Agreement between large language models and human coders on psychological and clinical text: a systematic review and multilevel meta-analysis

Objective: To synthesize agreement between generative large language models (LLMs) and human coders on psychological and clinical text across four tasks, and benchmark it against human–human agreement on identical texts.Method: Preregistered on OSF (https://doi.org/10.17605/OSF.IO/Z63BS). Nine bibliographic sources were searched to October 21, 2026, plus citation searching. Eligible studies evaluated a generative LLM and human coders on identical text units using the same scheme. Correlated-and-hierarchical-effects models with cluster-robust variance estimation (CR2) pooled effect sizes.Results: Across 377 studies (416 reports), pooled chance-corrected agreement (kappa family) was 0.62 (95% CI [0.58, 0.65]; 107 studies, 595 estimates; 95% prediction interval [0.15, 1.00]). Agreement did not differ across tasks (codebook coding 0.63, assessment scoring 0.58, clinical rating 0.62, helper evaluation 0.59). Correlation/ICC pooled at 0.71 [0.66, 0.75] (95 studies). In 23 matched studies, LLMs agreed less with humans than humans with each other (Δκ = −0.11 [−0.18, −0.03]); equivalence within ±0.10 could not be established for models from 2025 (Δκ = −0.12, 90% CI [−0.19, −0.05]). Agreement did not improve with release date (+0.02/year, p = .282) or within-study model contrasts (−0.01 [−0.06, 0.03]); it tracked human reliability (slope = 0.63, p < .001) and was higher for manifest than latent constructs (+0.09, p = .028). Percentage agreement exceeded kappa by 0.24 in the same configurations. Certainty (GRADE) was low to very low.Conclusions: Generative LLMs achieve moderate-to-substantial agreement but remain, on average, below human inter-rater reliability. Agreement did not change across model release dates. Validation studies must report chance-corrected metrics and human benchmarks on identical units.