Levels of automation

Before anybody argued about AI, engineers had a scale for how much of a task a machine does, and a model that splits automation into four stages so you can automate one and leave another alone. It is the most directly usable allocation instrument in the literature and almost nobody in the AI conversation uses it.

The four stages, which is the part that transfers#

The scale alone invites a single dial, high or low, and that framing is the reason most adoption arguments are unproductive. The stage model says something more useful: a workflow can have heavy automation of gathering and analysing, and none at all of choosing, and that combination is a design rather than a compromise.

• Acquisition. What is collected and put in front of the person. Filtering here is invisible and shapes everything downstream.
• Analysis. Summarising, projecting, ranking. Where most current tools actually sit, whatever the marketing says.
• Decision selection. Choosing the option. The stage organisations think they have withheld and frequently have not, because a ranked list with a recommendation at the top is a choice made elsewhere.
• Action. Executing. The stage agents now touch, and the one where a stop control either exists or does not.

What the framework does not give you#

It contains no rule for choosing the level. Parasuraman and colleagues propose evaluating a candidate design against human performance consequences, mental workload, situation awareness, complacency and skill degradation, which is a set of criteria rather than an algorithm. The choice stays a judgement, made by people, which is the recursion this whole territory keeps producing.

It was also built for deterministic control systems: aircraft, plants, undersea vehicles. Whether it transfers cleanly to stochastic generative systems is an open question that is usually assumed rather than argued. A language model does not sit at a fixed level; it drifts between analysis and decision selection depending on how the prompt is written and how the output is read, which is a property the 1978 scale has no vocabulary for.

Why it belongs in a discussion about judgement#

Because it turns "how much AI" into four separate questions with four separate answers, and because the costs it names, situation awareness, complacency and skill degradation, are the costs this research is about. Bainbridge's Ironies of Automation made the argument in 1983 in four pages: automate the easy parts and you leave the person the hardest residue, with less practice at it.

The estate's own instrument for the same job is the delegation boundary map, which asks the question stage by stage in the language of a leadership team rather than of a human-factors journal. Where a decision should stay human entirely, see keep, share, hand over.

Key sources

• Parasuraman, R., Sheridan, T. B. and Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man and Cybernetics, 30(3), 286-297.
• Sheridan, T. B. and Verplank, W. L. (1978). Human and Computer Control of Undersea Teleoperators. MIT Man-Machine Systems Laboratory.
• Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6).