AI Misconception and Adoption Among Indonesian K-12 Teachers

Synthesis: A 13-item mixed-format survey of 132 Indonesian K-12 teachers finds a striking mismatch between adoption and understanding: 96.2% support teacher AI use and 84.6% support student AI use, yet most describe Generative AI tools through high-level or mistaken mental models. Asked to explain what ChatGPT does with an instruction, 69.2% gave vague input-output descriptions and 21.5% held Misconceptions about AI that it searches a database or the internet. Teachers recognize the risks, dependency, skill erosion, and threats to Academic Integrity, but this awareness does not reduce their support. The study frames the gap as a Teacher AI Competency problem: educators rely daily on a black box they cannot evaluate, using Large Language Models (LLMs) tools for materials (32.3%), grading (26.9%), and planning (19.2%), while perceiving students' most common use as direct answer seeking (44.6%). Because the sample is one country, drawn from a self-selected pool of Self-Report Measures, the authors present it as a preliminary Global South case whose measurement approach and readiness tensions transfer to other structurally constrained systems.

Key Findings

1. Support outpaces understanding: 96.2% of the 132 teachers support AI use by teachers and 84.6% by students, yet 69.2% could describe ChatGPT's operation only as vague input and output.
2. Mechanism misconceptions are common: 21.5% of teachers said ChatGPT searches the internet or a database, and only 4.6% mentioned technical concepts such as tokenization, generative AI, or decoding.
3. AI is embedded across the teaching pipeline: teachers reported using it for generating materials (32.3%), grading and assessment (26.9%), information curation and administrative tasks (21.5%), and teaching planning and design (19.2%).
4. Teachers see students using AI mainly as a shortcut: 44.6% said students use it for direct answering, against 10.8% who saw tutor-like support and 9.2% who reported no use or no knowledge.
5. Efficiency drives support: 50.9% of supporters cited reduced teacher workload and 43% cited improved student learning, while the most common reason for opposing student use was dependency and learning erosion (34.6%).
6. Guardrails, not bans, define the boundary: teachers who opposed student AI use conditioned support on monitoring and clear limits (10.8%), and answer seeking and AI-written reports were the most contested classroom uses.

Conceptions of AI

Asked to define AI, 55.4% of teachers described a technology that mimics human intellectual ability and 40% defined it through practical uses, while only 4.6% framed it through risks or governance. The items probing mechanism exposed thinner understanding. When asked to explain what ChatGPT does after receiving an instruction, 69.2% described it as simply processing the instruction and returning an answer, often restating the question; a further 21.5% held a knowledge-retrieval misconception that the model searches the internet or a data center, and just 4.6% used technical terms such as tokenization, generative AI, or decoding. The authors read this as uneven AI Literacy: the tools are already in use, but their internal logic stays opaque. Teachers also anchored AI in their own professional world, since 45.4% named education and academic uses when asked about AI generally, and the most-cited reason people use AI was efficiency and productivity (55.4%).

How Teachers and Students Use AI

AI was reported across the whole teaching pipeline: 32.3% of teachers used it to generate materials such as slides and posters, 26.9% for grading or generating rubrics and answer sheets, 21.5% for information curation and administrative work, and 19.2% for planning and design such as teaching modules and learning steps. Teachers' picture of student use was far more lopsided. The single most common perception was direct answering for homework and assignments (44.6%), followed by information retrieval (19.2%) and content generation (16.2%); only 10.8% saw tutor-like support, and 9.2% reported no use or no knowledge at all. Most respondents viewed answer-seeking unfavorably, yet that judgment did not reduce their broader support for student AI use. The authors note that the familiarity is firmly professional: even when asked about AI in general, teachers reached first for academic and administrative examples, and the tool they described most often was a Conversational AI assistant.

Support, Risk, and Pedagogical Boundaries

Support was broad but not uncritical. Teachers backed AI for efficiency (50.9% cited reduced workload) and for improving student learning (43%), and opposed it mainly over dependency and learning erosion (34.6% for student use) and academic cheating and plagiarism (23.8%). A striking pattern is that risk awareness coexisted with permissive attitudes; the authors call this a permissive environment where convenience outweighs caution. When asked task by task, the most contested uses were answer seeking in performance-based tasks and AI-written reports, while boundaries softened for ideation and other stages. Conditional support was framed as Guardrails, meaning monitoring, verification, and clear limits rather than bans, and a minority worried that AI could replace the Teaching or reduce the development of social and emotional skills. The authors stress that Framing AI Use for Students should separate the cases where AI supports learning from those where it substitutes for the process.

Policy Context and Scaling Gap

Indonesia's national initiatives aim to mainstream AI and coding in schools from the 2025-2026 academic year, under the National Strategy for Artificial Intelligence 2020-2045. Readiness is uneven: roughly 65% of schools reportedly lack stable internet and 35% operate without reliable electricity, and in 2025 the ministry deployed 2,915 facilitators to train 62,898 teachers across 53,568 schools, against a system of 3,857,718 K-12 teachers in 276,415 schools. The authors frame the mismatch as an Educational AI Policy problem, since policy timelines move faster than teacher preparation. In their account the fix is not specialist expertise but functional Teacher AI Competency, the capacity to judge when a Generative AI output needs verification or risks undermining learning. They recommend longer, practice-oriented programs that require demonstrated responsible use, school-level AI committees, and rules stated in the syllabus with permissible purposes and AI Use and Disclosure Statements requirements.

What this means for practice

• Instructors. Judge AI tools by what they do to learning, not by what they save: teachers who most valued efficiency (50.9%) also most feared dependency and erosion (34.6%), so state plainly which tasks permit AI and which forbid it.
• Instructors. Correct the mechanism misconception directly, since 21.5% of teachers believed ChatGPT searches a database and only 4.6% named tokenization or decoding; inaccurate models shape monitoring and the normalization of AI-supported work.
• Instructors. Require students to document their AI process, because 44.6% of teachers saw students using AI mainly for direct answers and answer seeking plus AI-written reports were the most contested tasks.
• Faculty developers. Replace awareness training with performance-based programs: the ministry trained 62,898 teachers through 2,915 facilitators, yet most respondents still held superficial or mistaken mental models.
• Administrators and policymakers. Write AI rules into the syllabus with permissible purposes, disclosure requirements, and prohibited uses, and form school AI committees to adapt national guidance to local conditions.

Limitations

• The sample is not nationally representative: 132 teachers, 61.4% (n=81) from West Java, with limited private-school representation outside West Java (9.8% private), and two participants flagged as AI-generated were omitted from the final analysis.
• Recruitment ran through professional word-of-mouth networks and self-selected Google Form registration, which may bias the pool toward technology-engaged educators; outside West Java, participants were drawn by random sampling because the pool of registrants there was limited.
• The study relies entirely on self-reported survey responses rather than observed classroom behavior.
• The analysis is limited to descriptive patterns and reports no inferential subgroup comparisons.
• Qualitative themes come from a human-in-the-loop workflow (LDA topic generation, refinement with ChatGPT GPT-5.2, and rubric application by GPT-5 mini) reporting 95.6% agreement with human judgment on a 10% sample, so coding judgment rests with the researchers.

Citation

Aji, A. F., Amriani, A., Chevi, R., Purwarianti, A., & Wijaya, D. (2026). AI Misconception and Adoption Among Indonesian K-12 Teachers. arXiv:2610.26758.