In 1951, Paul Fitts edited a report on human engineering for air-navigation and traffic-control systems. One of its lasting artifacts was a deceptively simple allocation device that came to be known as the Fitts list: what humans are better at and what machines are better at (Fitts, 1951). In shorthand, it became HABA-MABA: Humans Are Better At, Machines Are Better At.
For decades, HABA-MABA offered a clean way to think about human-machine systems. Humans were better at judgment, improvisation, recognizing meaningful patterns, and dealing with novelty. Machines were better at speed, precision, repetition, measurement, calculation, and operating under conditions that would bore, fatigue, or endanger people.
That model was never perfect. It was more of a heuristic than a law of nature. It also encouraged a particular kind of complacency. Give the machine the things it does well, leave the human with the things the machine cannot handle, and assume the system will somehow balance itself. Bainbridge (1983) showed why that assumption can fail. When automation removes people from routine operation but leaves them responsible for rare and difficult interventions, the human role can become harder rather than easier. The operator monitors more, practices less, and is expected to recover the system when context is thinnest.
That critique remains applicable and relevant in the current AI era because the old trap now appears one layer earlier in the workflow. The human is no longer only being asked to step in after automated execution fails. The human is increasingly being asked to step in after machine-mediated interpretation has already shaped the nature of the problem.
Classical automation and contemporary AI are both machine-mediated, but they are not the same kind of thing. Many machine components are better understood as bounded systems. They may be statistical, heuristic, or probabilistic internally, but they operate inside explicit constraints, validation rules, state transitions, and procedural checks. AI systems sit in a different place when they are used to classify, rank, summarize, detect patterns, or generate candidate explanations from messy inputs. Some of those systems are generative, such as the current generation of LLMs, while others are not. A fine-tuned classifier, a recommendation engine, a risk-scoring model, and a large language model do not behave identically, but they can all shape interpretation before a person reaches judgment. That additional layer suggests a third term:
HABA: Humans Are Better At, MABA: Machines Are Better At, AIABA: AI Is Better At
Treat the acronym as an object of discussion, rather than a formal attempt to extend the old model. It names the part of the workflow where AI systems, generative or not, increasingly participate in interpretation rather than only in execution.
The underlying issue is function allocation. The binary construct of “What are humans better at, and what are machines better at?” increasingly has more nuance along the lines of “Where should functions like judgment, interpretation, verification, and accountability reside?”
The previous split may no longer hold
The original HABA-MABA formulation worked best when “machine” meant something closer to mechanical, electromechanical, or tightly bounded computation (Fitts, 1951; de Winter & Dodou, 2014). Machines could calculate, regulate, measure, sort, repeat, and execute. Humans could perceive, adapt, judge, and improvise.
Those boundaries were always porous, but they were at least understandable. A radar display, a flight control system, a database, a sensor network, and a scheduling algorithm all occupy different places in a system, but they share an important property. Their behavior is usually bounded by explicit design. They may be complex and they may fail, but their outputs can usually be traced through rules, states, inputs, and procedures.
Generative AI complicates that most visibly. It is not the only system in this category, but it is the limiting case because it can present interpretation in fluent language rather than only in labels, scores, or rankings. It is not just a faster calculator or a better control loop, it is an interpretive system trained on patterns. It can operate across ambiguity, but not with the same kind of boundedness that procedurally constrained systems provide. It can produce analysis that is often persuasive, but not because it possesses truth. It can summarize, classify, infer, and explain, but those outputs remain probabilistic artifacts that need context and validation. Those traits make generative AI capable, but it also makes it easy to misallocate responsibility.
If we treat generative AI as simply a better machine, we may assign it deterministic authority it does not possess. If we treat it as a junior human, we may assign it judgment, intent, or accountability it should not bear. Both mistakes are common because generative systems can mimic the surface features of human work while still being machine components inside a larger system.
Where the interpretive layer fits
AI systems are often better than humans at scale-bound interpretive tasks. They can search across large bodies of text, identify recurring themes, normalize inconsistent phrasing, assign labels, rank likely matches, generate summaries, cluster related material, and translate between natural language and structured representations. They can make a large corpus legible. They can propose hypotheses that a person may not have surfaced manually, at least in a reasonable amount of time. They can turn a vague question into a set of plausible analytical paths. Generative AI is the most visible current case because it performs that role in language, often fluently enough to sound like judgment rather than preprocessing.
These systems also work well as connective layers. A person may describe a goal in ordinary language. An AI system may help translate that goal into a query, workflow, schema, prompt, checklist, or code scaffold. A machine system can then execute the resulting procedure, producing outputs that are logged, tested, measured, and reviewed.
AI systems are not better at owning the mission or deciding acceptable risk. They are not better at bearing responsibility or knowing whether the question being asked is the right question. They are not better at determining whether an answer is ethical, proportionate, or appropriate in context. They can help with each of those tasks, but helping is not owning.
