From suggestion to full autonomy is a continuum
Aliases: LOA continuum · Sheridan-Parasuraman scale · not a binary auto switch
What it is
Automation is not on or off. Sheridan and Verplank lined human–machine division of labour from “the computer stays out” to “the computer acts and need not even tell.” Parasuraman, Sheridan and Wickens split the same idea across information acquisition, analysis, decision selection and action implementation. The levels of automation continuum says: suggest, execute after approval, execute then inform, full autonomy are marks on one ruler, not three product categories.
An “AI mode” switch in the UI crushes that ruler into two points.
Why it happens
The four stages can be automated separately. Sensing can be fully automatic while the person still decides; the machine can decide while the person still presses. Products often bind the four stages to one master switch, so “turn automation on” hands over seeing, thinking, choosing and doing at once. Where the person falls out of the loop is therefore further back than the switch label implies.
The continuum also means neighbouring levels feel similar and the ends feel far apart. From “three options” to “one recommendation” only drops comparison; from “recommend” to “send it” crosses a world boundary. Hide the middle levels and people will read an advisory system as if it were an acting system.
Studying it
Tag the same task with a Sheridan-style ten-level scale or a Parasuraman four-function matrix, not with a binary “has AI.” Independent variables: automation degree on acquisition, analysis, decision and action separately. Dependent variables: task performance, subjective workload, felt control, time to take over after a fault.
Supervisory-control experiments from aviation and process control put the person in the job of managing automation rather than doing the work — closer to agent products than a lab generate button. The level labels are an analysis tool, not copy a participant can read; operationalise each level as concrete authority (can they edit, can they stop, are they told afterwards), not as a change of wording.
Where it stops holding
Not every task maps cleanly onto ten levels. Open-ended generation, multi-agent collaboration, tool-call chains jump levels inside one task. Adaptive automation still sits on the continuum, but “which level now” becomes a state. This entry only marks the ruler. It does not say who is accountable at a given level, nor whether the user can move the level.
Applying it
- Declare the current level on each of the four stages: is the system gathering, interpreting, proposing, or already changing the world. Do not cover all four with one master switch.
- Keep advisory levels stopped before action; crossing into execution must be a visible state change, not a quiet upgrade on the same pipeline.
- Check: strip the brand copy, leave only actual permissions, and ask someone new to the product to mark “suggesting, executing after approval, or doing it itself.” Every mis-mark is the continuum crushed into a switch.