AI threat linked to deeper agent use among 387 Chinese knowledge workers

Workers who perceive that artificial intelligence could undermine their jobs or skills were associated with deeper use of AI agents in a cross-sectional study of 387 Chinese knowledge workers. Rather than showing a simple pattern of avoidance, the research found that perceived AI threat was statistically associated with stronger confidence in using agents, greater motivation to protect one’s professional position, and more intensive engagement after initial adoption. These relationships operated through two parallel pathways, with no statistically significant direct association between threat perception and engagement once the two proposed mechanisms were included in the model.

The study, published in Frontiers in Psychology, examined post-adoptive engagement rather than the initial decision to try AI. The researchers defined this deeper engagement through five behaviors: frequent daily use, exploration of new platforms or features, participation in customization or development, integration into core workflows, and recommending the tools to colleagues. Participants were active users who had incorporated AI agents into their weekly routines for at least six months and used them for a minimum of 10 minutes per week. The final sample was concentrated in research, digitalization and technical work, and management, with 93.28% holding a bachelor’s degree or higher and 83.98% reporting monthly incomes above 10,000 yuan.

Investigators first conducted 13 semi-structured interviews with people drawn from agent-learning and developer communities, including five researchers, three corporate information-technology or digitalization technicians, two middle managers, two legal or consulting professionals, and one freelancer. Two researchers independently took field notes during the 10-to-20-minute interviews and reconciled their records afterward. Grounded-theory coding yielded 86 initial concepts and six broad categories. Across the interviews, users described both a growing belief that they could master AI agents and an external push arising from fears of elimination, budget cuts, poor performance evaluations, or falling behind colleagues.

The qualitative findings informed four multi-item scales rated on seven-point Likert-type scales. AI threat perception covered job stability, skill depreciation, replacement risk, competitive disadvantage, and uncertainty about career prospects. AI self-efficacy measured confidence in learning complex tools, understanding operating logic, applying AI to work, solving problems independently, and participating in customization or development. Defensive motivation captured the intention to learn in order to avoid unemployment or marginalization, manage uncertainty, respond to outside pressure, and preserve organizational value. After exploratory factor analysis removed one item from each of two scales, the remaining 18 indicators showed good reliability and construct validity. Cronbach’s alpha values ranged from 0.819 to 0.883, while average variance extracted values ranged from 0.549 to 0.572.

The researchers then used structural equation modeling to test a model connecting threat perception, self-efficacy, defensive motivation, and engagement. The model fit was strong, with a comparative fit index of 0.998, a Tucker–Lewis index of 0.998, and a root mean square error of approximation of 0.010. Threat perception was positively associated with AI self-efficacy, with a standardized coefficient of 0.486, and with defensive motivation, with a coefficient of 0.460; both associations had p-values below 0.001. Self-efficacy and defensive motivation were each similarly associated with post-adoptive engagement, with standardized coefficients of 0.411 and 0.409, respectively. By contrast, the direct path from threat perception to engagement was small and non-significant, at a standardized coefficient of −0.059 with p = 0.383.

Bootstrap analysis with 5,000 resamples supported both parallel indirect pathways. The estimated association between threat perception and engagement through self-efficacy was 0.186, with a bias-corrected 95% confidence interval from 0.122 to 0.271. The corresponding estimate through defensive motivation was 0.176, with an interval from 0.109 to 0.261. Neither confidence interval included zero. A proposed serial pathway running from threat perception through self-efficacy and defensive motivation to engagement was not supported: its estimate was 0.011, its confidence interval ranged from −0.013 to 0.037, and p was 0.318. The results therefore fit an indirect-only statistical pattern within this model, rather than establishing that threat perception has no direct influence on behavior.

Multi-group analyses identified differences in particular links rather than across the entire model. The relationship between threat perception and self-efficacy was stronger among workers aged 25 to 45 than among a pooled comparison group of respondents under 25 or over 45, with standardized coefficients of 0.580 and 0.230 and a group-difference p-value of 0.002. That age analysis was exploratory and combined two theoretically distinct outer groups to preserve adequate statistical power. Education showed a different boundary condition: threat perception was more strongly associated with defensive motivation among postgraduates than among workers with undergraduate education or less, with coefficients of 0.581 and 0.323 and p = 0.039 for the difference.

Region also modified one pathway. Participants in eastern China showed a stronger association between threat perception and self-efficacy than those in central and western China, with standardized coefficients of 0.603 and 0.397 and a difference test yielding p = 0.034. The authors interpret this as a possible contextual pattern associated with differences in digital development, but they emphasize that the study did not directly measure local AI penetration, provincial digital-economy indicators, or individual environmental stressors. Other tested paths did not differ significantly across groups. Gender, occupation, and income were described as non-significant moderators, although the limited representation of women and lower-income or less-educated participants constrains those comparisons.

The central result is a counterintuitive coexistence: the same sense of technological vulnerability that might discourage initial adoption was statistically linked to both capability-building and loss-prevention orientations among established users. Self-efficacy represents confidence that one can learn, operate, apply, and customize AI agents, while defensive motivation represents purposeful engagement aimed at countering perceived workplace losses. The two paths carried similar indirect-effect estimates and were not significantly related to each other in the cross-sectional model. This pattern does not prove that threat causes either mechanism, nor does it show that anxiety itself improves productivity. Because all variables were measured at one time through a single-source self-report survey, reciprocal relationships and common-method effects remain possible, and exploratory and confirmatory factor analyses used the same sample.

Several further limitations constrain the findings. Recruitment through convenience and snowball sampling produced a self-selected, relatively elite group of active users whose demographic range was narrow. The education comparison was effectively between postgraduate and bachelor’s-level respondents because only 6.72% had a junior college education or lower. The age analysis pooled younger and older workers, while the regional analysis combined central and western respondents, obscuring variation within those areas. The theoretical model also used downstream self-efficacy and defensive motivation rather than directly measuring distinct challenge and threat appraisal states, and it did not include formal Protection Motivation Theory components such as response efficacy and response costs. Longer studies using independent samples, repeated measurements, and objective records of AI use would be needed to determine whether fear-driven learning precedes growing confidence, follows it, or develops alongside it.

Subject of Research: Psychological pathways linking perceived AI workplace threats to post-adoptive AI agent engagement

Article Title: How AI threat perception relates to post-adoptive AI agent engagement: the dual paths of self-efficacy and defensive motivation

Article References: Yang, G., Feng, M., & Yang, W. (2026). How AI threat perception relates to post-adoptive AI agent engagement: the dual paths of self-efficacy and defensive motivation. Frontiers in Psychology, 17, Article 1887571. https://doi.org/10.3389/fpsyg.2026.1887571

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1887571

Keywords: AI agents, AI threat perception, self-efficacy, defensive motivation, technology adoption, knowledge workers, human-AI collaboration, workplace psychology, structural equation modeling, threat, perception, relates