A Critical Narrative Synthesis of Psychological Correlates, Measurement, and Reported Findings on Conversational AI Engagement and Dependence-Related Constructs

Synthesis: Deyu Yan (2026) presents a critical narrative synthesis of 51 peer-reviewed records (2021–2026) on the psychological correlates, measurement, and reported findings of conversational AI engagement and dependence-related constructs. The review separates conceptual and measurement work from 22 first-order empirical records and argues that labels such as trust, reliance, attachment, problematic use, and dependence refer to different processes that are frequently conflated in a rapidly growing body of scales. Cross-sectional studies mainly associated stronger AI-engagement or dependence-oriented scores with loneliness, social anxiety, depressive symptoms, low self-esteem, attachment insecurity, academic stress, escapism, anthropomorphism, fatigue, weaker critical thinking, procrastination, and lower Well-Being — associations that do not establish prediction or consequence. The synthesis develops a provisional framework separating instrumental-cognitive from relational-emotional use and treating AI Regulation in Education as a distinct dimension, and it calls for careful construct separation and design-matched claims.

Key Findings

• Constructs are routinely conflated: Trust (expectations about system competence), reliance (behavioral delegation), over-reliance (delegation without scrutiny), attachment/companionship (emotional closeness), and problematic use/dysregulated dependence (impaired control) refer to different processes yet are often treated as one continuum; several distinct instruments share overlapping labels without a stable object of measurement.
• Cross-sectional correlates, not causation: Mainly cross-sectional studies linked stronger engagement or dependence-oriented scores to poorer well-being, fatigue, weaker critical thinking, procrastination, loneliness, and academic stress — but none of these findings establish prediction or consequence.
• Both supportive and harmful experiences are reported: Qualitative and quasi-experimental work documents both relational support (perceived disclosure, lower loneliness in some settings) and distress linked to emotionally significant use, with limited, context-dependent evidence of short-term reductions in loneliness and social anxiety.
• Evidence is largely low leverage: Of 22 first-order empirical records, 16 (72.7%) provided lower-leverage evidence for directional claims and six moderate-leverage evidence; none supported a firm causal model, and direct cross-cultural and platform-comparative research remains rare.
• A provisional framework: Separating use orientation (instrumental-cognitive vs. relational-emotional) from regulation offers an interpretive — not validated — way to organize the field; frequent use, reliance, or attachment should not be labeled dependence without impaired control or functional harm.

What this means for practice

• Researchers. Match the instrument to the construct before measuring: trust (expectations about system competence), reliance (behavioral delegation), over-reliance (delegation without scrutiny), attachment, and dysregulated dependence refer to different processes that current scales often conflate under shared labels.
• Do not label frequent use, reliance, or attachment as dependence without documented impaired control or functional harm; only 22 of the 51 retained records supplied first-order empirical evidence, and 16 of those (72.7%) were lower-leverage for directional claims.
• Design for temporal ordering, because the corpus is dominated by cross-sectional surveys in which modeled mediation cannot establish that the mediator preceded the reported outcome.
• Educators. Distinguish instrumental-cognitive use (explanation, thinking support) from relational-emotional use (companionship, disclosure) when monitoring student engagement, and route emotionally significant use toward well-being support rather than treating it as educational over-reliance.
• Keep educational delegation claims separate from relational attachment and clinical addiction constructs when interpreting classroom chatbot use.

Limitations

• This is a critical narrative synthesis, not a systematic review: complete search histories and duplicate-removal logs were not retained, the search strategy was not preregistered, and the authors state the design claims neither PRISMA compliance nor systematic-review reproducibility.
• A single author completed eligibility checking, extraction, classification, appraisal, and synthesis; no second independent coder was used and no inter-rater reliability statistic is reported, so the sensitivity analysis does not correct selection bias or single-author subjectivity.
• Of 51 retained records, only 22 supplied first-order empirical evidence and 16 of those (72.7%) provided lower-leverage evidence for directional claims; none supported a firm causal model.
• Coverage concentrates on generic conversational AI, ChatGPT and other productivity systems, Replika, and companion chatbots, which the authors say limits generalization across age groups, cultures, and platform types; the instrumental/relational framework itself is an interpretive step after coding, not a factor-analytically derived or causally tested model.

Citation

Yan, D. (2026). A critical narrative synthesis of psychological correlates, measurement, and reported findings on conversational AI engagement and dependence-related constructs. Frontiers in Psychology, 17, 1827795.