Fifty years of topic development, thematic combination, and AI's evolving reach in Computers & Education

Synthesis: Wu and colleagues model the first fifty years of Computers & Education — 5,739 records assembled, 5,387 eligible articles and reviews — with an eighteen-topic structural topic model, tracing prevalence across five historical periods and the ordered pairing of primary and secondary topics inside each article. The ten most frequent topic pairs fall from 74.4% of documents in 1976–1989 to 16.5% in 2010–2019, a diversification the authors treat as the journal's real intellectual history. A dictionary-based, human-validated AI corpus of 390 records (7.2%) spreads across 17 of the 18 topics, and annual AI prevalence accelerates sharply after 2018. The authors insist this records research attention, not demonstrated effectiveness.

Key Findings

1. A fifty-year corpus, modeled as eighteen topics. The authors assembled 5,739 records, retained 5,387 eligible articles and reviews, and fitted an eighteen-topic structural topic model over 2,412 terms and 471,226 tokens.
2. Four concerns anchor the topic structure. The four most prevalent topics — learning achievement, teacher integration, computer-assisted learning in engineering and mathematics, and educational tool design — account for 33.9% of modeled text, led by technology interventions and learning achievement at 9.6%.
3. Within-article combinations diversified sharply. The ten most frequent primary–secondary topic pairs represented 74.4% of documents in 1976–1989 but only 16.5% in 2010–2019, while observed ordered pairs rose from 67 to 261.
4. The recent period is slightly less diverse, not more concentrated. In 2020–partial 2026 the archive shows 211 observed pairs and a top-ten share of 18.7%, a modest decline from the 2010s peak that the authors attribute partly to the shorter window.
5. AI is distributed across the curriculum of concerns. The validated AI corpus contains 390 records (7.2%) and appears in 17 of the 18 topics; self-regulated learning and personalized support carries 85 of 299 records (28.4%), intelligent tutoring 81 of 342 (23.7%), learning analytics 54 of 253 (21.3%), and language learning 23 of 120 (19.2%).
6. AI publication prevalence turned upward after 2018. The annual odds that an article was AI-related rose by 21.8% per year from 1976 to 1989 (OR = 1.218), declined by 3.7% annually through 2018 (OR = .963), then increased by 40.1% annually to 2025 (OR = 1.401), reaching a 30.9% share in 2025 and 52.1% in partial 2026.

A journal history told through topic mixtures

The corpus covers the journal from its founding through June 16, 2026. Web of Science supplied 5,715 records from 1978 onward; 24 more came from the journal website to cover 1976 and 1977. Eligibility screening left 4,960 articles, 258 records indexed as both article and proceedings paper, and 169 reviews. Titles and abstracts were combined where an abstract existed, and 84 documents remained title-only. The model treats each document as a mixture of eighteen topics rather than assigning it to one, allowing later analysis of co-occurrence.

The eighteen topics are broad but interpretable: computer-assisted learning in engineering and mathematics, educational tool and web resource design, intelligent tutoring and adaptive testing, learning behavior analytics, self-regulated learning and personalized support, and digital literacy and the digital divide. The authors group them into three orientations — technical systems and methods (45.2%), educational activity and experience (26.5%), and the human and social conditions of technology use (28.3%) — and read the balance as evidence that the journal never organized itself around a single technology.

The historical trajectories follow the technology chronology only loosely. Computer-assisted learning exceeded 40% of modeled text in the late 1970s before declining through the 1980s; educational tool and web resource design peaked in the mid-1990s; technology interventions and learning achievement peaked in the mid-2010s; and self-regulated learning and personalized support became the most prominent topic by the end of the series. The authors align these movements with a technology chronology but refuse causal reading.

Diversification inside the article

The paper's distinctive move is to keep each article's full topic mixture and analyze the ordered pair formed by its largest and second-largest topics. That pair accounts for more than half of modeled text in 64.73% of documents.

