Human Agency Framework Offers Four-Part Test for AI in Education

A PRISMA-guided systematic review has produced a four-part framework for asking whether artificial intelligence in education expands people’s freedom to learn and work rather than simply automating tasks. The proposed Human Agency Framework, or HAF, organizes its assessment around purpose, control, collaboration and conditions: the valued capabilities learners develop, their influence over pathways and data, the quality of human interaction, and the institutional environment supporting autonomy. The authors describe the result as a conceptual diagnostic tool, not a validated measure of educational or employment outcomes.

The framework was developed through a systematic review of 35 sources published between January 2018 and March 2026, supplemented by targeted searches of UNESCO, OECD and European Union repositories and by backward and forward citation tracking. Searches produced 2,910 records before deduplication and 2,255 records screened by title and abstract. Of 525 sources assessed in full text, 35 met the inclusion criteria. Eligible work explicitly examined agency, autonomy, empowerment, learner control, capability or participation in AI-mediated education or workforce development.

Two reviewers independently screened and coded the evidence, resolving disagreements through discussion and, when necessary, adjudication by a third reviewer. Coding took place in NVivo 14 through open, axial and selective stages. The researchers treated a sentence, paragraph or group of consecutive sentences expressing a complete idea as the unit of analysis, a choice designed to capture agency as a contextual concept rather than as an isolated keyword. The final codebook contained 62 codes, which were consolidated into 12 axial categories and then four higher-order dimensions. Agreement on a random sample representing 20% of coded text was assessed with Cohen’s kappa, yielding a value of 0.83, described as substantial.

The authors combined this synthesis with a method they call Diagnostic Design Synthesis. It builds on framework synthesis and thematic analysis by adding an expert review and refinement stage. Five practitioners—two K–12 educators, a corporate learning and development manager, an educational technology developer and a policy adviser from a national skills agency—reviewed the proposed dimensions, questions and indicators. They recommended adding attention to data portability and long-term career pathways, applied the draft framework to two initiatives and commented on its clarity and usefulness. The practical experience therefore informed the framework’s design, but it did not constitute formal validation.

Purpose asks what substantive capabilities an AI initiative enables people to achieve, drawing on the capability approach associated with Amartya Sen and Martha Nussbaum. Rather than treating test scores or technical proficiency as sufficient, this dimension considers whether learners can set their own goals and pursue outcomes they value, such as critical thinking, creativity, civic participation or personal well-being. Indicators of a strong purpose include learner involvement in goal-setting, curricula that emphasize broad capabilities, and documented improvements in self-efficacy and sense of meaning. The central distinction is between the efficiency of a system and the freedom it gives its users.

Control examines who directs learning pathways, data and design decisions. Diagnostic questions ask whether learners can select different routes or reject algorithmic recommendations, how consent and deletion are handled, and whether learners and educators helped design the system. The framework treats meaningful choice, transparent data policies and participatory design as signs that people retain influence over systems that may shape their education. This dimension also incorporates the source review’s attention to human oversight and transparency, including regulatory approaches that classify educational AI as high risk, while reframing legal protections as conditions that can enable agency rather than as compliance obligations alone.

Collaboration examines agency with whom. The framework extends human–AI collaboration typologies beyond the replacement or augmentation of teacher judgment to include relationships among learners, peers, educators and AI systems. Questions ask whether AI replaces or augments human interaction, whether educators are prepared to co-design learning experiences with it, and whether it supports peer-to-peer learning and collective problem-solving. Relevant indicators include tools that facilitate group work, professional development in co-orchestration and evidence of collaboration among students. Here, agency is treated as relational and distributed: an efficient system that isolates its users may still weaken their capacity to learn together.

Conditions addresses the structural setting in which agency occurs. Even a well-designed tool may offer little practical autonomy if leadership does not support it, access is unequal, data protections are weak or training ends when a pilot does. The proposed indicators include explicit leadership support for learner autonomy, documented equity and privacy policies, long-term technical support, professional development and multi-year sustainability plans. The dimension also broadens the temporal horizon of assessment, asking not only whether an initiative works at launch but whether users can navigate learning and career pathways over time, including through the data portability added during expert refinement.

To demonstrate how the HAF might be used, the authors applied it illustratively to two initiatives using publicly available documents and secondary sources: a UNESCO–Huawei AI capacity-building project in Arabic-speaking countries and Central Asia, and Coursera’s AI-powered career guidance tool, which recommends courses and career paths from user profiles and activity. The first combined teacher training, student resources and locally relevant content, while the second offered individual and institutional workforce-development users algorithmic guidance. The ratings were interpretive assessments of publicly visible evidence, not empirical evaluations, performance rankings or measurements of agency outcomes.

Several limitations constrain the proposal. The framework has not been tested across diverse real-world settings, and the review included only English-language sources, potentially omitting culturally distinct accounts of agency, especially from the Global South. Its 35-source corpus was relatively small and was not subjected to formal quality appraisal, although peer-reviewed publications and reports from established international organisations were preferred during selection. Coding still depended on researchers’ judgments, despite the 0.83 kappa, and the five-member practitioner panel was too small to represent broad professional experience. The literature and case examples were also weighted toward Western and Global North settings. The next stage would need applied testing in schools, universities, corporate training and informal learning, standardised quantitative measures, cultural adaptation and examination of how more autonomous AI systems may alter the balance of power between users and machines.

Subject of Research: A conceptual framework for evaluating human agency in AI-mediated education and workforce development.

Article Title: A diagnostic human agency framework for AI driven education from a PRISMA guided systematic review

Article References: Acharya, K., & Gondaliya, M. (2026). A diagnostic human agency framework for AI driven education from a PRISMA guided systematic review. Discover Education, 5(1), Article 1206. https://doi.org/10.1007/s44217-026-02173-7

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02173-7

Keywords: artificial intelligence in education, human agency, learner autonomy, educational technology, systematic review, AI ethics, workforce development, human-AI collaboration, data governance, diagnostic, human, agency