Explaining teacher adoption of AI through the Technology Acceptance Model

A model from 1989, pointed at a technology from 2023

When researchers want to predict whether teachers will use an AI tool, they almost always reach for the Technology Acceptance Model: people adopt what they find useful and easy to use.

That model was built for spreadsheets and email, technologies that are stable, legible, and do what they are told. Generative AI is none of those things. It is opaque, it changes underneath you, and it can do a version of the job the teacher thought was theirs.

So the question is whether the old model still earns its keep.

What went into the pool

A PRISMA systematic review identified 38 empirical studies of teacher AI adoption: 15,919 teachers across 19 countries.

Twenty-four of those reported correlations that could be extracted and pooled (N = 11,768). Those were transformed to Fisher-z, combined under a random-effects model, and converted back for reporting.

The other fourteen studies were not discarded. They carry the qualitative synthesis that turns out to matter more than the averages.

The averages look reassuring

All three core relationships are positive and moderate to strong:

Ease of use ↔ Usefulness   r = .53
Ease of use ↔ Intention   r = .54
Usefulness ↔ Intention   r = .62

Usefulness is the strongest link to intention. Read this far and the conclusion writes itself: convince teachers a tool is useful and they will adopt it.

Then look at the spread

The narrow bar is the confidence interval, showing how precisely the average is known. The wide bar is the prediction interval: where the association would plausibly land in the next school you walk into.

For every one of the three, that interval starts below zero and runs almost to one. The average is well estimated. The next case could fall almost anywhere along that range.

What the variation is actually made of

The qualitative synthesis of all 38 studies locates five conditions behind the spread:

1. AI literacy and trust. A usefulness judgment needs a working model of what the system can do.
2. Identity, not just function. Acceptance shifts from "does it work" to "what does it make me".
3. Institutional context and social norms. Mandates and peer practice.
4. Tool-specific logics. A chatbot and a grading system are not adopted the same way.
5. Ethical risk. A values-based gate that can override a positive usefulness judgment.

What to do with this

If you are deciding whether teachers in your institution will take up an AI tool, the usefulness-and-ease survey will give you a confident average that may not transfer to your setting.

The questions that carry the variance are different ones: do these teachers have a working model of what the system does? Does it threaten or extend how they see their job? What does the institution actually reward? And where is their ethical line?

The model is a compass, not a map. It gives you the direction; it does not give you the terrain.

PS