Evidence & accuracy

The useful comparison isn't whether AI always matches one examiner exactly - qualified examiners don't always match each other exactly either. That's why we benchmark Marking.ai against independently marked student work and compare how closely it agrees with qualified examiners against how closely those examiners agree with each other. Across our published accuracy studies, we report performance by subject and question type rather than relying on a single headline percentage. We then add further safeguards through specialised marking strategies, evidence behind marks, confidence indicators, and teacher review before results are shared.

Because different marking tasks behave differently. Agreement on a short, points-based question is not directly comparable with agreement on an extended response where professional judgement plays a larger role. We therefore report accuracy by subject, assessment and question type, state the measurement being used, and compare Marking.ai with the relevant human examiner benchmark. As new studies are published, they are added to the evidence base rather than combined into one potentially misleading number.

General-purpose AI can be prompted to mark student work. The difference is the system built around the model. Marking.ai applies specialised marking strategies, uses the relevant marking guidelines, provides evidence and feedback, manages real class marking workflows, and gives teachers a structured review process. Where our studies include a controlled comparison with the underlying model, we publish those results too. This allows us to measure what the full Marking.ai system adds rather than simply claiming that it performs better.

Every published study is dated because both AI models and the Marking.ai marking system continue to evolve. Each study reflects the production system used at the time of measurement. As the system changes, we re-run relevant benchmarks and publish updated results rather than allowing old figures to stand indefinitely.

Marking.ai does more than send a student's answer and marking guidelines to an AI model. Depending on the assessment, the production system can: route different questions through specialised marking strategies; evaluate work against the relevant criteria; connect marks to evidence in the student's response; use confidence indicators to help focus review. We then test that system against independently marked student work and publish the results by subject and assessment type. The aim is not to claim perfect marking. It is to build a marking system whose performance can be measured transparently against the human standards schools already rely on.