The substitution myth

The substitution myth is the assumption that technology can be introduced as a straight swap of machines for people, leaving the system otherwise intact and better on some measure. Sidney Dekker and David Woods named it in 2002 and set out why it fails: automation transforms the work rather than subtracting a piece of it. Their sentence is the whole argument. "Capitalizing on some strength of automation does not replace a human weakness. It creates new human strengths and weaknesses, often in unanticipated ways." The term is theirs and belongs to cognitive systems engineering, not to this research.

Where it comes from#

The paper is called MABA-MABA or Abracadabra? Progress on Human-Automation Co-ordination, and the title is doing work. MABA-MABA stands for Men-Are-Better-At, Machines-Are-Better-At: the family of lists that allocate tasks by comparing what each party does well. The archetype is the Fitts list, from a 1951 report on air navigation and traffic control, and it has been reinvented in identical shape in every wave of automation since.

Dekker and Woods argue that quantitative "who does what" allocation cannot deliver coordination, "because the real effects of automation are qualitative: it transforms human practice and forces people to adapt their skills and routines". Underneath the lists sits the assumption they name, borrowing Erik Hollnagel's phrase, function allocation by substitution: the belief that new technology substitutes for people while "preserving the basic system while improving it on some output measures (lower workload, better economy, fewer errors, higher accuracy, etc.)".

The prediction it makes, which is testable#

The myth is not a complaint about optimism. It generates a specific prediction: allocating a function creates new functions for the other party that did not exist before. Dekker and Woods give the mundane example, "typing, or searching for the right display page". In an AI deployment the equivalents are writing the prompt, checking the citation, deciding whether the output is in scope, and holding the thread across a conversation that has no memory of yesterday.

That matters commercially, because almost every AI business case is written in substitution form. This took four hours, the model does it in ten minutes, therefore three hours fifty are saved. The prediction says the estimate is wrong in a known direction, and the estate's page on measurement timing records what happens when it is tested: self-reported speed gains and measured gains diverge, sometimes by tens of percentage points, and in at least one randomised trial developers estimated they were 20 per cent faster while being measured 19 per cent slower. That trial used early-2025 tooling and its authors withdrew the 19 as a current signal in February 2026; the divergence between belief and measurement is the part that stands.

The older version of the same argument#

Lisanne Bainbridge got there in 1983, from process control. Her Ironies of Automation observes that automating the routine parts of a task leaves the human with the hardest residue, monitoring and exception handling, while removing the routine practice that built the competence to handle it. Automation makes the remaining human role harder rather than easier. Dekker and Woods generalise the point beyond process control and name the assumption that keeps hiding it.

Cook, Render and Woods reached the same place from clinical medicine in 2000, describing the division of nursing work with less credentialed technicians. The economic benefit is real. Among the side effects, they write, "are restrictions on the ability of the individual nurse to anticipate and detect gaps in the care of the patients", because the nurse now has more patients to track and less of the direct contact from which anticipation was built. Three literatures, three decades, one finding.

What follows for anyone designing the work#

Dekker and Woods offer a replacement question rather than a better list: "The question for successful automation is not 'who has control over what or how much'. It is 'how do we get along together'." In practice that turns three habits around.

• Cost the new work. If a case claims a saving, it should name the verification, prompting and coordination that appears alongside it, and say who does it. A case that shows only subtraction has assumed the myth.
• Design the boundary, not the split. The failures concentrate where work passes between the parties, which is the subject of how humans and agents divide work across a process.
• Expect the residue to be harder. Bainbridge's point is the uncomfortable one for workforce planning: the tasks left after automation are the ones the designer could not automate, and they are typically the judgement-bearing ones, now performed by someone with less practice.

What the paper is, and is not#

This page rests on it, so its status needs saying precisely. Dekker and Woods is an argument, not a study. It contains no data, no participants and no experiment, and its supporting accident examples are cited rather than analysed. It should be read as the field's best articulation of a problem rather than as evidence about the size of that problem. The measured evidence for the prediction sits elsewhere, in the productivity literature and in the handover research, and is linked above.

Related SuperSkills research#

On the design consequence, how humans and agents divide work across a process and the delegation boundary map. On the residue, the invisible work of oversight, who supervises work they cannot do and deskilling. On the measurement, how to measure AI adoption properly and how long before you know if an AI investment worked.

Key research and primary sources

• Dekker, S. W. A. and Woods, D. D. (2002). MABA-MABA or Abracadabra? Progress on Human-Automation Co-ordination. Cognition, Technology & Work, 4(4), 240-244.
• Bainbridge, L. (1983). Ironies of Automation. Automatica, 19(6), 775-779.
• Cook, R. I., Render, M. and Woods, D. D. (2000). Gaps in the continuity of care and progress on patient safety. BMJ, 320(7237), 791-794.
• Model Evaluation and Threat Research (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The substitution myth is Dekker and Woods' term, function allocation by substitution is Hollnagel's, and the ironies of automation are Bainbridge's. The Fitts list is credited to Paul Fitts as editor of the 1951 National Research Council report rather than as its sole author. Nothing on this page is a SuperSkills coinage.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.