AIEd Knowledge BaseAIEd KB
On this page

Research Article

Where Artificial Intelligence Enters Teacher Work

added 2026-09-21 · updated 2026-09-21

Synthesis: Using TALIS 2024 data from 56,669 teachers, with a primary allocation analysis of 24,058 AI users across 46 education systems, this study separates AI adoption from task allocation: which parts of a teacher's work the tool actually enters. AI use rises with perceived task demand, and within a teacher it tracks the task they find harder — but the pattern is not uniform. Lesson planning and Special Education support attract AI (OR = 1.084 and OR = 1.442), while Assessment and marking pull away from it (OR = 0.921), the opposite of what a simple effort-reduction story predicts.

Key Findings

  1. Adoption rises with demand but only modestly. A one-point increase in average task demand across three matched domains was associated with 8.4% higher odds of reporting any AI use (OR = 1.084, 95% CI [1.050, 1.120], p < .001).
  2. Allocation is within-teacher. A focal task one point above the same teacher's own mean demand was associated with higher odds of AI use on that task (OR = 1.163, 95% CI [1.127, 1.200], p < .001).
  3. Task effects diverge sharply from one another (F(2, 48) = 52.53, p < .001): lesson planning OR = 1.084, assessment/marking OR = 0.921, and special-education support/adaptation OR = 1.442.
  4. Predicted allocation spreads widely. Moving from one point below to one point above a teacher's mean demand raised predicted planning use from 68.5% to 71.6%, lowered assessment/marking use from 37.1% to 33.6%, and raised special-education support use from 34.0% to 50.3%.
  5. The prespecified task mapping beat the alternatives. The intended three-pair demand-to-task correspondence produced the strongest within-teacher alignment (OR = 1.167) of all six possible assignments, ahead of the next strongest at OR = 1.118.
  6. The analyzed sample was filtered for stability. Systems entered the allocation analysis only with at least 150 AI adopters, 30 sampled schools, and AI-module observation for 20% of sampled teachers across 46 nonoverlapping systems.
  7. Most AI use is preparation, not grading. Among AI users, 64% reported generating lesson plans or activities with AI while 26% reported using it to assess or grade student work.

Adoption and allocation are different questions

The paper's methodological point is that asking whether teachers use AI tells you nothing about where the tool lands in their work. Adoption here is a teacher-level yes/no; allocation is a within-teacher comparison between tasks, which is why the analysis can separate a teacher who uses AI for everything from one who uses it only where they feel most stretched. The design also holds teacher and system constant by construction, so the estimated task effects are not contaminated by the fact that teachers who feel overloaded differ from those who do not. The strongest system-average predictor remained modest, which the authors read as evidence that demand shapes use within a teacher's week more than it sorts teachers into adopters and non-adopters.

The assessment inversion

The finding with the most direct implication is that assessment and marking are the one domain where higher perceived demand predicts less AI use — a negative slope, not a null one. The plausible reading is that grading carries accountability, judgment and integrity stakes that planning does not, so a teacher under time pressure may still keep the marking in their own hands. Whatever the mechanism, the direction matters for policy: mandates that encourage AI in assessment are pushing against a pattern that teachers are already following, and the paper's own institutional framing acknowledges that guidance must account for where staff themselves draw the line.

Special-education support as the demand-driven case

The largest allocation effect, OR = 1.442 for special-education support and adaptation, identifies the work where teachers most reach for AI when it feels demanding. Predicted use there rises by more than 16 percentage points across the demand range. This is a double-edged result: the same adaptation task that AI makes tractable is the one where generic outputs are least likely to be adequate, since support plans are individualized and legally consequential. It also points to a training gap, because teachers are most likely to use AI unsupervised in the most sensitive part of their practice.

What this means for practice

  • Policymakers. Target guidance at the tasks where use concentrates. Support and adaptation planning is where teachers reach for AI under pressure, and it needs the most explicit standards rather than blanket encouragement or prohibition.
  • Administrators. Treat planning and grading differently in professional development. The data show teachers already keep high-stakes marking human, so training effort belongs on judgment calibration for the adaptation work they are doing with AI.
  • Researchers. Study allocation, not just adoption; a teacher-level adoption indicator cannot distinguish the preparation assistant from the grading assistant and will misread both.

Limitations

  • TALIS is cross-sectional and self-reported: demand is the teacher's perception of their work, not an observed time-use measure, so the analysis is correlational and cannot establish that demand causes AI use.
  • The allocation sample covers 46 nonoverlapping education systems after size rules of at least 150 AI adopters, 30 schools and 20% AI-module observation, which trades system coverage for estimate stability.
  • Country-level and system-level moderation is summarized rather than interpreted case by case, so pooled slopes may not describe any single education system.
  • Task demand is measured at the domain level, so a teacher's single rating stands in for a whole category of work such as all assessment and marking.

Connected Concepts

Connected Articles

Citation

Holster, J. (2026). Where Artificial Intelligence Enters Teacher Work. EdArXiv.

Metadata

Study designSecondary Analysis
LevelSecondary
Intended audiencePolicymakers, Administrators, Researchers
Learning and instructionCareer Development and Readiness
Technologies and techniquesGenerative AI
Assessment and measurementAutomated Assessment
Research methods and evaluationQuantitative Research
Institutions and policyEducational AI Policy
Foundations of AI in educationTeaching
Page kindEvaluation
SourceHistoryEditChat

Embed this page

Copy the code below to embed a chromeless version of this page in a learning management system or other website. The embedded view hides the site header, navigation, and footer.