Levels of Automation and the Case for Intermediate Automation¶

Status: established
Last updated: 2026-06-15
Sources: Kaber Endsley 1997 Intermediate Levels Of Automation
Tags: [levels-of-automation, supervisory-control, function-allocation, out-of-the-loop, situational-awareness, human-automation-teaming, process-control, automation, intervention-strategies, remote-operations]

Summary¶

Level of automation (LOA) is a function-allocation approach that divides the work of a control task between human and computer so that both stay involved, rather than handing as much as possible to the machine. Kaber and Endsley (1997) frame human supervisory control, monitoring, and passive information processing as three forms of out-of-the-loop (OOTL) performance, each carrying vigilance, complacency, situation-awareness, and skill-decay costs that worsen during failures. They reproduce the Endsley and Kaber ten-level LOA taxonomy, defined over four generic functions — monitoring, generating, selecting, and implementing — and summarise an experiment in a simulated dynamic control task showing that intermediate levels outperform both manual control and high automation. The pattern depends on which function the human keeps: computer aiding in implementation raised throughput, joint human-computer option generation cut overlooked tasks, recovery from failures was best when the human kept the implementation role, and comprehension improved when the human was relieved of strategy selection. The paper makes the safety and cost argument for intermediate automation in process control.

Body¶

Context¶

Kaber and Endsley (1997) is a conceptual paper with an embedded empirical summary, written for a process-safety audience and motivated by accidents attributed to operators losing awareness of system state — Three Mile Island, the 1989 US Air B-737 at LaGuardia, and similar aircraft cases (PDF p. 1, orig. p. 126). It extends the out-of-the-loop argument of Endsley and Kiris (1995) compiled in Out Of The Loop Performance Problem, generalising from a five-level scheme for a discrete decision task to a ten-level taxonomy for dynamic control. Within this knowledge base it is the supervisory-control reference for function allocation and intervention strategies, sitting alongside the use/misuse account in Use Misuse Disuse Abuse Of Automation and the complacency mechanisms in Automation Complacency And Bias.

Key Points¶

Three OOTL roles. The paper treats supervisory control, monitoring, and passive information processing as variants of the same problem: in each the operator is removed from direct, real-time control and so loses awareness and manual skill (PDF pp. 1–3, orig. pp. 126–128). Supervisors observe a computer controller and agree or disagree with it, but their low taskload under normal conditions erodes the skills needed for recovery. Monitors wait to detect critical events while scanning many indicators — a role humans fill poorly because of vigilance and complacency limits. Passive information processors accept or reject another controller's actions without acting on the process themselves, and perform poorly when intervention is finally required. The shared consequence is loss of situation awareness, to which the cited accidents are attributed.

LOA as function allocation. Level of automation allocates system functions to human and computer based on the capabilities of each under both normal and failure conditions, with the explicit aim of keeping both involved (PDF p. 3, orig. p. 128). Kaber and Endsley contrast this with a technological approach that assigns as much as possible to the machine purely on computer capability, which pushes the operator into one of the OOTL roles and produces exactly the performance problems at issue. LOA instead aims to maintain appropriate operator taskload, reduce errors from poor automated decision making by retaining human judgement, and still gain the throughput of computer processing.

The ten-level taxonomy. The Endsley and Kaber taxonomy defines ten levels over four functions — monitoring (scanning to perceive status), generating (formulating options or strategies), selecting (deciding on an option), and implementing (carrying it out) — each assigned to the human, the computer, or both (PDF p. 4, orig. p. 129):

The levels build on earlier schemes, notably the ten-point scale of Sheridan and Verplanck (1978) for teleoperation, but are organised by which of the four functions each party performs rather than by a single ordinal scale.

