When Help Becomes a Habit: What Science Really Knows About Over-Reliance on AI

Artificial intelligence has quietly woven itself into the fabric of everyday cognition. Students draft essays with chatbots, professionals summarise reports with large language models, and millions of people now turn to conversational agents for everything from recipe ideas to emotional reassurance. As these tools grow more capable and more accessible, a pressing question has moved from the margins of technology commentary into mainstream psychology: what happens to the human mind when it leans too heavily on machine intelligence? A new integrative review published in the International Journal of Mental Health and Addiction offers the most comprehensive attempt yet to answer that question, and its conclusions are more nuanced, and in some ways more surprising, than the alarmist headlines that often accompany discussions of so-called AI addiction.

The review, led by Batoul Fneich of Lancaster University with colleagues at the University of Derby, the University of Tabuk, and beyond, synthesised 28 empirical studies published between 2023 and 2026. The team searched Scopus, PubMed, and PsycINFO from database inception to January 2026, screening nearly 1,500 records and supplementing the search with forward and backward citation tracking. Following established integrative review methodology, the researchers combined quantitative, qualitative, mixed-methods, and measurement studies, appraising each with the Mixed Methods Appraisal Tool and synthesising findings narratively rather than statistically, because the heterogeneity of designs, populations, and definitions made meta-analysis inappropriate. The result is a map of a young, fast-moving, and conceptually fragmented research field.

The first striking finding is that nobody agrees on what AI over-reliance actually is. The included studies used a bewildering array of overlapping labels: reliance, dependence, dependency, problematic use, compulsive use, psychological dependence, and even AI addiction. These constructs differ in whether they emphasise functionality, perceived need, controllability, adverse consequences, or functional impairment, and they are often used inconsistently across studies. Reliance, for instance, generally refers to incorporating AI into learning, work, or decision-making and is not inherently maladaptive. Dependence typically involves a perceived need for AI assistance or difficulty completing tasks without it, but does not necessarily imply impaired control. Problematic and compulsive use add recurrent, difficult-to-control engagement with adverse consequences, while addiction-like engagement further encompasses salience, tolerance, withdrawal, conflict, relapse, and functional impairment. Crucially, AI addiction is not recognised as a diagnostic category in either the DSM-5-TR or the ICD-11.

Beneath the terminological chaos, the reviewers identified three recurring dimensions of over-reliance. Cognitive over-reliance, the most frequently represented, appeared in roughly half of the studies and involves delegating thinking, learning, decision-making, or information retrieval to AI, sometimes accompanied by reduced verification and analytical engagement. Behavioural over-reliance, identified in about 40 percent of studies, involves habitual, excessive, or compulsive engagement and a perceived inability to function effectively without AI assistance. Emotional over-reliance, present in just under a third of studies, involves attachment to AI or its use for companionship, reassurance, validation, or emotional regulation, a dimension particularly evident in research on conversational and social chatbots. These dimensions frequently overlap within individual studies, but the evidence does not establish whether they represent components of a single construct, distinct forms of reliance, or sequential stages of problematic engagement.

Measurement is equally unsettled. Some studies relied on single-item ratings of perceived dependence, while others adapted instruments originally designed for computer, internet, smartphone, or gaming dependence. A handful of AI-specific scales have emerged, including a Researcher AI Addiction Scale assessing compulsive behaviour, overdependence, functional impairment, withdrawal, and tolerance, and a Reliance on ChatGPT Scale for nursing students spanning task dependency, ethical concerns, social influence, and privacy dimensions. Similar construct labels were operationalised through different indicators, and differently named instruments often measured overlapping features. Few studies established thresholds distinguishing functional reliance from practically significant over-reliance, meaning the available instruments are not directly interchangeable and cross-study comparisons remain fraught.

So what does the evidence say about psychological consequences? Mental health symptoms and psychological distress formed the largest evidence domain, examined in 11 of the 28 studies. Ten of those 11 identified some association between greater AI dependence, problematic use, or addiction-like engagement and psychological symptoms. Problematic ChatGPT use and chatbot dependence were linked to depressive symptoms among university students and adult users, while compulsive use was associated with anxiety, burnout, and sleep disturbance. Two qualitative studies described acute distress when AI access was lost, with nursing students in one Turkish study coining the memorable term AIlessphobia to describe fear and panic at the anticipated loss of AI support. Yet the strength and direction of these associations varied considerably, and some dimensions of addiction-like engagement showed weak or non-significant relationships with mental health outcomes.

