Problematic engagement with generative AI measurement drivers contexts and outcomes

Generative AI tools and conversational chatbots are increasingly embedded in education, work, and daily life, raising concerns about engagement that becomes difficult to self-regulate. This review synthesizes emerging research on problematic generative AI use, treating addiction-like language as part of a broader problematic-use construct rather than a clinical diagnosis. Following PRISMA 2020 reporting, a Scopus search executed identified 188 records. Included studies were characterised by research design and evidence type rather than scored with a formal risk-of-bias instrument. Measurement work is expanding quickly, but instruments remain partially fragmented, blurring the boundary between high engagement, habitual reliance, and impairment. Mechanism studies converge on attachment processes, anthropomorphism and perceived empathy, loneliness and social anxiety, immersive flow, and self-regulatory strain, suggesting that social presence and immersion amplify reliance for some users. In learning and knowledge work, performance pressure and incentive structures can normalize overreliance, with reported risks for cognitive offloading, independent thinking, academic integrity, and skill development. Outcome studies associate compulsive or dependent use with technostress, psychological distress, and lower wellbeing, while mitigation discussion highlights reinforcement-friendly interface cues, privacy and memory governance, and sustainability trade-offs. Because the corpus is almost entirely cross-sectional and self-reported, these relationships are associations rather than causal effects. The key distinction is not between users and non-users, but between bounded, purpose-driven use and escalating reliance with reduced self-regulation. An Antecedents–Decisions–Outcomes structure is used only to frame a forward-looking agenda for longitudinal and intervention research. A sensitivity analysis comparing studies published in 2026 with those published earlier showed that no conclusion was reversed: the design profile and principal gaps held, while the most recent longitudinal studies strengthened the evidence on temporal order and outcomes.