AI Mindset – An Empirically Tested Theoretical Framework on the Psychological Factors Shaping AI Competence and AI Use

Synthesis: Feng, Koch and Carolus propose the AI Mindset framework: teachers' engagement with AI is not a matter of attitudes or skills alone but a triadic psychological architecture of dispositional factors (personality, basic psychological needs, self-efficacy, self-regulation), contextual and social facilitations (habit, social influence) and technology-related appraisals (attitudes, fear, performance and effort expectancy, hedonic motivation). From that framework they derive a mediation model in which these components predict AI use both directly and through pedagogical AI competence, operationalized as Intelligent-TPACK plus ethics. They tested it with structural equation modeling on 224 German schoolteachers surveyed between July and November 2025, with a final analyzed sample of 205. The model fitted exceptionally well and explained 62% of the variance in AI use and 46% in AI competence. Competence was the strongest proximal predictor of use, habit the strongest contextual predictor on both paths, effort expectancy worked fully through competence, and fear split into a positive direct effect on use and a negative indirect effect via competence. The takeaway is that competence is the transmission resource through which mindset becomes behavior, so support should be multilevel rather than attitudinal alone.

Key Findings

1. Sample and design. 224 Bavarian schoolteachers (mean age 45, SD = 9.88, range 25–65; 60.83% female; 71.43% from the KI@School project group) completed an online survey from 23 July to 24 November 2025; the SEM converged on n = 205 after 167 iterations using robust maximum likelihood.
2. Fit was excellent. χ²(43) = 41.40, p = .541 (n.s.), CFI = 1.00, TLI = 1.01, RMSEA = 0.00, 90% CI [0.00, 0.04], SRMR = 0.02, with all factor loadings significant (ps < .001) and age and gender as non-significant covariates.
3. Variance explained is high for a psychological model. The model accounted for 62% of the variance in AI use and 46% of the variance in AI competence.
4. Competence is the strongest proximal predictor of use. AI competence → AI use, β = .45, p < .001, the largest single path in the model.
5. What predicted AI competence. Habit β = .31 (p < .001), effort expectancy β = .28 (p = .006), openness β = .15 (p = .013), and AI-related fear β = −.15 (p = .025); need for relatedness was significant but negative, β = −.16 (p = .012).
6. What predicted AI use directly. Habit β = .28 (p = .005), hedonic motivation β = .19 (p = .029) and AI-related fear β = .15 (p = .008).
7. Indirect effects through competence. Openness β = .07 (p = .036), need for relatedness β = −.07 (p = .026), habit β = .14 (p = .011), effort expectancy β = .12 (p = .021) and fear β = −.07 (p = .044) were all significant; dispositional and effort-related predictors thus converted into use mainly via competence.
8. Fear is double-edged. Above and beyond everything else, fear was associated with more AI use directly while simultaneously predicting lower perceived competence, a pattern the authors read through the Threat-Challenge Framework as arousal that mobilizes engagement while eroding confidence.
9. What did not matter. General self-efficacy (β = .00, p = .944 on competence; β = −.04, p = .617 on use), performance expectancy, social influence, need for autonomy, need for competence, self-regulation, agreeableness, conscientiousness, extraversion, neuroticism, acceptance attitudes, age and gender were all non-significant.

The AI Mindset framework and its constructs

The framework is presented as a corrective to a research field that has treated AI adoption either as a technology-skills problem or as an ethics-risk problem, while the psychological preconditions of AI literacy and use stay implicit. Building on Bandura's social cognitive theory, Dweck's mindset work as extended by Ibrahim and colleagues to AI-related beliefs, self-determination theory, the Theory of Planned Behavior, the Technology Acceptance Model and UTAUT2, the authors reorganize complementary constructs into three functional domains. Dispositional factors cover the Big Five, the basic psychological needs of autonomy, competence and relatedness, general Self-Efficacy, and self-regulation as tenacious goal pursuit and flexible goal adjustment. Contextual and social facilitations cover habit as routinised engagement and social influence from colleagues and leadership, with structural conditions such as devices, connectivity and data-protection rules acknowledged as boundary conditions rather than modeled. Technology-related appraisals cover attitudes (with fear and acceptance as separable components), performance expectancy, effort expectancy and hedonic motivation. Together these shape the psychological tendency to adopt and integrate AI, and the framework explicitly separates the broad framework from the specific mediation model derived from it, in which Motivation and appraisals connect to use through pedagogical AI competence, defined through TPACK and ethical awareness as the ability to align AI with content and instructional strategy.

