The numbers first
Our research corpus now holds 13,150 papers, 11,721 of them fully analyzed. The range covers 1984 through September 2026, collected from 1,125 arXiv preprints and 3,289 journal publications. All 160 taxonomy cells (20 topics, 8 research aspects each) are filled.
For context: earlier posts occasionally cited 9,653 or 11,640 papers. That was a counting bug in an earlier pipeline version, not a different corpus.
The interesting part is not the volume but the distribution: which topics the field works on, and which it skips.
What the research talks about: feelings before measurement
The most published topic is attitudes, trust and acceptance. Not competency models, and not measurement.
At the bottom sit Assessment & Measurement (152 papers) and Program Evaluation (177). Roughly speaking, the field spends more than twice the effort on how people feel about AI than on whether training programs work at all. Anyone planning training should know this: sentiment is well covered, effectiveness is nearly untouched.
An extremely young field
Almost all of the evidence is recent. 77 % of all dated papers are from 2025 or 2026.
Until 2021, "AI literacy" was a niche topic with single-digit annual output. In practice this means references age quickly and meta-analyses are almost entirely missing. Whoever builds topical authority now can keep it for a long time.
The evaluation gap: 2.2 %
We assign every paper one of eight research aspects. The distribution captures what may be the single most important finding of the corpus:
Only 2.2 % of papers evaluate anything. The thinnest cells in the taxonomy are almost all evaluative: Organizational Implementation × Evaluation, 1 paper. Lifelong Learning × Evaluation, 1 paper. SME Training × Evaluation, 2. Even Compliance × Evaluation, 2.
So anyone launching an AI-literacy program today is deciding largely without effectiveness evidence. Robust studies on whether and how AI training changes workplace behavior simply don't exist. Our evaluation framework targets exactly this gap.
Where the field is heading: agentic AI and the AI Act
In the 12-month window (1,477 papers since August 2025), a few keywords burst well above their long-term levels:
| Keyword | Papers (12 mo) | Burst factor |
|---|---|---|
| Agentic | 48 | 2.5× |
| Article 4 (AI Act) | 6 | 2.0× |
| Productivity | 57 | 1.7× |
| Readiness | 48 | 1.7× |
| EU AI Act | 46 | 1.7× |
| Compliance | 46 | 1.7× |
| Trust | 98 | 1.5× |
The category level tells the same story. Compliance & AI Act tripled within the 12-month window (214 → 652 papers, +205 %), the fastest of any major area, followed by Assessment (+155 %). The field's question is shifting from "What is AI literacy?" to "How do we get compliant, and what does it pay?"
Companies need this discussion anyway. Art. 4 of the EU AI Act has applied since August 2024 and requires role-appropriate AI competence. That the research is turning there now is no coincidence.
What we're building from it
Three guides are in the works:
- Implementation playbook, for organizations setting up AI-literacy programs (development, at 4 %, is a white space of its own)
- Evaluation framework measuring at Kirkpatrick levels 3–4 (behavior change and results), for the 3 % white space
- Assessment instrument spec, task-oriented and occupation-aligned across levels 1 to 4, instead of pure self-report scales
Keep reading
Quarterly re-scans keep the corpus current. The next trends post comes with a fresh 12-month window. If you want to get practical: the 16 lessons of the course are built on exactly these evidence blocks.