Teachers are being handed AI faster than anyone can train them, or govern it.

Teacher AI adoption is routinely predicted with a model built in 1989 for static, deterministic tools. Does it still explain adoption of a technology that is opaque, adaptive, and capable of unsettling professional identity?

A PRISMA systematic review of 38 empirical studies covering 15,919 teachers in 19 countries, with a random-effects meta-analysis of the 24 studies reporting extractable correlations.

The core associations hold and are moderate to strong, but on their own they say little about what will happen in a particular setting. Heterogeneity is very high (I² = 97.9–98.6%) and prediction intervals run from negative to near-unity.

An adoption survey built on usefulness and ease of use will not, by itself, tell an institution whether its teachers will use a tool. The synthesis points to trust, AI literacy, professional identity and ethical risk as the conditions that shape that decision.

Dozens of frameworks now claim to define AI literacy for educators. Do they agree on what they are describing, and does any of them tell you how to measure it?

An umbrella review, meaning a review of reviews. 28 prior studies (26 systematic or scoping reviews, 2 design and resource-mapping studies) published 2020 to 2025, drawn from six databases and screened under PRISMA 2020.

The field is fragmented. Constructivism is the dominant theoretical base, with TPACK and AI4K12 close behind, but the treatment of ethics varies enormously between frameworks. Explainability, which is what decides whether a teacher can trust an AI judgment in front of a class, is explicitly addressed in exactly one of the 28.

Adopting a framework off the shelf means inheriting whatever it happened to leave out. There is still no standardised, teacher-specific way to assess AI literacy, so most programmes cannot demonstrate that they worked.

Generative AI is arriving in teacher professional development faster than anyone can say what it is good for. What does it actually afford, and what would a defensible programme contain?

A critical review of 20 peer-reviewed studies published 2023–2025, analysed for framework development, implementation strategy and pedagogical affordance.

Four affordance categories emerge: functional, cognitive, social, reflective. But existing frameworks inherit an assumption that the technology is under the teacher's control, and generative AI violates that assumption through probabilistic output and confident error.

It produces the GenAI-TPD framework: five competency domains across four developmental stages, written as testable propositions and observable indicators. Something to build and assess against, not another reading list.

Two different things are now called AI governance: external regulation that audits a system after the fact, and ethical constraints built into how a model behaves while it runs. In education, which one is doing the work?

A comparative analysis of the EU AI Act and national governance guidance against design-embedded ethical frameworks, read through a temporal lens: when in a system's life does the constraint actually bite?

Regulation operationalises explainability retrospectively, through documentation, auditability and conformity assessment. Design-embedded frameworks operationalise it at runtime, as a property of the interaction. Neither substitutes for the other.

The Act treats educational AI used for admissions, evaluating learning outcomes, placement, or monitoring students during exams as high-risk. Transparency and human oversight stop being aspirations and become legal obligations.

India runs some of the world's largest AI-in-education programmes. How do its policies actually conceptualise AI literacy, and how far do they align with international benchmarks?

Structured document analysis of national policy, strategic frameworks and CBSE curricular materials, scored on a 1–5 rubric across six dimensions against UNESCO and OECD frameworks.

Strong on curriculum integration and student competencies. Weak on governance structures, teacher professional development and data protection. Those are the three dimensions that decide whether curriculum ambition survives contact with a classroom.

The gap is not one of intent. Policy specifies what students should learn without specifying who governs the systems, who trains the teachers, or who holds the data. Any large system can check itself against the same six dimensions.

Teacher AI training is normally evaluated by measuring attitudes afterwards. If a cohort already believes AI is useful before you start, what is the training for, and how would you find the real gap?

Pooled data from two hands-on ChatGPT workshops at two different institutions in and around Bengaluru (n = 122 before, n = 119 after), profiled across five readiness dimensions.

Most dimensions sit near the ceiling before any training happens. The two exceptions, perceived ease of use and AI self-efficacy, are precisely the two that then shift (d = 0.43 and 0.33). The change lands on concrete task affordance, not abstract value.

It produces a five-character readiness code and the AIT3 diagnostic, which a trainer can run on a cohort beforehand, so a programme aims at the dimension that is genuinely low rather than the one easiest to measure.