Between Human and Machine: University Teachers’ Accounts of Judgment, Metacognitive Defense, and Cognitive Offloading in Generative AI Use

Synthesis: Qiang and Wang (2026) analyzed multi-source accounts from 25 university teachers in China — 22 interviews plus three documentary cases — to ask how teachers describe their own generative AI use and the learning conditions they arrange for students. A shared boundary around what AI cannot do (emotion, embodiment, metacognition, tacit knowledge, judgment) sat alongside a relocation of professional value from knowing to judging, with verification carrying clear evidence in 24 of 25 cases. On the "intellectual degeneration" proposition the protocol supplied, 9 teachers endorsed it, 11 made it conditional on user agency, 2 rejected it, 1 was ambivalent, and 2 took no position. The authors frame offloading as something teachers regulate on both sides of the desk, and their belief–strategy–practice cycle as a heuristic rather than a demonstrated mechanism.

Key Findings

1. The boundary was consistent, but it was elicited. Emotion and embodiment drew clear or partial evidence in 24 of 25 cases (13 clear, 11 partial, one counter-case in Teacher A); metacognition, tacit knowledge and judgment drew 25 of 25 (14 clear, 11 partial). Both codes followed a direct question about what AI cannot replace, so the agreement describes how teachers answered that question rather than a concern they raised unbidden.
2. Most teachers renamed the role, and the names were theirs. 15 of the 25 cases (60.0%) no longer treated "knowledge transmitter/knowledge expert" as their primary self-positioning, adopting guide, designer, organizer, coach, companion, gatekeeper, navigator, goalie or trainer. The other 10 (40.0%) kept the expert label, most of them reinterpreting it as producer, systematizer, gatekeeper or critic of knowledge rather than reciter.
3. Judgment became the defended core. Verification carried clear evidence in 24 of 25 cases, and at least ten participants named professional expertise as its foundation. Reported checks ran from cross-model comparison (5 explicit adopters) and authoritative sources (4) to knowledge-base constraint (3), follow-up interrogation (2) and implementation testing (2).
4. Instrumental and partnership views were not alternatives. An instrumental formulation appeared in every case; a partnership formulation, which no protocol item introduced, appeared in 10 of 25 (seven clear, three partial). The partnership cases reported 1–3 h of use per day (median 2) against 1–4 h (median 2) in the other fifteen, so the vocabulary did not track intensity.
5. Teachers split on degeneration. Endorsement drew 9 cases (36.0%), conditional positions 11 (44.0%), explicit rebuttal 2 (8.0%), prolonged ambivalence 1 (4.0%) and no direct position 2 (8.0%). The distribution records positions toward a claim the interviewers supplied verbatim.
6. Defense was described more often than it was evidenced. The category had clear evidence in nine cases and partial evidence in sixteen. The specific moves — completing work independently before checking, retaining a right to err, forced thinking, deliberate avoidance in selected contexts, attention management — were volunteered inside a framed topic.
7. Irreplaceability was built out of the boundary. It appeared in 24 of 25 cases, but the filter-and-polarization idiom reached 23 of 25 coverage with clear evidence in only six, so the authors report it as a shared vocabulary rather than a firmly held belief.

What the study did

The design is a basic qualitative study analyzed with codebook thematic analysis. Materials were collected between January 2025 and May 2026; 29 teachers were approached, four declined, and 25 cases entered the analytic corpus. Twenty-two contributed semi-structured interviews of 60–120 minutes, and three (Teachers D, E and F) rested on field notes plus chat records, product documents or published essays, with no interview. Self-narratives added 11 ten-minute presentations from 10 participants delivered on 26 January 2025; the first author's own reflective writing is held out of every count as reflexive material.

The sample is 22 women and three men, aged 33 to 60 (Mdn = 43), with 7 to 30 years of teaching (Mdn = 16), working in 13 provinces and municipalities across institution types that extend well beyond the elite tier. Foreign-language teachers make up 10 of the 25 cases. Recruitment ran through purposive and snowball sampling inside an AI-focused learning community founded in December 2023, which the first author belonged to, so the study is explicitly insider research.

Coding used 20 codes (B1–B11 and C1–C9) assigned to belief, cognition, identity, affect and reported practice. Agreement on primary codes was 88.9% (40/45) with Cohen's κ = 0.881; excluding the first-author meaning unit it was 88.6% (39/44) with κ = 0.878. A 25 × 20 code-by-case matrix recorded clear, partial, counter or absent evidence per case, and a reporting convention — clear or partial evidence in at least three-quarters of cases and clear evidence in at least one third — was adopted after the matrix was compiled. Seventeen codes qualified under it. The authors also audited their own protocol and classified seven codes as prompted, eight as framed and five as unprompted.

Reading the claims against the protocol

The protocol audit is the most useful part of the method for anyone borrowing these findings. Participants were asked directly what AI cannot replace, whether they agreed that over-reliance causes "intellectual degeneration", and how they verify output. The boundary list, the degeneration positions and the verification talk therefore carry the shape of answers to supplied questions. What was volunteered — the partnership language, the metaphors for how AI works, the efficiency–depth tension, the learning-community talk and the renamings themselves — is where the study's more original claims sit.

The authors are unusually explicit that convergence has a competing explanation. Twenty-five teachers who meet regularly to discuss the same problem will describe it in shared terms, and part of what looks like independent agreement may be the circulation of a vocabulary. They also note that every participant was already using GenAI, that foreign-language teachers are over-represented at 10 of 25, and that disenchantment therefore appears in these accounts as a stage within continued use rather than as a reason to stop.

The belief–strategy–practice cycle in the discussion is offered as an interpretation of what teachers say, with arrows recording what participants describe as following from what. Nothing in a retrospective, largely cross-sectional, self-reported design tests direction of effect. There are no inferential statistics anywhere in the paper, and no student learning outcome is measured; the claims concern how this AI-engaged sample accounts for its work.

What this means for practice

• Instructors: the transferable move is task design that keeps thinking visible — an answer produced before any AI check, a justification of what a prompt contained, or an oral defense — because that is how the teachers here say they distinguish engagement from delegation.
• Administrators: professional development aimed at judgment, disciplinary depth and the metacognitive habit of monitoring participation fits what participants describe needing. The same accounts report that AI increased rather than reduced workload, so tool training alone addresses the wrong constraint.
• Designers: verification is the workflow these teachers describe running, and the five strategies named here are concrete requirements rather than abstractions. Cross-model comparison, bounded corpora and traceable sources are the features that would support them.
• Researchers: the next step is observational. Whether teachers enact the arrangements they describe, and what those arrangements do to students' monitoring and judgment, cannot be answered by asking teachers again.

Limitations

The sample is purposive and snowballed through one learning community, so teachers who considered GenAI and rejected it, and teachers who never encountered it, are absent, and the shared vocabulary cannot be separated from the convergence reported. Everything rests on self-report: stated beliefs are not observed practice. The design is largely cross-sectional and retrospective, so it can neither establish trajectories nor fix the direction of influence between belief and strategy. Several propositions were put to participants by the protocol, and broad code definitions inflate coverage. Three cases have no interview, and an absent code in a thinly documented case records the limits of the material rather than the absence of a belief. English quotations were translated by the corresponding author without an additional bilingual check.

Citation

Qiang, Z., & Wang, Z. (2026). Between Human and Machine: University Teachers’ Accounts of Judgment, Metacognitive Defense, and Cognitive Offloading in Generative AI Use. Journal of Intelligence, 14, 254. https://doi.org/10.3390/jintelligence14100254