The single most important caveat, and arguably the review’s headline finding, concerns temporal direction. Almost all of the evidence is cross-sectional, capturing snapshots that cannot establish whether AI over-reliance causes psychological distress or vice versa. The only longitudinal study included in the review, a cross-lagged panel analysis of adolescents, found that baseline depression and anxiety predicted subsequent AI dependence, whereas AI dependence did not predict later depression or anxiety. Escape and social motivations helped explain the pathway from earlier distress to later dependence. This pattern is consistent with Compensatory Internet Use Theory, which proposes that people turn to technologies that offer distraction, reassurance, or emotional regulation when experiencing psychological difficulty. Generative and conversational AI, continuously available and endlessly responsive, may be especially suited to this compensatory function. If pre-existing vulnerability drives dependence, then interpreting depression or anxiety purely as consequences of AI use risks reversing the arrow of causation, and interventions focused solely on reducing AI engagement may fail to address the underlying need the behaviour serves.

Findings on well-being and self-related functioning were decidedly mixed. Two of four well-being studies found no direct association between AI dependence and mental well-being, with one showing that purpose of use mattered: dependence was positively associated with well-being among people primarily using chatbots for information retrieval, but not among those using them for other purposes. Self-efficacy results were similarly contradictory. Greater AI dependence was associated with lower academic self-efficacy in one study, yet positively associated with reported academic self-efficacy in another, even as that same study found dependency predicted lower actual academic achievement. Among teachers, ChatGPT dependence correlated with greater work self-efficacy and effort. One workplace study captured the double-edged nature of the phenomenon elegantly: dependence on intelligent machines increased perceived progress toward work goals while simultaneously heightening self-esteem threat related to competence and intelligence. The reviewers suggest, drawing on Social Cognitive Theory, that confidence in completing a task with AI may be psychologically distinct from confidence in one’s unaided ability, a distinction future research should measure directly.

Cognitive findings reinforce the tension between augmentation and substitution. Addiction-like ChatGPT use was associated with greater cognitive miserliness, a shift toward rapid, low-effort heuristic processing rather than deliberate analytical thinking. Qualitative studies described perceived reductions in critical thinking, creativity, independent problem-solving, and metacognitive engagement among heavy users, though these reflect participants’ perceptions rather than objectively measured decline. The review frames the relevant distinction not as AI use versus non-use, but as augmentation, where AI supports performance while users retain understanding, monitoring, and independent capability, versus substitution, where functions users would otherwise perform are increasingly delegated to the machine. Notably, the evidence does not demonstrate that substitution causes permanent cognitive atrophy; its significance likely depends on what is delegated, how often independent capability is exercised, and whether users retain active control. Automation bias, the tendency to favour automated recommendations despite contradictory information, compounds the risk by reducing verification of AI outputs.

Social and relational consequences remain the field’s blind spot: only one of the 28 studies directly examined them as outcomes, finding that compulsive ChatGPT use was associated with greater loneliness and social avoidance among students, with psychological distress mediating the pathway. Other studies treated loneliness, attachment, and companionship needs as antecedents rather than consequences, suggesting the current evidence is stronger for emotional vulnerabilities contributing to AI dependence than for AI dependence causing social withdrawal. The reviewers conclude that AI over-reliance is best understood as an overarching, multidimensional umbrella concept rather than a unitary behavioural addiction, with different theoretical frameworks, cognitive offloading and trust calibration for task reliance, compensatory and attachment processes for emotionally motivated use, and the I-PACE model only where impaired control and functional impairment are evident, applying to different forms of engagement. Their practical message is equally clear: frequent AI use is not itself evidence of pathology, and the meaningful questions are what functions are being delegated, why, whether users retain control and independent capability, and whether reliance produces genuine distress or impairment. Progress, they argue, now demands standardised measurement, clearer conceptual boundaries, and longitudinal and experimental research capable of disentangling augmentation from dependence.

Subject of Research: Psychological outcomes and correlates associated with over-reliance on artificial intelligence tools

Article Title: Psychological Outcomes and Correlates Associated with Over-Reliance on Artificial Intelligence Tools: An Integrative Review

Article References: Fneich, B., Ahadi, N., Ali, J., & Alkadhimi, F. (2026). Psychological Outcomes and Correlates Associated with Over-Reliance on Artificial Intelligence Tools: An Integrative Review. International Journal of Mental Health and Addiction. https://doi.org/10.1007/s11469-026-01724-1

Image Credits: AI Generated

DOI: 10.1007/s11469-026-01724-1

Keywords: artificial intelligence, AI over-reliance, AI dependence, mental health, cognitive offloading, ChatGPT, problematic technology use, depression, anxiety, self-efficacy, automation bias, integrative review