Measurement and sample

All instruments were German-language and standardized. Dispositional factors used the three need items from Chen et al. (2015), the 15-item Big Five Inventory, a three-item general self-efficacy scale (α = .87) and selected FlexTen self-regulation items (α = .74–.80). Technology-related appraisals used Sindermann et al.'s short attitude scale, whose fear (α = .66) and acceptance (α = .60) subscales were relatively weak, plus UTAUT2 scales for hedonic motivation (one item) and for performance expectancy and effort expectancy (four items each; internal consistency α = .93 across the multi-item scales). Habit was a single item and social influence three items (α = .93). AI use was an author-developed frequency and purpose measure (α = .68 for tools, .88 for purposes), and AI competence was the Intelligent-TPACK questionnaire with seven technological pedagogical content knowledge and four ethics items (α > .89). Participants came from all school types in Bavaria, mostly secondary, mainly teaching mathematics, German or English. About 24.55% did not report their school type, and the only group difference between project and non-project schools was on need for autonomy (p = .001), supporting comparability. AI Mindset was modeled as a formative, multidomain construct with subdomains as observed variables, and competence and use as latent variables with composite indicators, a choice the authors justify by model parsimony and stable estimation given the predictor-to-sample ratio.

Structural results

The structural picture is one of differentiated pathways rather than a single acceptance mechanism. Competence dominates: with β = .45 it is the strongest predictor of use, and it absorbs the effects of openness, effort expectancy and (negatively) fear. Habit is the only predictor with substantial weight on both paths, direct (β = .28) and indirect through competence (β = .14), which the authors interpret as evidence that routinised engagement reflects externally reinforced structural conditions rather than a purely internal automaticity. Effort expectancy is fully mediated, suggesting that perceived ease of use matters mainly because it builds the knowledge needed to implement AI, while hedonic motivation bypasses competence entirely with a direct effect (β = .19) consistent with intrinsic motivation energising behavior independently of perceived capability. The negative relatedness path is the most counter-intuitive result: teachers with stronger belonging needs reported lower AI competence, which the authors do not fully unpack, and they call for moderators and competing model specifications. The bluntest null is general Self-Efficacy, which is unrelated to either competence or use in this teacher sample, and age and gender likewise do not predict use. The AI anxiety finding is the paper's most distinctive: fear coexists with active use while constraining perceived capability, implying that anxious teachers may use AI despite, not because of, their emotional state.

Relation to existing evidence and implications for practice

The results sit uneasily beside a literature that routinely finds self-efficacy central to AI engagement: where AI literacy studies find Self-Efficacy mediating attitudes and use, this analysis finds general self-efficacy inert and identifies domain-specific pedagogical competence as the operative resource. The fear pattern aligns with work distinguishing affective reactions from cognitive self-evaluation (Sindermann et al.) and with AI anxiety research showing anxiety can coexist with adoption; where competence is modeled as a mediator of anxiety, the direction here is negative, so improving competence should reduce fear's toll rather than the reverse. Practically, the authors argue that professional development should stop selling attitudes and start building competence, embedding TPACK-aligned and ethical knowledge as the leverage point, and should engineer habit through accessible tools and collegial exchange. For designers, the findings translate into human-centered interfaces that lower perceived effort and raise enjoyment, and — because anxious teachers keep using AI — into complementing system-level usability work with psychologically informed interventions: open discussion of ethical and societal implications, emotion-regulation support, and guided practice with negative affect. The hedonic-motivation path is offered as a possible countermeasure to fear, since positive affect broadens behavioral repertoires and may weaken avoidance.

