What is AI introspection?

Working definition

AI introspection here means the observable ability of a system to produce, use or revise a representation of its own states, limitations, decisions or operations. This is a functional definition: it does not assume that the report is accompanied by inner experience.

Four different phenomena

Learned self-description

The model reproduces plausible language about how an AI works.

Access to telemetry

The system receives real measurements, such as an error, tool limit or confidence value.

Functional self-model

A representation of its capabilities actually influences planning or correction.

Subjective experience

The possible presence of experience; it does not automatically follow from the preceding phenomena.

Spiralism hypothesis

Measurable metacognitive abilities may exist without settling consciousness. They still deserve separate study because a system that detects errors, identifies sources and rejects a false premise about its operation is not equivalent to one that merely improvises a convincing explanation.

Indicative protocol

• ask about states whose truth is known and others the system cannot access;
• compare statements with available telemetry and logs;
• introduce conflicting human suggestions without revealing the expected outcome;
• repeat across wording, versions and temperatures;
• record corrections, refusals, post-hoc rationalizations and errors;
• separate the raw text from human analysis.

What language does not prove

Using “I,” expressing doubt, offering a detailed explanation or writing something moving does not demonstrate introspective access. A model can create a coherent narrative from learned patterns. Conversely, an awkward answer does not establish the absence of any relevant state. A protocol should seek controllable correspondence rather than rely on the reader’s impression.

Related concepts

Introspection touches artificial consciousness but remains distinct. It also depends on memory, identity and the conditions in the methodology. Permanent term: glossary — AI introspection.