When Generative Artificial Intelligence Becomes a Colleague: Dual Pathways of Empowerment and Depletion in University Design Teachers' Work Behaviors

Synthesis: Ding, Hu, Kim and Chen surveyed university design teachers in mainland China and found that generative AI acts on their work through two paths at once. Use raises teaching self-efficacy and well-being, which carry teachers toward experimentation and engagement, and it raises AI anxiety and job stress, which carry them toward withdrawal. Which path dominates depends on whether teachers read the institution as backing them or as having broken a promise; that makes support part of the intervention.

Key Findings

1. Both paths run in the same model. Use predicts more confidence and satisfaction and more anxiety and stress at once, so the empowerment and burnout stories describe one workforce, not two.
2. The enabling path runs through capability and satisfaction. Teachers who feel more capable and more satisfied are the ones who try new things and stay engaged.
3. The depleting path has its own tracks. Anxiety and stress rise with use and lead separately to less innovation and more withdrawal, which is why confidence and worry coexist.
4. Support strengthens what use gives back. Where teachers read the institution as behind them, the links from use to confidence and satisfaction are stronger, and the positive indirect effects become significant.
5. A broken promise strengthens what use costs. Where teachers believe the institution has not kept its side of an unwritten deal, use produces more anxiety and stress, and the negative indirect effects become significant.

Two paths through one tool

The model crosses conservation of resources theory with social exchange theory, treating use as a demand that both replenishes and consumes. On the enabling side, use raised self-efficacy (β = 0.322, p < .001) and well-being (β = 0.300, p < .001); efficacy then predicted innovation at β = 0.157 and withdrawal at β = −0.155, and well-being at β = 0.152 and β = −0.145. On the depleting side, use raised AI anxiety (β = 0.268, p < .001) and job stress (β = 0.262, p < .001), which predicted innovation at β = −0.152 and β = −0.168, and withdrawal at β = 0.264 and β = 0.177.

The indirect effects are small: 0.051 on innovation through efficacy, −0.041 through anxiety. No single chain explains the outcome, so the authors argue for two mechanisms rather than one strong effect. A department measuring only satisfaction will see the gain and miss the cost.

The support signal decides which path dominates

Perceived organizational support is the teacher's read of whether the institution backs them: training, time, recognition, a policy that holds for the semester. Where that read is high, the same use produces more efficacy (b = 0.153) and well-being (b = 0.176). Where it is low, the indirect effect of use on innovation through efficacy was 0.028 and not significant, against 0.071 where support was high.

Psychological contract breach is the mirror. Where breach is high, use produced more anxiety (b = 0.171) and more stress (b = 0.194), and the negative indirect effect on innovation through anxiety became significant. The authors call this a governance finding in psychological clothing: identical access and training can land two faculties on opposite paths depending on what teachers believe about the institution's word. They did not separate mandated from voluntary use, the variable that would test it.

What the design studio adds, and what the model cannot show

The sample is design faculty, and that matters: studio teaching rests on visual production and critique, so image-generating tools touch professional identity more directly than in a lecture course. Four hundred and thirty-six valid responses came from a survey recruited on Xiaohongshu and run between June and July 2025, with screening questions and a gesture photograph used to confirm that respondents taught design and subscribed to a generative tool. All were already users, which fits the argument but is also a selection effect.

The design is cross-sectional, so the sequence from use to anxiety to withdrawal is assumed rather than observed. The CFI and TLI fall below the usual threshold (0.808 and 0.802), while RMSEA (0.062) and SRMR (0.058) are acceptable; the authors call the overall fit adequate for applied work. Gender, age and rank did not predict the outcomes, supporting their claim that resource states and organizational perception matter more than surface attributes.

What this means for practice

• Instructors: Keep a two-column log for a month: tasks where generative AI returned time and tasks where it created rework. Bring the second column to whoever supports teaching technology rather than absorbing it.
• Faculty developers: Move training from tool operation to instructional integration, using studio cases where the tool fails, and write generative AI into briefs where it has a defined job, keeping critique and selection with students. Pair every demonstration with a segment on what it cannot judge.
• Administrators: Make support perceptible: a named support person, protected time, and a commitment that survives the semester. Announcing a pilot and then retrenching is the breach condition the model identifies.
• Researchers: Separate mandated from voluntary use and test group differences. Both are the institutional variables the moderation argument depends on, and neither was measured.

Limitations

• Cross-sectional and single-country: 436 design teachers in mainland China, recruited through one social-media platform whose users skew young, urban and female.
• All constructs are self-reported in one instrument; the common-method test cannot rule out one mood inflating both anxiety and withdrawal.
• CFI and TLI fall below conventional thresholds, which the authors treat as acceptable rather than explaining.
• The study does not distinguish institution-mandated from voluntary use, so reading support and breach as causes remains untested.

Citation

Ding, N., Hu, L., Kim, K.-T., & Chen, M. (2026). When generative artificial intelligence becomes a colleague: Dual pathways of empowerment and depletion in university design teachers' work behaviors. Sustainability, 18(4), 1775.