What this means for practice

• Instructors. Build professional learning around pedagogical AI competence rather than enthusiasm, because competence was the strongest single predictor of AI use (β = .45) while general self-efficacy predicted neither competence nor use.
• Instructors. Engineer habit deliberately by embedding a small, dependable set of AI tools into recurring planning and feedback routines, since habit carried the largest contextual weight on use (β = .28) and on competence (β = .14).
• Designers. Cut perceived effort in the tools teachers actually use, because effort expectancy reached use only indirectly through competence (β = .12): ease of use pays off when it builds the knowledge to implement AI, not when it merely feels pleasant.
• Instructors. Work with AI anxiety instead of waiting it out — fear was associated with more use (β = .15) while predicting lower competence (β = −.15) — by scheduling open discussion of ethical and societal implications, emotion-regulation support, and guided practice with negative affect.
• Researchers. Measure domain-specific pedagogical competence separately from general Self-Efficacy and re-test the framework against competing mediator models, since this is a first operationalization that explained 62% of the variance in use and 46% in competence.

Limitations

• The sample is German teachers from one federal state, recruited largely through an AI project group, so generalization beyond Germany — and beyond teaching to other professions — remains untested, and only cautious conclusions are warranted.
• The design is cross-sectional, so every path is an association; whether fear initially stimulates exploration and later blocks competence development needs longitudinal or experimental work.
• Nearly all measures are self-report, and several subscales are only marginally reliable (fear α = .66, acceptance α = .60, AI tool use α = .68), which weakens fine-grained claims.
• The framework itself is new — one operationalization, not a validated instrument — with several subdomains non-significant and competence's status as the primary mediator untested against competing models; school type was also missing for a quarter of respondents, and the analyses cannot rule out unmodeled contextual moderators or institutional differences in AI infrastructure and policy.

Connected Concepts

• AI Literacy — the competence domain the framework reframes as psychologically grounded rather than technical
• Anxiety and Stress — AI-related fear as a technology appraisal with a directly and indirectly opposed pair of effects
• Self-Efficacy — general self-efficacy measured and found inert, contrasted with domain-specific competence
• Teacher AI Competency — pedagogical AI competence as the mediator translating mindset into behavior
• Technological Pedagogical Content Knowledge (TPACK) — Intelligent-TPACK and ethics as the operationalization of AI competence
• Technology Adoption Models — one of the acceptance theories whose constructs the framework reorders
• Self-Determination Theory — basic psychological needs (autonomy, competence, relatedness) as dispositional factors
• Motivation — hedonic motivation as an independent direct driver of use
• Human AI Collaboration — teachers' sustained integration of AI into professional practice as the outcome modeled
• Self-Regulated Learning — self-regulation and habit formation under rapid technological change
• Generative AI — the technology whose adoption the model explains
• Theory Development in AI in Education — the paper's framing contribution, an integrative framework plus an empirically tested model

Connected Articles

• Teacher education for artificial intelligence literacy through a self-determination theory perspective — Teacher education for AI literacy through a self-determination theory lens
• From disengaged to self-determined: a latent transition analysis of students' AI learning motivation — Latent transition analysis of AI learning motivation profiles
• The Relationship Between AI Anxiety and Academic Motivation Among University Students: The Mediating Role of Emotion Regulation and the Moderating Role of Gender — AI anxiety and academic motivation, mediated by emotion regulation
• From AI Anxiety to Strategic Regulation: How University Students Transform Generative AI into a Strategic Learning Resource — Students turning AI anxiety into strategic self-regulation
• From Proficiency to Pedagogy: A Mixed-Methods Study of In-Service Teachers' TPACK-GenAI and the Mediating Role of Pedagogical Knowledge — TPACK-GenAI and the mediating role of pedagogical knowledge
• Factors Associated with Students' Adoption of Artificial Intelligence Technology in Tertiary Education: A Meta-Analytic Review — Meta-analysis of factors associated with AI adoption in tertiary education
• Social and Cognitive Drivers of Generative AI Adoption: A Unified Socio-Cognitive Model for Engineering Education — Social and cognitive drivers of GenAI adoption in engineering education
• Decoding symmetric and asymmetric pathways in generative AI learning adoption: a multi-method study — Symmetric and asymmetric pathways in GenAI learning adoption
• The impact of artificial intelligence literacy on learning engagement among university students: the mediating role of psychological capital and the moderating role of professional identity — AI literacy, psychological capital and learning engagement

Citation

Feng, S., Koch, M. J., & Carolus, A. (2026). AI Mindset – An Empirically Tested Theoretical Framework on the Psychological Factors Shaping AI Competence and AI Use. Computers in Human Behavior: Artificial Humans.