Level of Automation (LoA)

The foundational taxonomy for human-automation interaction, ranging from Level 1 (human does the whole job) to Level 10 (computer does the whole job and ignores the human). Key decision points include:

• Levels 1-4: The human retains authority over action selection
• Levels 5-6: The computer executes an action only after human approval
• Levels 7-9: The computer executes autonomously but must inform the human
• Level 10: Full autonomy with no human notification This scale explicitly separates information acquisition, analysis, decision selection, and action implementation.

An industry-specific taxonomy adopted globally for vehicle automation, defining six levels from Level 0 (No Automation) to Level 5 (Full Driving Automation). The critical distinction occurs between:

• Level 2: The driver must constantly supervise the dynamic driving task
• Level 3: The system handles all driving under limited conditions, but the human must be ready to intervene immediately upon request
• Level 4: No human intervention required within a defined operational design domain
• Level 5: Unconditional, driverless operation everywhere This taxonomy is legally binding in many jurisdictions and directly informs liability frameworks.

Modern LoA frameworks decompose automation not as a single axis but across distinct functional dimensions:

• Information Acquisition: Automating the sensing, filtering, and prioritization of data
• Information Analysis: Automating prediction, inference, and situation assessment
• Decision Selection: Automating the choice between alternative courses of action
• Action Implementation: Automating the physical or digital execution of the chosen action A system may operate at Level 8 for analysis (fully autonomous threat detection) while remaining at Level 3 for decision selection (requiring human authorization for countermeasures). This granularity prevents mode confusion.

Static LoA assignments are insufficient for complex, dynamic environments. Sliding autonomy allows the system to dynamically shift control along the automation spectrum based on:

• Task criticality: Escalating to human control when risk thresholds are breached
• Operator cognitive load: Increasing automation when the human is overloaded
• System confidence: Handing off to the human when the model's prediction confidence falls below a calibrated threshold
• Environmental complexity: Reducing autonomy in unmapped or degraded conditions This requires a robust deferral policy and seamless handoff mechanisms to prevent control gaps.

LoA taxonomies serve as the technical substrate for legal accountability. The EU AI Act and similar regulations use automation levels to determine:

• Whether a system qualifies as high-risk and requires conformity assessment
• The degree of meaningful human control mandated by law
• The allocation of liability between the manufacturer, deployer, and human operator For example, a system classified at Level 6 (computer executes after human approval) places liability on the approving human, while Level 9 (computer executes and informs, human can veto) shifts significant liability to the system designer for ensuring adequate notification latency and interface clarity.

Each LoA level imposes distinct requirements on the human-machine interface (HMI) to prevent automation bias and mode confusion:

• Levels 1-4: The interface must make the system's information analysis transparent so the human can make informed decisions
• Levels 5-6: The interface must clearly present the proposed action, the rationale, and a time-bound approval mechanism
• Levels 7-9: The interface must provide salient, non-intrusive notification of actions taken, with an unambiguous override mechanism
• Level 10: The interface must clearly communicate that no human intervention is possible, preventing futile control attempts Poor HMI design at any level leads to automation complacency and mode confusion, two leading causes of automation-related incidents.