitive or manual. Routine tasks are those whose rules can be written down and executed step by step — a clerical procedure, an assembly-line operation, an accounting rule. Non-routine tasks require judgment that no procedure fully specifies, whether physical (a plumber diagnosing an unfamiliar leak) or cognitive (an analyst synthesizing a recommendation from ambiguous evidence).^[13]

Middle-skill jobs are occupations concentrated in the routine quadrants of this map — clerical work, production and craft, administrative support, data processing, and routine sales. They sit in the middle of the wage distribution because their tasks can be learned without exceptional education but are not particularly scarce.

The defining economic property is codifiability: the rules of a middle-skill job can be written down, which means a sufficiently capable machine can execute them.

The result, documented across US commuting zones and time periods, was a hollowing out of the middle. Between 1980 and 2010, employment grew at both ends of the skill distribution and contracted in the middle. Professional, technical, and managerial roles — concentrated in non-routine cognitive tasks — expanded. In-person service roles — personal care, food preparation, cleaning, building security — also grew, because their non-routine physical steps require presence and perception that automation could not provide. Routine jobs in both quadrants shrank: clerical work, production, and craft were displaced by software and industrial robots.^[17]

The polarization pattern is why the current wave of automation matters more than earlier ones. A large and growing share of labor now sits in roles dominated by non-routine cognitive steps — the tier that grew throughout the twentieth century. Earlier computerization displaced middle-skill routine cognitive jobs. The LLM wave reaches a different tier: exposure is highest in professional and analytical occupations, the ones that survived and grew through all prior waves of automation.^[15]

What Remains Distinct in Human Cognitive Labor

Four features of human cognitive labor bear on which steps stay human:

• Many cognitive outputs have no fixed answer set. Checking an analysis of a loan file takes judgment of the same kind that produced it. Herbert Simon called decisions of this kind "nonprogrammed."^[18]
• People connect words to things they have encountered. Stevan Harnad argued that a system that only manipulates symbols has no such connection on its own.^[19]
• Much expertise is tacit. Michael Polanyi wrote that "we can know more than we can tell."^[20](e.g., "Polanyi's Paradox")
• A person can be held responsible for a decision, and can reflect on and endorse or reject their own reasons for it.^[21]

The cognitive steps most likely to stay human are those whose outputs someone must verify or answer for. Overseeing automated cognitive work takes effort for the same reason, as "Labor Automation" and Part 9 describe.

---

References

1. Marx, K. (1867). Capital: A Critique of Political Economy, Vol. I, ch. 7. (S. Moore & E. Aveling, Trans., 1887.) marxists.org
2. National Center for O*NET Development. Summary Report for 53-3052.00, Bus Drivers, Transit and Intercity. O*NET OnLine. Accessed 12 October 2026. [work activities ranked by importance] onetonline.org
3. Taylor, F. W. (1911). The Principles of Scientific Management. Harper & Brothers.
4. Braverman, H. (1974). Labor and Monopoly Capital: The Degradation of Work in the Twentieth Century. Monthly Review Press.
5. Attewell, P. (1987). The deskilling controversy. Work and Occupations, 14(3), 323–346. doi:10.1177/0730888487014003001
6. Smil, V. (2017). Energy and Civilization: A History. MIT Press. ["60–100 W of sustained (daylong) useful labor"; horsepower as 745.7 W]
7. Wrigley, E. A. (2010). Energy and the English Industrial Revolution. Cambridge University Press. doi:10.1017/CBO9780511779619
8. Raichle, M. E., & Gusnard, D. A. (2002). Appraising the brain's energy budget. Proceedings of the National Academy of Sciences, 99(16), 10237–10239. doi:10.1073/pnas.172399499
9. Raichle, M. E. (2010). Two views of brain function. Trends in Cognitive Sciences, 14(4), 180–190. doi:10.1016/j.tics.2010.01.008 See also Raichle, M. E., & Mintun, M. A. (2006). Brain work and brain imaging. Annual Review of Neuroscience, 29, 449–476. doi:10.1146/annurev.neuro.29.051605.112819
10. Kahneman, D. (1973). Attention and Effort. Prentice-Hall.
11. Kool, W., & Botvinick, M. (2018). Mental labour. Nature Human Behaviour, 2(12), 899–908. doi:10.1038/s41562-018-0401-9 See also Kool, W., & Botvinick, M. (2014). A labor/leisure tradeoff in cognitive control. Journal of Experimental Psychology: General, 143(1), 131–141. doi:10.1037/a0031048
12. Autor, D. H. (2015). Paradox of abundance: Automation anxiety returns. In S. Rangan (Ed.), Performance and Progress: Essays on Capitalism, Business, and Society (pp. 237–260). Oxford University Press. doi:10.1093/acprof:oso/9780198744283.003.0017
13. Autor, D. H., Levy, F., & Murnane, R. J. (2003). The skill content of recent technological change: An empirical exploration. Quarterly Journal of Economics, 118(4), 1279–1333. doi:10.1162/003355303322552801
14. Moravec, H. (1988). Mind Children: The Future of Robot and Human Intelligence. Harvard University Press. [often cited as p. 15]
15. Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023). GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. arXiv:2303.10130. The published version, Science, 384(6702), 1306–1308 (2024), doi:10.1126/science.adj0998, reports different headline measures.
16. Autor, D. H. (2015). Why are there still so many jobs? The history and future of workplace automation. Journal of Economic Perspectives, 29(3), 3–30, at p. 10. doi:10.1257/jep.29.3.3
17. Autor, D. H., & Dorn, D. (2013). The growth of low-skill service jobs and the polarization of the US labor market. American Economic Review, 103(5), 1553–1597. doi:10.1257/aer.103.5.1553
18. Simon, H. A. (1960). The New Science of Management Decision. Harper & Row. [nonprogrammed decisions, pp. 5–6]
19. Harnad, S. (1990). The symbol grounding problem. Physica D, 42(1–3), 335–346. doi:10.1016/0167-2789(90)90087-6
20. Polanyi, M. (1966). The Tacit Dimension. Doubleday.
21. Frankfurt, H. G. (1971). Freedom of the will and the concept of a person. Journal of Philosophy, 68(1), 5–20. doi:10.2307/2024717