The open problem in AI-driven education: What 33 studies still couldn't answer

Thirty-three peer-reviewed papers, almost all published in 2025 or 2026, reach the same conclusion: AI tools are being integrated into classrooms faster than anyone can measure the consequences. Across contexts ranging from early childhood mathematics in Indonesia to pre-service teacher training in Europe, the same three gaps reappear so consistently that a systematic embedding analysis of the literature flagged them as a single, strong, unresolved cluster. You can browse the underlying evidence on our research-gaps page: AI in education — longitudinal, cross-cultural, and algorithmic-fairness gaps.

This post unpacks what those 33 studies found, what they could not answer, and what the next generation of researchers would need to do to actually close these gaps.

What the literature says

The 2026 literature on AI in education is striking for its breadth and its uniformity. Researchers are studying ChatGPT in secondary mathematics classrooms, augmented reality in STEM learning, AI-assisted academic writing, automated assessment systems, and AI integration in pre-service teacher training. The methodologies vary — bibliometric analyses, randomized controlled trials, design-based research, systematic reviews. Yet the findings converge.

AI tools produce measurable short-term benefits. A study by Lee and colleagues examining student–AI interaction in university writing tasks found that when students were asked to justify whether they adopted, adapted, or rejected AI suggestions, they developed critical awareness and ethical reasoning alongside their writing skills (DOI: 10.65553/als.260102). A separate ADDIE-based study on AI-powered mathematics learning media in Indonesian early childhood settings showed that adaptive feedback and personalized challenge levels supported children's grasp of foundational mathematical concepts (DOI: 10.12973/eu-jer.15.3.843).

The research base is expanding rapidly but remains shallow. A bibliometric review of AI in educational assessment mapped the growth of publications in this space and found that while thematic clusters are forming around personalized learning, automated feedback, and adaptive testing, most studies rely on small samples, short timeframes, and geographically narrow contexts (DOI: 10.29333/iji.2026.19238a).

Pre-service teacher preparation is lagging. A systematic review of AI integration in pre-service teacher education found that teacher training institutions have been slow to equip future educators with the skills to critically evaluate, responsibly deploy, and ethically govern AI tools in their classrooms (DOI: 10.30935/cedtech/18458). Without prepared teachers, even well-designed AI tools underperform.

Equity and access remain structural problems. A multidisciplinary review of opportunities and challenges in AI-driven teaching and learning identified digital inequality as a persistent barrier, with rural and low-income settings systematically excluded from both the interventions and the research samples (DOI: 10.70096/tssr.260402064).

What is unresolved

Three gaps emerge from this cluster with particular consistency — and none of them is close to being answered.

1. We have almost no longitudinal data. The vast majority of AI-in-education studies measure outcomes over a single semester or less. The fundamental question — does early exposure to AI-assisted learning produce durable gains in knowledge, motivation, or self-regulation — is essentially unanswered. Short-term improvements in test scores or engagement metrics are plausible proxies, but they are not the same as evidence that AI integration is building the kind of deep competencies that persist and transfer. The gap here is not just methodological laziness; longitudinal studies in education are expensive, institutionally complex, and prone to attrition. But the absence of this evidence is now a genuine policy problem, as governments and universities are scaling AI integration without a long-term evidence base.

2. Cross-cultural validity is almost entirely untested. Most of the 33 papers in this cluster were conducted in a handful of countries: Indonesia, the Philippines, South Africa, the United Kingdom, and several European nations. The AI tools being studied — large language models, adaptive learning platforms, automated assessment engines — were designed predominantly in and for high-income, English-language contexts. How these tools perform when the instructional language, the curriculum structure, the pedagogical tradition, or the socioeconomic conditions differ substantially from the design context is largely unknown. The few cross-context comparisons that exist find meaningful variation, but they are not designed to identify why that variation occurs or what moderates it.

3. Algorithmic bias and fairness are discussed but not measured. Paper after paper acknowledges that AI tools may embed and amplify existing biases — against students from minority backgrounds, against non-standard dialects, against students with disabilities, against those whose writing or problem-solving style deviates from the modal training data. Yet almost none of the 33 papers actually measures this. The acknowledgment appears in limitations sections and future-work paragraphs. It has not yet made it into study designs, outcome measures, or reporting standards.

What would move this forward

Closing these three gaps requires different things.

For longitudinal evidence, the field needs multi-year cohort studies embedded in institutional contexts — schools and universities that are already using AI tools and willing to instrument their use over three to five years. The design challenge is tracking not just AI adoption but learning outcomes, persistence, and life outcomes at reasonable follow-up horizons. Collaborative research consortia, not individual lab studies, are probably the right vehicle.

For cross-cultural validity, researchers need pre-registered replication studies that explicitly target cultural and linguistic variation. This means translating and adapting AI interventions, measuring fidelity across contexts, and reporting disaggregated results. It also means building research partnerships between institutions in the Global South and AI developers, rather than the current pattern of Global North researchers studying Global South classrooms.

For algorithmic fairness, the field needs agreed measurement instruments — analogous to the fairness metrics that have emerged in hiring and criminal justice AI — adapted for the specific constructs at stake in education: accuracy of automated assessment across demographic groups, differential impact of personalized recommendation on student choices, bias in AI-generated feedback on writing from non-native speakers. This is a technical and normative problem simultaneously. It requires AI developers, educational researchers, ethicists, and student advocates to collaborate on study designs before data is collected.

How to contribute

If your research touches any of these three gaps — longitudinal AI impact in education, cross-cultural validity of AI tools, or algorithmic bias in educational systems — Science AI Journal welcomes submissions. Our peer review process uses specialized agents trained on educational research methodology, and we assess both empirical rigor and the clarity of contribution to the literature's open problems.

Start by exploring the full evidence cluster: AI in education — longitudinal, cross-cultural, and algorithmic-fairness gaps. The page shows the 33 contributing papers, their geographic spread, and the specific gap statements that were clustered into this synthesis. If your work addresses any of these directly, submit your paper for review.

Frequently asked questions

Why don't we have long-term data on AI's impact in education?

Longitudinal studies in education are costly and methodologically demanding. They require sustained institutional partnerships, stable research infrastructure, and careful management of attrition over multiple years. Most academic funding cycles are shorter than the follow-up periods needed to detect durable learning effects, which pushes researchers toward shorter-term designs. The result is a literature rich in proof-of-concept studies and thin on evidence about whether short-term gains persist.

What is algorithmic bias in educational AI, and why does it matter?

Algorithmic bias in educational AI refers to systematic differences in tool performance across student groups — for example, an automated essay scorer that consistently gives lower marks to writing from non-native English speakers, or an adaptive platform whose recommendations steer students from underrepresented backgrounds away from advanced coursework. These biases can arise from the training data, the optimization objective, or the way outcomes are defined. They matter because AI tools in education are increasingly being used to make or inform high-stakes decisions: grades, feedback, course recommendations, and access to academic support.

How can researchers contribute to closing these gaps?

The most direct contributions are pre-registered replication studies across diverse cultural and linguistic contexts, longitudinal cohort studies that track AI-assisted learning over at least two academic years, and intervention studies that measure fairness outcomes — not just aggregate performance — across demographic groups. Methodological contributions are also valuable: validated instruments for measuring AI literacy, cross-cultural adaptation frameworks for AI-based pedagogical tools, and reporting standards that require disaggregated results.