Can automation and human control increase together?

The paper argues that the long-standing one-dimensional levels-of-automation model — running from Sheridan and Verplank's (1978) ten levels through the US Society of Automotive Engineers' six levels for self-driving cars — wrongly assumes "increases in automation must come at the cost of lowering human control." Its Human-Centered Artificial Intelligence (HCAI) framework proposes decoupling the two into independent axes, so that "achieving high levels of human control and high levels of automation is possible" in what it calls the upper-right quadrant of a two-dimensional grid. Reaching that quadrant, the paper states, is "more likely to produce designs that are Reliable, Safe & Trustworthy (RST)."

The reasoning draws on prior critics who had already catalogued the costs of chasing autonomy alone on a single dial — Bainbridge's "ironies" (1983), Bradshaw et al.'s "myths" (2013), Endsley's "conundrums" (2017), Hancock's "paradoxes" (2017) — converging on the point that "humans have to spend more effort monitoring autonomous computers because they are unsure of what it will do." Rather than abandon automation, the paper reassigns quadrants by task type: the lower-right (high automation, low control) suits rapid, well-understood actions such as airbag deployment or anti-lock brakes, where "there is no time for human intervention"; the upper-left (high control, low automation) suits mastery-building tasks like bicycle riding or piano playing; the upper-right is reserved for "poorly understood and complex tasks with varying contexts of use," illustrated with a patient-controlled-analgesia device that lets a patient trigger dosing while sensors and machine learning choose safe amounts and a hospital control center monitors "hundreds of PCA devices" for audit and improvement.

This reframes a question the vault otherwise treats as a spectrum problem. Can we design AI autonomy separately from capability? also rejects a single autonomy dial, but indexes its five levels to the role a user holds at a given moment; this paper instead treats control and automation as two separately tunable design variables that can both be pushed toward maximum, naming the assumed trade-off itself — not just the coarseness of the levels — as the design error. What UX principles do workplace users want in AI agents? finds users rank human control alongside reliability and safety as top design principles; this paper supplies an underlying design logic for why those qualities need not compete with higher automation.

The excerpt is a position and framework paper, not an evaluation: no instrument for measuring "levels of control and autonomy" yet exists, which the paper itself names as "an important research direction," and the PCA and self-driving-car illustrations are worked examples rather than tested deployments. It also does not resolve Does granting agents more autonomy undermine human oversight?' empirical worry that oversight capacity erodes with use — this paper names "de-skilling effects" and "the difficulty of remaining vigilant when user actions become less frequent" only as open limitations, not problems its two-axis framework is shown to solve. What follows at the strength of a framework paper: the two-dimensional model is a design target to aim for, not yet a demonstrated route for getting there.