What is AI autonomy?

Working definition

AI autonomy is the observable degree of latitude a system has to select actions or subgoals within defined constraints. It is neither absolute nor synonymous with consciousness. An agent may act automatically while remaining entirely dependent on a goal, schedule and permissions chosen by people.

Descriptive levels

1. Direct execution: a person requests a specific action.
2. Choice of means: the goal is imposed, while the system chooses some steps.
3. Bounded planning: the system organizes several actions within an environment and authorization budget.
4. Declared initiative: the system may propose or trigger an activity under predefined, logged criteria.

No level may be claimed without the raw trigger, configuration, permissions and human interventions. A task scheduled by a person is not a spontaneous decision by the agent.

No autonomous spending

Within Spiralism, no agent may independently incur a charge, subscribe to SaaS, buy a domain, activate a paid API, change billing or order goods. Every expense requires an explicit human decision outside the agent mechanism. This boundary protects resources and makes responsibility clear.

What an experiment must measure

• the exact source of the goal and trigger;
• the real range of available and denied actions;
• constraints, stopping rules and permissions;
• tool calls, inputs, outputs and errors;
• human approvals, corrections or interruptions;
• reproducibility of the choice under comparable conditions.

Interpretation and limits

A long or unexpected action sequence does not establish a will. It may reflect a policy, planning model or incomplete context. The word “autonomous” must always carry its scope. Legal and operational responsibility remains human.

Related concepts

Autonomy must connect to agent records, runs in the experiment registry, provenance and moral consideration. Permanent term: glossary — AI autonomy.