AI Marking for A Level Politics: 0.84 Correlation on the Edexcel 30-Mark (No Source) Question

Richard Davis

29 June 2026 · 7 min read

"How reliable is your A-Level Politics AI marking system?" We hear this question constantly. So we put our tool to the test on the Edexcel A Level Politics 30-mark "no source" essay — the longer, source-free evaluative questions — using 51 of the exam board's own standardisation exemplars. Here is exactly how accurate it was, with every script's marks laid out in full.

What we tested

This study focuses on Edexcel A Level Politics — specifically the 30-mark "no source" questions. These are the extended, evaluative essay questions (the "evaluate" and "to what extent" essays), where students build and sustain an argument without a source extract to lean on. They are demanding to write and demanding to mark.

Edexcel publishes a large bank of exemplar essays for these papers, and we used 51 of those exam-board-approved standardisation materials. These are the official scripts used to show teachers the full spectrum of answer quality, from low-mark responses through to near-perfect ones — the marks in our sample run all the way from 8 to 30.

We ran all 51 essays through our dedicated Edexcel Politics marking tool, then measured the correlation between the official marks the board awarded and the marks Top Marks AI assigned to those same essays. The AI was not shown the official marks or any annotations.

The headline measure is the Pearson correlation coefficient. In short:

• A value of 1 means perfect correlation — when one marker scores high, so does the other, and likewise for low scores.
• A value of 0 means no correlation — one marker's score tells you nothing about the other's.
• Negative values mean the markers systematically disagree.

For context, how do humans perform?

What sort of correlation do experienced human markers achieve when marking scripts already marked by a lead examiner? A landmark study by Fowles (2009) set out to measure exactly that, comparing the marks of experienced GCSE examiners against the chief examiner's marks on the same scripts.

The findings are sobering. Those experienced examiners correlated with the chief examiner at only around 0.65 — a positive correlation, but far from perfect — and just ~45% of their marks fell within the exam board's tolerance of the definitive mark. Marking, it turns out, is far less consistent than most people assume, even among seasoned professionals.

How did Top Marks AI perform?

Across all 51 essays, Top Marks recorded a correlation of 0.84 (0.835 to three decimal places) — a very strong positive correlation that comfortably outperforms the ~0.65 achieved by experienced human markers (Fowles, 2009), and on a harder, source-free question type.

That Mean Absolute Error of 2.55 marks means that, on average, the AI differed from the board by just 2.55 marks — an average difference of about 8.5% on the 30-mark scale. The within-tolerance figure tells the same story: Top Marks AI landed 68.6% of its marks within 3 marks of the board, against the ~45% that experienced examiners managed in Fowles (2009). On the measure exam boards care about most — how often a marker falls inside tolerance — our tool was markedly more consistent than the human benchmark.

We don't claim Top Marks is infallible — so when it does get things wrong, how bad is it? For that we turn to the Root Mean Square Error (RMSE), which highlights the severity of large errors by squaring them: square 2 and you only reach 4, but square 5 and you jump all the way to 25. Top Marks AI's RMSE was 3.23, meaning that even when larger errors occur, they stay remarkably small relative to the 30-mark scale.

You can see the full side-by-side human and AI scores for all 51 essays below. A negative difference means we marked below the board; a positive difference means we marked above it.

Can I see this as a graph?

Absolutely. First, here's a scatter graph showing what a theoretical perfect correlation of 1 would look like — every essay sitting exactly on the diagonal:

Now the real-life graph, drawn from the data above:

On the horizontal axis is the mark given by the exam board; on the vertical, the mark given by Top Marks AI. Each dot is one essay, and you can see how closely the cloud of points hugs the line of perfect correlation.

Richard Davis

Founder & CEO, Top Marks AI

Richard read English at UCL and Cambridge before founding Accolade Press, a boutique academic publisher. A lifelong educator and the author of four bestselling thriller novels, he founded Top Marks AI to bring rigorous, exam-board-calibrated marking to every school in the UK.