Prototype: This dashboard is in active development. Data and features will be refined through workshops with end users.

What Can AI Tell Us About Education Policy-and What Shapes Its Answers?

AI is shaping real decisions in education systems - including the issues that get attention. But the choices shaping AI outputs aren't always visible. Builders can't easily show them, decision makers can't assess them, and those affected can't question them.

AtlasED analyses thousands of education policy documents to surface what's being debated, who's shaping the conversation, and how these patterns change over time. But it also does something most AI tools don't: it shows you what changes when the choices behind the model change.

Why This Matters A policymaker asks AI what the biggest issue in education is. The model says 'learning outcomes,' so funding follows. Change one model setting, and that single category splits into reading scores, maths attainment, progress rates, and early years development - each needing different action. The result wasn't wrong. It depended on how the question was set up.

The Core Model

The model used here is deliberately simple - because every choice is visible. The number of topics, the data sources, the text processing: change any one and watch the findings shift. In more complex AI systems, those same choices still exist but are harder to see - making it even more important to test whether results would change if the system were built differently.

4,170
Total Documents Analysed
2023 – 2026
Time Period
30
Policy Topics Identified
6
Data Sources
97%
Model Stability
See How It Works
1
Data collection
~4,000+ documents across six organisation types.
What's included shapes what's found.
2
Text processing
Text is cleaned and simplified.
These choices shape the model's language.
3
Model
We use NMF - a transparent method with visible settings.
In more complex AI, these choices still exist but are harder to see.
4
Quality checks
The model is consistent and passes standard tests.
But these don't show how results change under different choices.
What's Next → On the next page explore what the core model found.