Humans are better at purpose, accountability, values, institutional context, and deciding what matters. Machine systems are better at execution, measurement, repeatability, persistence, and enforcing constraints. AI systems are better at interpretation (in context), synthesis, semantic search, pattern discovery, classification, ranking, and generating candidate explanations across messy inputs and large corpora, especially when time is a constraint. The challenge is to keep those roles from bleeding into one another in ways that hide risk.
A three-part model
A modern allocation model looks less like two columns and more like a distribution of responsibilities.
Humans define the mission, set the risk tolerance, adjudicate exceptions, approve consequential actions, and own the outcome. AI systems help interpret evidence, generate options, summarize context, detect patterns, rank possibilities, and explain alternatives. Machine systems execute specified workflows, maintain state, enforce constraints, record provenance, and provide repeatable measurement. An updated approach could be to allocate authority to humans, interpretation to AI systems, and execution to machine systems, with verification crossing all three.
Verification is the missing piece. It cannot live only in the human layer. Humans get tired, distracted, overloaded, and overly trusting of systems that sound confident. It cannot live only in the AI layer because model output, whether scored, ranked, classified, or generated, may still be wrong in ways that are difficult to detect. It cannot live only in the machine layer because procedural checks can validate compliance but not meaning.
Verification has to be designed across the system. In practice, that means a few things. It means preserving provenance at every intermediate step rather than only at the final output. It means making the chain of evidence inspectable in a form a reviewer can actually interrogate. It means defining thresholds that trigger human review rather than waiting for obvious failure. It means exposing uncertainty, disagreement, or competing interpretations instead of collapsing them into one polished answer. It means running domain-specific validation checks before action, even when the generated narrative appears coherent.
For example, an AI system might summarize a set of reports, classify planning documents by likely risk, or rank likely infrastructure dependencies. A machine system might preserve the source documents, track citations, validate extracted entities against known schemas, and record each transformation. A human analyst might decide whether the dependencies matter for the mission, whether the evidence is sufficient, and what action should follow. A failed verification in that chain is not only a hallucinated sentence. It may also be a missing citation, an untraceable transformation, a schema violation, a brittle classifier, an unsupported inference, or a reviewer who cannot reconstruct why the system reached its recommendation.
The AI doesn’t own the conclusion, the machine system doesn’t own the judgment, and the human doesn’t have to manually inspect every token generated along the way. Each part contributes something, and each part constrains the others.
The new irony of automation
Bainbridge’s (1983) old warning still applies. Automation can leave humans with harder work than before. AI can intensify that problem because it does not merely automate action, it also automates parts of interpretation.
That creates a new version of the monitoring trap. A person may no longer be watching dials or gauges. They may instead be watching summaries, explanations, recommendations, and generated narratives. The work feels more cognitive, but the trap is similar. If the AI layer handles the routine interpretation, the human may be asked to intervene only when the case is ambiguous, high-risk, or already off the rails. That is ceremonial oversight rather than meaningful human control.
That is not to say we should keep AI away from important work. The work should be designed so the human remains meaningfully engaged with the evidence, the assumptions, and the decision criteria. That includes ensuring AI exposes uncertainty rather than smooths it over. It may also mean requiring the system to retain competing interpretations. It may mean designing around source material and decision points rather than summaries, and it may mean giving the human fewer outputs but better ways to interrogate them.
The hard part is when the layers conflict and the model says one thing, the machine system validates only part of it, the log is incomplete, and the human has to decide anyway. That is the point at which the system either preserves judgment or collapses into ritual. Under time pressure, the primary question is not whether a human remains nominally in the loop, but whether the system has preserved enough evidence and reversibility to assert human authority.
Why this matters in geospatial
Geospatial work is a good example because it already sits at the junction of measurement, interpretation, and consequence. Sensors collect, databases persist, coordinate systems align measurement. Software renders, analyzes, transforms, and distributes. People still decide what the result means in relation to land, infrastructure, jurisdiction, policy, and risk.
The relevance is not only that the problem is visible there. It is that geospatial outputs carry an implicit authority that comes from their association with measurement. A map, a coordinate, a hotspot, or an overlay result often feels like it reports the world rather than interprets it. That makes the boundary between the interpretive layer and the machine layer harder to see in geospatial work than in a text summary, where the act of reduction is obvious.
AI enters close to the point where technical output becomes institutional meaning. It can summarize planning documents, translate user intent into spatial queries, generate code, classify imagery, rank likely anomalies, or connect spatial evidence with non-spatial context. A model may identify a road, a building, a damaged roof, or a suspicious pattern. A machine system may store the geometry, measure the area, run the overlay, and enforce the schema. But a person, or an accountable institution acting through people, still has to decide what level of confidence is sufficient and what action is justified.
A concise illustration is PredPol, a place-based predictive policing system that generated algorithmically produced patrol boxes on a map (Brayne, 2020), representing areas where the system assessed crime was likely. PredPol made that dynamic unusually visible. Officers followed the patrol boxes because the boxes looked like they reported where crime was, not where a model guessed crime might be. In that setting, the geospatial version of ceremonial oversight was not a signature at the end of a document. It was a GPS trace inside a polygon (Brayne, 2020).
That is not unique to geospatial. But spatial representation can disguise interpretation more effectively than many other output formats, which makes the problem easier to institutionalize and harder to notice.
