Survey links AI dependence to a peak in employee innovation

Employee dependence on generative artificial intelligence showed an inverted U-shaped association with innovation performance in a multi-study research program, with the strongest results occurring at intermediate levels of dependence rather than at either extreme. The finding, reported in BMC Psychology, indicates that greater reliance on generative AI was linked to better innovation only up to a point, after which the association became less favorable. Because the main test used a three-wave time-lagged survey, the results do not establish that dependence causes changes in innovation performance.

The research program addressed a concept that the authors say had received little empirical attention and lacked a validated measurement instrument. It combined qualitative interviews, psychometric scale development and a time-lagged survey to examine employee dependence on generative AI, workplace thriving, mindfulness and innovation performance. Its overall design was intended both to define dependence and to test whether the relationship with innovation followed a nonlinear pattern.

In Study 1, the researchers used an abductive grounded-theory approach with 24 knowledge workers. They conducted 21 initial interviews and then three additional interviews to assess theoretical saturation. This qualitative work identified three interrelated and potentially progressive dimensions of generative AI dependence: instrumental, psychological and cognitive.

The published abstract does not provide detailed definitions for those dimensions, but it presents them as components of the broader construct measured by the new scale. The authors do not present the dimensions as a formally demonstrated stage-based progression. Study 2 developed and evaluated a 20-item instrument called the GenAID across three survey rounds with 233, 271 and 269 respondents.

The researchers applied item-total correlations, exploratory factor analysis and confirmatory factor analysis, supplementing these approaches with content-validity indices, comparisons among correlated three-factor, second-order, bifactor and one-factor models, omega coefficients and heterotrait-monotrait ratios. The analysis provided preliminary evidence for the scale’s internal structure, reliability and discriminant validity, while model comparisons supported the use of a total dependence score. The authors nevertheless describe the instrument as initial, so the evidence does not demonstrate that the scale is equally reliable across occupations, sectors or national settings.

Study 3 supplied the principal test of the proposed double-edged-sword relationship. A three-wave time-lagged survey involving 342 employees examined the curvilinear association between generative AI dependence and innovation performance. The researchers used hierarchical regression and the MEDCURVE macro, followed by turning-point analyses, analyses of individual dimensions and robustness checks.

The estimated turning point fell within the range of the observed data. Slopes were significantly positive at one standard deviation below the mean and significantly negative at one standard deviation above it, supporting the inverted U pattern within the sampled data rather than demonstrating a universal threshold. Workplace thriving was tested as a mediator in the relationship between generative AI dependence and innovation performance, while mindfulness was tested as a moderator.

The statistical model indicated that mindfulness flattened the downward slope and shifted the turning point toward higher levels of dependence. This means that the shape of the association varied with mindfulness, although it does not show that deliberately increasing mindfulness would improve innovation outcomes. Mindfulness-based programs are proposed by the authors as candidate interventions for future testing, not as interventions already shown to be effective.

When the three dimensions were analyzed separately, all showed significant negative curvature, and their estimated turning points occurred progressively earlier from instrumental through psychological to cognitive dependence. However, the researchers report that the complete set of inverted-U conditions was not uniformly satisfied for every dimension. This qualification matters because it prevents the strongest dimension-level findings from being interpreted as identical in strength or shape.

The program provides an initial conceptualization of employee dependence on generative AI, a preliminary measurement scale and observational evidence consistent with a curvilinear relationship between that dependence and innovation performance. Its survey-based findings identify workplace thriving as a statistical mediator and mindfulness as a moderator, but they do not determine the direction of causation among those variables. The results also remain limited to the populations and settings represented in the three studies.

Further research is needed to evaluate the scale more broadly, examine the proposed mechanisms and directly test whether interventions informed by these findings can change employee innovation performance.

Subject of Research: Survey-based associations among employee generative AI dependence, workplace thriving, mindfulness and innovation performance

Article Title: The double-edged sword of generative AI dependence: scale development and its curvilinear impact on employee innovation

Article References: Chai, M., & Zeng, Q. (2026). The double-edged sword of generative AI dependence: scale development and its curvilinear impact on employee innovation. BMC Psychology. https://doi.org/10.1186/s40359-026-05530-1

Image Credits: AI Generated

DOI: 10.1186/s40359-026-05530-1

Keywords: generative AI, employee innovation, AI dependence, workplace thriving, mindfulness, curvilinear relationship, knowledge workers, psychometric scale, time-lagged survey, BMC Psychology, double-edged, sword