Problematic engagement with smartphones, social media, online games and other digital platforms may leave behind a detailed behavioural signature—one that researchers increasingly believe could help identify people at risk of behavioural addictions before serious harm becomes visible in clinical settings. A new review in Nature Reviews Psychology examines how machine-learning systems can analyse these digital traces, transforming everyday interactions with apps and devices into signals that may support earlier, more targeted mental-health interventions. The central idea is not that technology can diagnose addiction on its own, but that patterns hidden in ordinary digital activity could help clinicians recognize when online engagement is becoming compulsive, difficult to control or disruptive to daily life.
The review focuses on a growing category of mental-health concerns involving excessive or harmful engagement with digital activities. These include problematic smartphone use, problematic social-media use, gaming disorder and related clinically relevant conditions. Unlike many traditional psychiatric assessments, which depend largely on interviews, questionnaires and retrospective reports, digital-trace research examines behaviour as it unfolds. A smartphone can record when applications are opened, how long sessions last, how frequently users return to a platform and how often they move between digital activities. When collected responsibly, these observations may provide a more continuous picture of behaviour than a single appointment or survey completed weeks after the fact.
Machine learning is particularly suited to this kind of information because digital traces are high-dimensional and highly temporal. A system can be trained to identify relationships among dozens or thousands of variables, including session duration, time of day, frequency of use, pauses between interactions and changes in activity across days or weeks. Researchers may then use supervised-learning methods, in which algorithms learn from participants whose clinical status is already known, to estimate the likelihood that a new pattern is associated with problematic engagement. Other approaches, including clustering and anomaly detection, can search for naturally occurring behavioural profiles or identify abrupt changes from a person’s usual routine. The resulting models may be used to prioritize individuals for further assessment rather than to deliver an automatic diagnosis.
The timing and structure of online sessions could be especially informative. Regular, purposeful use may look very different from repeated checking that continues late into the night, interrupts work or study and becomes difficult to stop. An algorithm might examine whether sessions are becoming longer, whether usage is concentrated during vulnerable periods such as overnight hours, or whether a person repeatedly returns to an application after attempting to leave it. It could also track “fragmentation,” a pattern in which users rapidly switch among apps or repeatedly reopen the same service. None of these behaviours proves the presence of an addiction, but combinations of features may reveal escalating loss of control or increasing interference with everyday responsibilities.
The content and style of digital interaction provide another potential source of evidence. Typing dynamics can include the speed and rhythm of keystrokes, pauses between words and changes in interaction patterns over time. Language analysis, meanwhile, can examine the words and themes people use online, although this area raises particularly sensitive questions about privacy and interpretation. Changes in emotional language, self-reference or expressions of distress might correlate with problematic engagement in some contexts, but algorithms must account for culture, age, language, platform norms and the possibility that a person is discussing a difficult experience without personally experiencing it. The meaning of a digital trace depends on context, making purely statistical interpretation potentially misleading.
The review emphasizes that machine-learning research in this field should be connected to established psychological models of addiction. Digital activity becomes clinically concerning not simply because it is frequent, but because it may involve impaired control, persistent use despite negative consequences, escalating involvement, withdrawal-like experiences or significant disruption to relationships, education, work and health. These concepts offer a framework for deciding which digital signals matter and why. Without such a framework, an algorithm could mistake intensive but healthy engagement—such as professional gaming, online study or social connection across distance—for pathology. Conversely, it might overlook harmful behaviour that occurs less frequently but has severe consequences.
A major challenge is determining whether apparently promising models work beyond the datasets used to create them. Many digital-trace studies rely on relatively small samples, short observation periods or participants recruited from a limited demographic. An algorithm trained primarily on young adults using one operating system or one social platform may perform poorly among older users, people in different countries or individuals with different patterns of internet access. Researchers therefore need larger and more diverse datasets, preregistered studies and external validation in independent populations. They also need clinically meaningful outcomes, such as structured assessments, functional impairment or treatment needs, rather than relying only on self-reported screen time or platform-specific measures.
Interpretability is another essential issue. A risk score may be statistically accurate while offering little explanation of how it was produced. Clinicians and users need to know whether a prediction is driven by nighttime activity, rapid app switching, unusually long sessions or some other feature. Explainable machine-learning techniques can help reveal the variables influencing a model, but transparency alone does not guarantee that the underlying association is causal or clinically useful. A pattern may reflect depression, social isolation, shift work, caregiving responsibilities or limited access to offline services rather than an emerging behavioural addiction. Digital traces should therefore complement professional judgment and conversation, not replace them.
Privacy, fairness and autonomy will determine whether these technologies gain public trust. Continuous monitoring can expose intimate information about a person’s routines, relationships, emotional state and vulnerabilities. Systems designed for early detection could be misused by employers, insurers, schools, platforms or advertisers if safeguards are weak. Consent must be meaningful, data collection should be limited to what is necessary and people should understand how their information will be analysed, stored and shared. Developers must also test for unequal performance across demographic groups and avoid turning predictive tools into systems of surveillance or punishment. The review calls for governance frameworks that protect individuals while allowing carefully controlled research into potentially valuable digital-health applications.
The emerging picture is therefore both promising and cautious. Digital traces may eventually help clinicians detect changing risk earlier, identify people who could benefit from support and monitor whether interventions are helping. Machine learning could make sense of complex behavioural patterns that would be impossible to evaluate manually, particularly when those patterns unfold over weeks or months. But the technology remains dependent on high-quality evidence, clinically grounded definitions and responsible oversight. The most credible future is not an automated machine that labels users as addicted, but a privacy-preserving decision-support system that recognizes meaningful changes, explains its uncertainty and invites a human conversation before problematic online engagement becomes more damaging.
Subject of Research: Machine learning and digital trace data for detecting risk of behavioural addictions associated with online activities.
Article Title: Machine learning of digital traces to detect risk for behavioural addictions online
Article References: Burleigh, T.L., Hao, Y., Schivinski, B. et al. “Machine learning of digital traces to detect risk for behavioural addictions online.” Nature Reviews Psychology (2026). https://doi.org/10.1038/s44159-026-00604-8
Image Credits: AI Generated
DOI: 10.1038/s44159-026-00604-8
Keywords: Behavioural addictions, problematic smartphone use, problematic social media use, gaming disorder, digital trace data, machine learning, passive sensing, digital mental health, early risk detection, privacy, fairness, user autonomy
