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

Metacognition is often described as “thinking about thinking.” It is the mental faculty that allows a person to sense when they are unsure of an answer, to judge the accuracy of a memory before acting on it, and to adjust behavior accordingly. In laboratory settings, metacognition is typically measured by asking participants to perform a perceptual or cognitive task and then to rate their confidence in each decision. Researchers then compute a metric known as metacognitive sensitivity — essentially, how well a person’s confidence ratings track their actual accuracy. A person with strong metacognitive sensitivity is confident when correct and doubtful when wrong; a person with impaired metacognition loses this correspondence, becoming unable to distinguish reliable internal signals from unreliable ones.

This capacity matters enormously in everyday life and, according to a growing body of addiction research, it may matter especially in the cycle of substance dependence. The predominant neurocognitive models of addiction hold that chronic drug use degrades the neural systems responsible for self-regulation, tipping the balance toward habitual, compulsive drug-seeking at the expense of deliberate, goal-directed behavior. If metacognitive monitoring is part of that self-regulatory architecture, then deficits in it could help explain one of the most perplexing features of addiction: the persistence of drug use despite obvious and repeated negative consequences. A person who cannot accurately appraise their own cognitive state may also be less equipped to appraise the mounting costs of their behavior, or to trust their own resolve when attempting abstinence.

Earlier studies in the field had reported reduced metacognitive accuracy in individuals with cocaine use disorder, but the reliability of those findings remained an open question. Small sample sizes, heterogeneous participant groups, and inconsistent methods for quantifying metacognition all left room for doubt about whether the observed deficits were genuine features of the disorder or artifacts of particular experiments. The new study was designed specifically to address this uncertainty through a confirmatory approach, meaning that it set out to test the previously reported association under carefully controlled and preregistered analytical conditions, using refined behavioral metrics that separate metacognitive sensitivity from the underlying task performance itself.

The research team assessed participants with cocaine use disorder alongside well-matched comparison participants, evaluating metacognition across cognitive tasks while also gathering detailed information about recent patterns of cocaine consumption. This dual focus proved crucial. The results indicated that metacognitive impairment was not simply a fixed trait of the disorder but was dynamically linked to the recency and intensity of drug use. Participants whose recent cocaine consumption was higher showed more pronounced deficits in their ability to monitor the accuracy of their own decisions, while those with longer periods of reduced use displayed comparatively better metacognitive performance. In other words, the internal compass that tells us how much to trust our own minds appears to be dulled by recent exposure to the drug and may partially recover as use declines.

Technically, the study relied on signal-detection-theoretic frameworks to disentangle the components of metacognitive performance. Simple accuracy on a task and confidence in one’s answers can be confounded — a participant who performs poorly on everything might also report uniformly low confidence, without any true metacognitive deficit. Modern metrics such as meta-d, which estimates metacognitive sensitivity relative to task performance, allow researchers to ask whether an individual’s confidence judgments are informative above and beyond their raw accuracy. By applying such analyses in a confirmatory framework, the researchers could demonstrate that the impairment in cocaine use disorder specifically involves the monitoring layer of cognition rather than a generalized cognitive slowdown or lack of engagement with the tasks. This distinction carries theoretical weight, because it points to disruption in neural circuits — often implicating the prefrontal cortex and its connections to parietal and limbic regions — that are thought to support self-directed evaluation.

The prefrontal cortex has long been identified as a region vulnerable to the effects of chronic cocaine exposure. Neuroimaging and neuropsychological studies have repeatedly documented alterations in prefrontal gray matter volume, white matter integrity, and metabolic activity in people with cocaine use disorder. Because these same frontal networks are central to metacognitive processing in healthy individuals, the convergence of evidence is striking: the brain systems that generate our sense of confidence and uncertainty are precisely those most affected by prolonged stimulant use. Recent drug use, by modulating dopaminergic signaling and prefrontal function acutely as well as chronically, may therefore exert a double burden — degrading the machinery of self-monitoring at the moment when recovering individuals most need it.

The clinical implications of this work are potentially far-reaching. Addiction treatment typically depends on the patient’s ability to recognize lapses in control, anticipate high-risk situations, and honestly evaluate progress. Metacognitive impairment could undermine each of these processes, explaining why some individuals struggle with self-reported insight into their condition — a phenomenon sometimes described clinically as impaired awareness of illness in addiction. If recent drug use exacerbates these monitoring deficits, then treatment programs might benefit from incorporating strategies that scaffold self-assessment: structured feedback, external monitoring tools, and therapeutic techniques such as metacognitive training or mindfulness-based relapse prevention, which explicitly aim to strengthen awareness of one’s own cognitive and emotional states. Conversely, the observed link between reduced recent use and better metacognition offers a hopeful message: the capacity for accurate self-reflection may be at least partially restorable during abstinence or sustained reduction in use.

At the same time, the researchers are careful to frame the findings within their limits. Confirmatory evidence strengthens confidence in the association between recent cocaine use and metacognitive impairment, but longitudinal studies remain essential to determine the direction of causality. It is plausible that impaired metacognition predisposes individuals to heavier use — poor self-monitoring could erode the brakes on consumption — while acute and chronic drug effects, in turn, deepen the impairment, creating a self-reinforcing loop. Disentangling these pathways will require repeated-measures designs that track metacognitive performance and drug use over time within the same individuals. Future research may also extend the paradigm to other substance use disorders to determine whether metacognitive dysfunction is a shared mechanism of addiction or a distinctive signature of stimulant dependence.

What the study already establishes, however, is a meaningful step forward for the cognitive neuroscience of addiction. By confirming, with methodological rigor, that cocaine use disorder involves a genuine and use-dependent impairment of metacognition, the work reframes the disorder not merely as a failure of willpower or reward processing, but as a disruption of the mind’s ability to audit itself. This perspective resonates with the lived experience of many people with addiction, who often describe acting on autopilot, surprised by their own behavior after the fact. Understanding that surprise itself — the failure to foresee and monitor one’s own lapses — has a measurable cognitive basis could reduce stigma and inform more compassionate, biologically grounded approaches to treatment. As research continues, the internal gauge of confidence and doubt may prove to be a promising therapeutic target, one whose recovery could mark a turning point in the journey out of dependence.

Subject of Research: Metacognitive impairment in cocaine use disorder and its relationship to recent drug use

Article Title: Metacognitive impairment in cocaine use disorder: confirmatory evidence for the effects of recent drug use

Article References: Moeller, S. J., McClain, N., Abeykoon, S., Alia-Klein, N., & Goldstein, R. Z. (2026). Metacognitive impairment in cocaine use disorder: confirmatory evidence for the effects of recent drug use. Translational Psychiatry. https://doi.org/10.1038/s41398-026-04406-7

Image Credits: AI Generated

DOI: 10.1038/s41398-026-04406-7

Keywords: cocaine use disorder, metacognition, substance abuse, prefrontal cortex, self-monitoring, Translational Psychiatry, addiction neuroscience, relapse prevention, cognitive impairment, dopamine, confidence judgments, abstinence