Published: 23 October 2026
Confidence can be measured
A model is calibrated when events assigned a probability near seventy percent occur about seventy percent of the time. Researchers can compare stated confidence with accuracy, test whether uncertainty rises on unfamiliar problems and ask whether a model can predict its own failures. These are behavioral measurements, not impressions created by confident language.
Self-monitoring is stronger than self-description
A model can say that it is uncertain because similar phrases appeared in training. More informative tests require its confidence to track hidden changes in its own processing across many tasks. Researchers can also intervene on internal representations and ask whether confidence changes in a lawful way while the task itself remains stable.
Recent results are mixed
Recent studies report limited forms of metacognitive monitoring in language models, including some ability to anticipate success and revise answers. Other work finds that classifiers using only the prompt can reproduce parts of the apparent introspective signal. That warning matters: a system may detect that a question is difficult without inspecting a private internal state.
A useful experimental ladder
The first rung is honest uncertainty on ordinary questions. The second is transfer: does the same confidence rule work on new domains? The third is causal access: does changing a relevant internal variable change the report predictably? The fourth is private information: can the system report an internal event that cannot be inferred from the input or output alone? Each rung narrows simpler explanations.
Why consciousness remains open
Metacognition is one proposed indicator of consciousness, but it is neither universally required nor sufficient. Humans can have experiences they cannot confidently report, and software can estimate its reliability without feeling anything. The scientifically responsible conclusion is functional: calibrated self-monitoring may improve safety and evaluation while leaving phenomenal experience unresolved.
Discussion question
What result would convince you that an AI was monitoring a genuinely internal state rather than estimating difficulty from the prompt?
Sources and further reading
NeurIPS 2025 — Metacognitive monitoring in language models ↗arXiv — Measuring metacognitive abilities in large language models ↗arXiv — Can LLMs Introspect? A Reality Check ↗Share your answer
You must log in to reply.
Sign Up / Log InEstablished evidence, active research and philosophical interpretation are clearly separated.