The distinction between correlation and causation sits at the heart of the research. Prior work has established that language models can report calibrated confidence estimates: when a model assigns a high probability to its own answer, that answer is more likely to be correct, and when the model expresses doubt, errors become more frequent. But such correlations leave open an ambiguity. A model might compute an internal confidence signal and then act on it, or the confidence estimate might simply be a by-product of the same computational processes that generate the answer, exerting no independent influence on behaviour. Disentangling the two requires more than observational measurement; it requires intervention.

That is precisely the approach taken in the new work. Rather than passively recording what models say about their confidence, the researchers manipulated the internal states associated with confidence and observed whether behaviour changed as a result. By intervening directly on the representations that carry confidence information within a model’s activations, the team could test whether those representations play a genuine causal role in driving the model’s subsequent outputs. The logic mirrors classic interventionist methods in neuroscience and psychology, where researchers establish that a neural signal causes behaviour by perturbing it and watching what happens, rather than merely noting that the two tend to co-occur.

The results, according to the study, show that when confidence-related internal states are amplified, models behave in ways characteristic of high-confidence systems: they commit more firmly to their answers, become less likely to hedge or express uncertainty, and shift their response patterns in measurable ways. When those same states are suppressed, the opposite occurs. Models become more tentative, more prone to qualify their statements, and more likely to distribute probability across multiple candidate answers. Crucially, these behavioural shifts occur without altering the model’s underlying knowledge or the input prompt, indicating that the confidence signal itself is doing causal work in the decision-making pipeline.

The implications extend well beyond academic curiosity. If a language model’s confidence genuinely drives its behaviour, then the way a model expresses certainty is not a cosmetic layer added on top of reasoning—it is an integral part of how the system decides what to say. This reframes long-standing concerns about overconfident AI. Models that state falsehoods with unwavering assurance may do so not because they lack a mechanism for detecting their own errors, but because the confidence signal that should temper their claims has been miscalibrated or overridden. Conversely, models that hedge excessively may be suffering from systematically deflated internal confidence, even when their answers are largely correct.

The findings also speak to a growing body of research on introspection in artificial systems. One of the most debated questions in AI safety is whether language models have meaningful access to their own internal computational states—whether, in some functional sense, they know what they know. Self-report alone is a weak guide, because models can be trained to produce plausible-sounding confidence statements without those statements reflecting anything real. The new causal evidence strengthens the case that something substantive underlies these reports: the confidence a model expresses is connected, through identifiable internal mechanisms, to the behaviour it subsequently produces. That connection is exactly what one would expect if the model is, in a limited but genuine sense, monitoring and acting on its own uncertainty.

Methodologically, the study represents a maturing of interpretability research. Early work in the field focused on locating where particular concepts are encoded in a network’s layers, producing maps of representation that were suggestive but often causally inert. The field has since shifted toward intervention-based techniques—activating, suppressing, or editing internal features and measuring the downstream consequences. The new research applies this interventionist toolkit to metacognition, the capacity of a system to represent its own cognitive states. Demonstrating that confidence representations pass causal intervention tests places them in the same evidentiary category as other well-established internal features, such as those implicated in factual recall and instruction following.

For practitioners deploying language models in high-stakes settings, the research carries practical weight. Calibration techniques, uncertainty quantification, and abstention mechanisms—methods that allow models to say when they do not know—are increasingly treated as essential safety infrastructure for medical, legal, and scientific applications. The new findings suggest these mechanisms are not merely bolted-on output filters but are entangled with the model’s core decision processes. Interventions that improve calibration at the level of internal confidence representations could therefore propagate through the entire behavioural stack, making the model more honest about its limits in a way that end-to-end fine-tuning alone may not reliably achieve. At the same time, the causal link cuts both ways: anything that distorts a model’s internal confidence—adversarial inputs, distribution shift, or aggressive post-training—could distort not just its stated confidence but its actual behaviour in hard-to-predict ways.

The study also raises questions that the authors and the broader field will now need to address. Confidence in language models is not a single scalar; it is distributed across layers, tokens, and computational pathways, and different components of a network may encode different, sometimes conflicting, estimates of reliability. Understanding which of these representations are causally potent—and which are epiphenomenal—will be essential for building systems whose self-assessments can be trusted. There are also open questions about generality: whether the causal role of confidence holds across model scales, architectures, and training regimes, and whether the same mechanisms govern other forms of self-monitoring, such as a model’s assessment of the safety or harmfulness of its own outputs.

What the research establishes, for now, is a conceptual milestone. Language models are not black boxes that merely emit text; they are systems whose internal states, including states that encode confidence, play demonstrable causal roles in shaping behaviour. That realization moves the field closer to a science of machine metacognition—one in which a model’s relationship to its own uncertainty can be measured, verified, and engineered rather than assumed. As language models are entrusted with ever more consequential decisions, knowing that they use their confidence to drive what they do is a crucial step toward ensuring that what they do remains worthy of the confidence we place in them.

Subject of Research: Causal role of internal confidence representations in driving the behaviour of large language models

Article Title: Causal evidence that language models use confidence to drive behaviour

Article References: Kumaran, D., Daw, N., Osindero, S., Veličković, P., & Patraucean, V. (2026). Causal evidence that language models use confidence to drive behaviour. Nature Machine Intelligence. https://doi.org/10.1038/s42256-026-01293-x

Image Credits: AI Generated

DOI: 10.1038/s42256-026-01293-x

Keywords: large language models, confidence, causality, interpretability, AI safety, metacognition, calibration, uncertainty, neural networks, machine learning, Nature Machine Intelligence, self-monitoring