Beyond ChatGPT: A Comparative Mapping of AI Research in STEM and Engineering Education and the Contemporary AI Tool Ecosystem

Synthesis: Terkowsky, Kaufhold, Meimers, Göring, Boettcher and Wilkesmann mapped the research visibility of 85 curated AI tools against the wider tool ecosystem in STEM and engineering education, using 340 standardized searches and semantic screening of titles and abstracts. Of the 85 tools, only 36 appeared anywhere in the literature, and Conversational AI and LLMs accounted for 73.7% of the 927 positive paper–tool connections across 592 publications, while only 23 publications had an established laboratory context. The authors call the mismatch a complexity gap: the tool ecosystem is functionally diverse and the research literature is not, and AI in education research is therefore studying a narrow slice of what students and laboratories could use.

Key Findings

1. Most of the curated ecosystem is invisible in research. Only 36 of 85 AI tools and systems appeared in the screened literature, so 49 tools had no positive paper–tool connection at all.
2. Conversational AI and LLMs dominate the literature. That category accounted for 73.7% of the 927 positive paper–tool connections identified across 592 unique publications.
3. Laboratory contexts are poorly represented. Only 23 of the 592 publications had an established laboratory context; the other 569 were classified within STEM or engineering education without one.
4. Retrieval was broad, screening was semantic. 340 standardized searches produced 14,378 raw retrieval events, reduced to 5,782 records, then to 6,582 unique paper–tool candidates representing 4,588 publications before semantic screening.
5. The design compares a curated ecosystem with its own research footprint. Because the same 85-item sampling frame drove retrieval and comparison, the study measures research visibility within a defined ecosystem rather than the whole AI-in-STEM literature.
6. Screening used titles and snippets, not full texts. Semantic coding relied on publication titles and consolidated Google Scholar snippets, which the authors name as the study's central limitation.

A mapping review, not an effectiveness review

The method is a mapping review with a matched sampling frame: the authors curated 85 contemporary AI tools, then searched the literature for each named tool and recorded whether a substantive educational connection could be established. That design answers a question effectiveness reviews cannot — not "does AI work" but "which AI is being studied at all" — and it produces the paper's cleanest number, the 36-of-85 visibility rate.

The concentration finding has a familiar shape. Chatbots and LLMs are cheap to deploy, familiar to researchers, and easy to study with a survey, so they accumulate publications; a simulation suite or a laboratory instrument requires infrastructure and specialist expertise, so it accumulates fewer. The authors read the pattern as a research-practice mismatch rather than as evidence about learning.

Why the laboratory gap matters

Laboratory learning is where the STEM disciplines make claims about equipment, procedure and practical competence, and it is precisely where AI support would need to be domain-specific. Finding 23 laboratory-context publications against 569 without one suggests the literature is largely studying AI in the classroom and lecture rather than in the lab, which limits what it can say about the settings engineering educators care most about.

The paper's recommendation follows directly: move from tool-centered to function-oriented and context-sensitive research, asking what form of AI support develops competence in a given setting rather than whether a chatbot helps.

How far the claim reaches

Two constraints bound the result. The sampling frame is 85 tools chosen by the authors, so tools outside it cannot appear and the study is a mapping of that ecosystem rather than of all AI in STEM education. And screening was semantic, on titles and consolidated snippets, so a paper that used a listed tool without naming it in the title or snippet would be missed, which biases the visibility estimate downward in ways the authors cannot quantify.

The ecosystem is also a snapshot: the curated set represents the tool landscape at the time of curation, and the tool landscape moves faster than the publication cycle.

What this means for practice

• Look past the chatbot when choosing what to study or adopt. 73.7% of research attention went to conversational AI, so an absence of evidence about a simulation, instrument or analytics tool is a gap in the literature, not evidence that the tool does not work.
• Fund and study laboratory-context AI explicitly. With 23 laboratory publications in the dataset, engineering programs evaluating AI for lab work have almost no comparative base to draw on.
• Read a mapping review as a map of attention. The 36-of-85 figure describes what researchers studied, not what institutions deployed or what students used.
• Ask function questions in local evaluations. "What support does this task need" is answerable within an institution even where the published literature is silent.

Limitations

• Screening used titles and Google Scholar snippets, not full texts. Snippets are generated for retrieval rather than evidence extraction, so eligible settings and substantive paper–tool connections may have been missed.
• A predefined 85-tool sampling frame. The matched design enables systematic comparison but is not an exhaustive search; papers on tools outside the frame were not captured.
• The ecosystem is a moving snapshot. The curated tool set represents one moment, and its relevance decays as the tool landscape changes.
• No effectiveness or learning-outcome data. The study maps research visibility, so it says nothing about whether any listed tool improves learning.

Citation

Terkowsky, C., Kaufhold, N., Meimers, A., Göring, J., Boettcher, K. E. R., & Wilkesmann, U. (2026). Beyond ChatGPT: A Comparative Mapping of AI Research in STEM and Engineering Education and the Contemporary AI Tool Ecosystem. Education Sciences, 16(11), 1834.