The experiment. Thirty university students performed a simulated dynamic control task adapted from Tulga and Sheridan (1980), eliminating coloured targets that moved toward a central deadline, across four ten-minute trials — two probing manual performance during simulated automation failures and two isolating SA via SAGAT freezes (PDF p. 4, orig. p. 129). The results are function-specific (PDF pp. 4–5, orig. pp. 129–130). Throughput (tasks addressed) was greatest at low-intermediate levels that added computer aiding to implementation — Action Support (2) and Batch Processing (3) beat Manual Control (1) and beat levels that aided strategy generation or selection instead. Overlooked tasks (target expirations) fell as automation rose, lowest under Rigid System (7) and Automated Decision Making (8), where option generation involved the computer. Recovery from automation failures was best when the human retained the implementation role and worst after levels that queued targets for computer processing (Batch Processing, Automated Decision Making), because operators focused on future tasks and missed the current failure. SA comprehension improved at levels that relieved the human of strategy selection, freeing attention for perceiving the system.

The safety and cost argument. Kaber and Endsley conclude that the distribution of control made a significant difference in the implementation and generation roles but not in selection, and that intermediate levels delivered the gains by combining computer data processing with human judgement (PDF p. 5, orig. p. 130). They argue these performance gains translate into safety gains — fewer overlooked events, faster recovery, less complacency and skill decay — and thence into cost reductions (less machine wear, waste, cleanup, and liability), with application to foundries, nuclear power, and petrochemical refining.

Conclusion¶

Kaber and Endsley (1997) hold that neither manual control nor high automation is optimal for dynamic control: manual control overloads the operator and leaves the most overlooked tasks, while high automation pushes the operator out of the loop and slows failure recovery. Intermediate levels of automation, allocating functions so that human and computer stay jointly involved, gave the best balance of throughput, error reduction, recovery, and situation awareness — but the specific level matters, because each function (monitoring, generating, selecting, implementing) trades off differently when shifted to the machine. The paper generalises the in-the-loop recommendation of Out Of The Loop Performance Problem into a design taxonomy and a process-safety rationale, making it the function-allocation reference for the supervisory-control and intervention-strategy themes of this knowledge base.

Related¶

• Out Of The Loop Performance Problem — Endsley & Kiris's foundational OOTL study and five-level control scheme that this taxonomy extends
• Use Misuse Disuse Abuse Of Automation — Parasuraman & Riley's patterns of automation use and human-centred automation
• Automation Complacency And Bias — the complacency and vigilance costs the LOA approach is designed to limit
• Active Control Vs Passive Monitoring Atc — the cost of relegating an operator to a monitor, in ATC
• Human In The Loop Automation Transparency — keeping the operator informed and involved in remote maritime supervision
• Multi Ship Remote Operations Workload Sa — workload and SA when one operator supervises many vehicles

References¶

Endsley, M. R. and Kaber, D. B. (1999) 'Level of automation effects on performance, situation awareness and workload in a dynamic control task', Ergonomics, 42(3), pp. 462–492. To be validated.

Endsley, M. R. and Kiris, E. O. (1995) 'The out-of-the-loop performance problem and level of control in automation', Human Factors, 37(2), pp. 381–394. doi: 10.1518/001872095779064555. endsley1995loop

Kaber, D. B. and Endsley, M. R. (1997) 'Out-of-the-loop performance problems and the use of intermediate levels of automation for improved control system functioning and safety', Process Safety Progress, 16(3), pp. 126–131. doi: 10.1002/prs.680160304. kaber1997intermediate

Sheridan, T. B. and Verplanck, W. L. (1978) Human and computer control of undersea teleoperators. Cambridge, MA: MIT Man-Machine Systems Laboratory, Technical Report. To be validated.

Tulga, M. K. and Sheridan, T. B. (1980) 'Dynamic decisions and work load in multitask supervisory control', IEEE Transactions on Systems, Man, and Cybernetics, SMC-10(5), pp. 217–232. To be validated.

Open Questions¶

• The empirical summary reports an experiment whose full methods and statistics are in the Endsley and Kaber (1999) Ergonomics paper, which is not yet held in RAW; the locator-level findings here should be reconciled with that primary when it is ingested.
• The taxonomy assigns functions categorically (human, computer, or both) but does not specify how the "both" cases share control moment-to-moment, which is where adaptive and adjustable automation schemes differ.
• The safety-to-cost translation is argued rather than measured; no field data on incident-rate or cost reductions from intermediate LOA in process plants is presented.