From task allocation to responsibility
HABA-MABA was a task allocation model. It asked what humans and machines were each better suited to do. That was a reasonable question for its time, and it still works as a starting point (Fitts, 1951; de Winter & Hancock, 2015). Later work on levels of automation made that problem more explicit by distinguishing among information acquisition, analysis, decision, and action rather than treating automation as a single block (Parasuraman et al., 2000). This essay is not a replacement for that literature. It is an argument that the analysis layer now deserves renewed attention because AI systems increasingly occupy it in ways that are persuasive to human decision-makers.
In AI-enabled systems, the central issue is how candidate interpretations are generated, how they are adjudicated, how evidence is maintained, how results are verified, and where accountability resides when the system is wrong. That moves past allocation into responsibility.
One way to visualize the shift is with a small RACI chart that keeps three distinct actors visible: human, AI, and machine. There is a practical distinction between second and third rows. Producing a candidate interpretation is not the same thing as adjudicating ambiguous evidence. The first is the creation of an artifact that may inform judgment. The second is the judgment act itself.
| Activity | Human | AI | Machine |
|---|---|---|---|
| Define mission and acceptable risk | A/R | C | I |
| Produce candidate interpretation or summary | A | R | C |
| Adjudicate ambiguous evidence | A/R | C | C |
| Execute specified workflow | A | C | R |
| Preserve provenance and audit trail | A | C | R |
| Approve consequential action | A/R | I | I |
This is intended as an illustration, not a universal template. The assignments will vary by system, domain, and risk level. In high-stakes settings, the human share of responsibility should usually expand, not contract. The point is simply to keep all three actors visible. AI does not replace the machine column. A machine system may be responsible for execution, logging, validation, access control, and persistence. In practice, that execution should be as deterministic as the domain allows, even when the broader system around it is only bounded rather than strictly deterministic. AI may be responsible for producing an interpretive artifact, such as a summary, classification, suggested workflow, ranking, or candidate explanation. A finer-grained RACI chart may show AI responsible for more discreet tasks. Humans remain accountable for mission, risk, approval, and consequence.
The table also has limits. RACI was built for actors that can bear responsibility in the ordinary organizational sense. AI cannot. So the table is best read as a visualization of where work is being done and where accountability must not be allowed to drift, not as a literal claim that a model is a member of the org chart.
In that sense, HABA-MABA-AIABA becomes less a list of capabilities and more a map of responsibility. AI can produce an interpretive artifact without owning the accepted interpretation. Machine systems can be responsible for execution without owning the judgment. Humans can be accountable for the outcome without manually performing every intermediate step. The primary considerations are how responsibility is distributed, how evidence is preserved, and how a consequential judgment is kept from becoming ceremonial oversight.
What must remain a human decision? What can be interpreted by AI but verified elsewhere? What should be executed only by deterministic systems? What evidence has to be preserved for review? What assumptions have to stay visible? What exceptions require human judgment rather than nominal human approval? Which parts of the workflow should be reversible, auditable, or constrained by policy?
These questions are less tidy than the old HABA-MABA list. They are closer to the modern systems we are now building. AI does not eliminate the human-machine allocation problem, but it makes that problem more explicit. It forces us to distinguish among judgment, interpretation, execution, and accountability. It also reminds us that capability is not the same thing as responsibility.
Accountability here is not as settled in practice as many governance documents imply. Law, organizational design, and institutional blame assignment are all still catching up. But that does not change the design problem. Advanced systems, even when they can produce shockingly human-like output, do not remove the need for someone to own the consequence.
A good AI-enabled system should not ask whether humans or machines are better in the abstract. It should ask what kind of work is being done, what kinds of failure are possible, what happens when the parts conflict, and where authority belongs.
That may be the real update to Fitts for the current AI era. Humans decide what matters, AI systems help produce candidate interpretations, machines execute what has been specified and preserve the evidence chain. The task is to keep those layers aligned before oversight becomes ceremonial.
References
Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8
Brayne, S. (2020). Predict and surveil: Data, discretion, and the future of policing. Oxford University Press. https://doi.org/10.1093/oso/9780190684099.001.0001
de Winter, J. C. F., & Dodou, D. (2014). Why the Fitts list has persisted throughout the history of function allocation. Cognition, Technology & Work, 16(1), 1–11. https://doi.org/10.1007/s10111-011-0188-1
de Winter, J. C. F., & Hancock, P. A. (2015). Reflections on the 1951 Fitts list: Do humans believe now that machines surpass them? Procedia Manufacturing, 3, 5334–5341. https://doi.org/10.1016/j.promfg.2015.07.641
Fitts, P. M. (Ed.). (1951). Human engineering for an effective air-navigation and traffic-control system. National Research Council, Division of Anthropology and Psychology, Committee on Aviation Psychology.
Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics – Part A: Systems and Humans, 30(3), 286–297. https://doi.org/10.1109/3468.844354
Header image: Petar Marjanovic, CC BY-SA 3.0 https://creativecommons.org/licenses/by-sa/3.0, via Wikimedia Commons
