Activity Theory and AI: Why the Tool Is Never the Whole Story

Most AI rollouts are judged on the tool: accuracy, speed, cost per call. Yet the same tool lifts one team and stalls the next. Activity theory explains why. It treats the whole activity as the unit of analysis, not the tool, and it expects friction wherever a new tool meets old rules.

Activity theory in brief

Cultural-historical activity theory (CHAT), as modelled by Yrjö Engeström, describes purposeful work as a system of six linked elements. Change one and the others push back.

• Subject — Who is doing the work? With AI: the learner, analyst or teacher using the system.
• Object — What are they working on, and toward what outcome? With AI: the essay, the incident, the hiring decision.
• Tools — What mediates the work? With AI: the model, the prompt, the dashboard, the score.
• Rules — What norms and policies govern it? With AI: academic integrity codes, audit duties, approval steps.
• Community — Who else has a stake? With AI: peers, managers, regulators, students' families.
• Division of labour — Who does which part? With AI: what the model drafts, and what the human checks and signs.

The key idea is contradiction: a tension between elements, such as a new tool that existing rules forbid or ignore. Contradictions are not bugs. They are where an activity learns and changes (Engeström, 2001).

What recent research shows

Studies from 2024 to 2026 show the same pattern. The AI tool lands quickly, and the rules and roles around it lag behind.

• New rules appear from below. In interviews with 25 engineering students, Leino Lindell and Stöhr (2026) found that tensions between traditional tools and generative AI pushed students to invent their own implicit norms for acceptable use. The division of labour between student and teacher shifted with them.
• The tool becomes the object. Bakogianni, Liljekvist and Bui (2026) interviewed ten mathematics academics. GenAI began as a mediating tool and grew into a "runaway object" that reshaped workflows, collaboration and departmental norms.
• Every element moves at once. In a CHAT case study of AI-based digital textbooks in two Korean high schools, Choi (2026) found that adoption changed all six components of the activity system. Contradictions around technical instability, data privacy, funding and equity became catalysts for redesign.
• Support, not substitution. Chapwanya (2025) studied further education teachers using ChatGPT and TeacherMatic. Efficiency and personalisation improved, but the tools supplemented teacher mentorship rather than replacing it.
• The tool works; the rules cap it. Susnjak, Mathrani and Bakhshov (2025) treated an AI success-probability score as a mediating tool for student advising. Supported students failed fewer courses, but fixed course loads and prerequisite chains limited completion gains.
• Community drives use. Across Nigerian universities, Essien and colleagues (2024) found that peer norms and ease of use shaped students' engagement with generative AI.

In each case the outcome depended on rules, community and division of labour, not only on the model.

Six questions to ask before you deploy AI

• Subject: Whose work changes, and did they help define the change?
• Object: What outcome are we improving, and what is the baseline?
• Tools: What does the model decide, suggest or only draft?
• Rules: Which policies now conflict with how the tool will really be used?
• Community: Who outside the team will feel the effect?
• Division of labour: Who checks the output, and who signs off when it is wrong?

The takeaway

Treat an AI rollout as a change to an activity system, not a tool install. Find the contradictions early, redesign the rules and roles around them, and measure the outcome, not the model.

References

Bakogianni, D., Liljekvist, Y., & Bui, P. (2026). GenAI as a runaway object in higher education: A socio-cultural view on AI-influenced academic practice in mathematics. Computers and Education: Artificial Intelligence.

Chapwanya, O. (2025). Exploring the teacher's role amid rising generative AI: An activity theory analysis in further education. Studies in Technology Enhanced Learning, 4(3). https://doi.org/10.21428/8c225f6e.4d227fbd

Choi, W. (2026). Analysis of contradictions in implementing AI-based digital textbooks (AIDTs) in Korean high schools: Cultural-historical activity theory (CHAT) perspective. Sage Open, 16(3). https://doi.org/10.1177/21582440261463040

Engeström, Y. (2001). Expansive learning at work: Toward an activity theoretical reconceptualization. Journal of Education and Work, 14(1), 133–156.

Engeström, Y. (2015). Learning by expanding: An activity-theoretical approach to developmental research (2nd ed.). Cambridge University Press. (Original work published 1987)

Essien, A., et al. (2024). Exploring socio-cultural influences on generative AI engagement in Nigerian higher education: An activity theory analysis. Smart Learning Environments, 11, 63. https://doi.org/10.1186/s40561-024-00352-3

Leino Lindell, T., & Stöhr, C. (2026). The AI disruption in engineering education: An analysis of changing student norms through cultural historical activity theory. Journal of Computing in Higher Education. https://doi.org/10.1007/s12528-025-09488-8

Susnjak, T., Mathrani, A., & Bakhshov, E. (2025). Stabilising learner trajectories: A doubly robust evaluation of AI-guided student support using activity theory. arXiv:2512.11154.