Concentration falls steeply and then stabilizes. In 1976–1989, the leading pair T3→T4 alone covered 20.1% of documents and T3→T7 another 15.6%; computer-assisted learning was the primary topic in eight of the ten leading pairs. By 2000–2009 the largest pair was only 2.6%, and in 2010–2019 the largest was 2.3%. Effective diversity, measured with Hill numbers, peaks in 2010–2019 at 177.4 for order 1 and 132.6 for order 2, with normalized entropy of .905. The final period slips to 145.7 and 110.0 but remains far above the early archive.

The authors read this as a change in research problems, not merely vocabulary. Later articles pair technical systems with teacher capacity, learner regulation, analytics, access, and feedback in many more configurations; no single concern organizes the archive the way computer-assisted learning once did. They are careful about what diversity can and cannot show: a broader distribution of topic pairs establishes a wider framing, not greater theoretical integration or better evidence.

AI as a layered presence rather than a recent arrival

AI-related research is identified with a controlled dictionary of 69 entries in five families, screened automatically then validated by human review, which excluded 21 of 411 candidates. The strict corpus of 390 records is bounded by a title-core count of 208 and an expanded count of 507; its overlapping families are umbrella AI (157), intelligent, adaptive, and agent systems (175), machine learning and perception (118), and generative and foundation models (70).

The distribution is historically legible. Intelligent tutoring and adaptive testing holds 62 intelligent-systems records, or 18.1% of that topic's articles; learning behavior analytics holds 43 machine-learning records (17.0%); self-regulated learning is the convergence point for all four families, with 45 umbrella, 39 intelligent-system, 18 machine-learning, and 30 generative records. Generative and foundation models appear only in the final period. Across the full archive, weighted AI incidence is highest in self-regulated learning (16.8%), intelligent tutoring (15.1%), learning analytics (14.4%), language learning (13.0%), and video lectures and feedback (10.2%).

The temporal pattern is the paper's sharpest correction to prevailing narratives. AI was not absent before generative models: it grew rapidly in the late 1970s and 1980s around student modeling, expert systems, and intelligent tutoring, declined as a share of a growing archive, and then accelerated after 2018 as generative models joined rather than replaced the older lineages. The authors state plainly that the acceleration measures publication attention and does not establish that AI-supported interventions became more effective.

What this means for practice

• Do not read topic growth as evidence of learning benefit. The paper's own framing separates prevalence from effectiveness; a rise in AI-related articles, including the 30.9% share in 2025, documents what researchers study, not what works for students.
• Evaluate AI inside the educational configuration it enters. Because AI records appear in 17 of 18 topics, the useful question is which function a system performs, for whom, under what pedagogy and assessment regime, and against which outcome — not whether a document mentions AI.
• Treat self-regulation and writing as the current pressure points. The concentration of AI work in self-regulated learning and personalized support (28.4% of that topic's records) and in language learning and higher-order thinking (19.2%) marks where questions about learner agency, evaluation, and authorship are being worked out.

Limitations

• The model reads titles and abstracts, not full texts, so topics describe stated emphases rather than what studies actually did; 84 records had no retrievable abstract and were represented by title alone.
• The eighteen-topic solution depends on stemming, vocabulary thresholds, a manual stop list, and model-selection judgments; adjacent solutions could merge or split related topics, so fine distinctions deserve less weight than broad trajectories.
• The five periods are interpretive windows of unequal length, and the last includes partial 2026; period comparisons, including the diversity decline, remain descriptive rather than tested.
• Primary–secondary pairs summarize only the two largest topic proportions and fully represent just 64.73% of documents; they are principal configurations, not complete content.
• AI identification depends on terminology visible in titles and abstracts, so earlier work described in domain-specific language may be missed; the strict count is best read as a transparent lower-bound estimate.

Citation

Wu, J.-Y., Shih, A.-C., Lee, Y.-H., Chai, C. S., & Tsai, C.-C. (2026). Fifty years of topic development, thematic combination, and AI's evolving reach in Computers & Education. Computers & Education, 255, 105755.