Development and validation of the AI-assisted psychoemotional regulation scale (ERP-IA-6).

BACKGROUND

Conversational and generative artificial intelligence tools are increasingly used to organize thoughts, clarify emotional experiences, and cope with everyday distress. However, brief instruments assessing perceived AI-assisted psychoemotional regulation remain scarce.

OBJECTIVE

To develop the AI-Assisted Psychoemotional Regulation Scale (ERP-IA-6) and examine its initial psychometric properties among Peruvian adults.

METHODS

The initial eight-item pool was reviewed by five psychologists, and all items obtained Aiken's V values above 0.80. A total of 538 adults participated (M age = 24.51 years, SD = 7.48; range = 18-81); 81.6% were aged 18-29 years, 52.0% were women, and 93.1% reported AI use. Using a reproducible sex-stratified procedure, the sample was divided into exploratory (n = 266) and confirmatory (n = 272) subsamples. Exploratory factor analysis used polychoric correlations and unweighted least squares; confirmatory factor analysis used WLSMV for ordinal indicators. Reliability, AVE-based convergent evidence, sensitivity analyses, shared-method diagnostics, and ordinal measurement invariance across sex were also examined.

RESULTS

Exploratory analyses supported reducing the scale from eight to six items based on conceptual redundancy and parsimony. The final version showed adequate factorability (KMO = 0.910; Bartlett's χ2 (15) = 1128.37, p < 0.001) and a one-factor solution explaining 67.4% of standardized item variance, with loadings from 0.750 to 0.872. In the confirmatory subsample, loadings ranged from 0.839 to 0.930. Robust incremental and residual fit were favorable (CFI = 0.981; TLI = 0.968; SRMR = 0.019), although robust RMSEA was elevated, 0.118, 90% CI [0.093, 0.145]. Sensitivity analyses showed that the one-factor structure persisted, while localized residual dependence partly accounted for the elevated RMSEA. Reliability was high (α = 0.935; ordinal α = 0.945; ω = 0.948), as were composite reliability (CR = 0.959) and average variance extracted (AVE = 0.794). Threshold, metric, and scalar models provided favorable but cautious evidence of invariance across sex.

CONCLUSION

The ERP-IA-6 provides favorable initial evidence as a brief measure with a dominant unidimensional structure for assessing perceived AI-assisted psychoemotional regulation among Peruvian adults. Scores should be interpreted as perceived regulatory support, not as evidence of clinical effectiveness, therapeutic safety, or equivalence to professional care.