Autonomy Levels — References

Graded automation as studied long before language models, and the governance frameworks that ask about supervision directly.

Grading automation

• SAE J3016: Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles — the best-known graded autonomy scheme, and a useful model for why each level must be defined by who is responsible for what rather than by capability. Cited by name.
• Human and Computer Control of Undersea Teleoperators (Sheridan and Verplank, 1978) — the original levels-of-automation scale, from the computer offering no assistance through to acting entirely on its own and deciding what to report. Cited by name.

What happens to the supervising human

• Ironies of Automation (Lisanne Bainbridge, 1983) — automating the easy parts leaves people with the hardest residual task while eroding the practice needed to perform it; the foundational argument against assuming a human reviewer is an adequate safeguard. Cited by name.
• The Out-of-the-Loop Performance Problem and Level of Control in Automation (Endsley and Kiris, 1995), Human Factors 37(2), 381-394 — measured degradation in operators monitoring an automated process rather than performing it. Cited by name.

Governance expectations

• NIST AI Risk Management Framework — the govern, map, measure and manage structure, including expectations about documented roles and oversight of deployed systems.
• EU Artificial Intelligence Act, Article 14 (human oversight) — the legal requirement that oversight measures be commensurate with risk and genuinely effective rather than nominal. Cited by name; consult the official consolidated text for the operative wording.

Enforcing a policy outside the model

• Open Policy Agent — decoupled policy evaluation with declarative rules, the established pattern for authorisation decisions made outside the component being governed.

No verified reference is recorded for autonomy-grading schemes specific to language-model agents. Published proposals exist but none has become canonical, and none is cited here.