"The pupils really liked the experience and found it much easier than writing on paper. It also felt very accessible for pupils with SEND - they seemed more at ease and confident with the format than with a paper test.”“The feedback was much more specific, detailed and in greater volume than a classroom teacher would generally identify from this kind of assessment – e.g. the AI picked up really effectively on phrases the students were using incorrectly."

Brittleness
‍We have conducted extensive testing and piloting using a wide range of assessment types, subjects and year groups. This ensures that our approach does not just work within limited, tested parameters, reducing brittleness.

‍Uncertainty
‍We are currently implementing an approach which involves the AI assigning a confidence score (1–10) to each mark, where 10 is full confidence and 0 is no confidence. Questions with a confidence score of 7 or below will be flagged as “priority for teacher review” in the product going forward.

‍Hallucinations
‍We are experiencing minimal levels of hallucination, partly because we automark using tightly defined marking guidance, which leaves less space for hallucination. That said, we still check for hallucinations in a number of ways. Teachers moderate automarking, allowing them to spot and correct hallucinations. A sample of teacher adjustments are then checked by us to scrutinise for hallucinations, amongst other things.Hallucinations are more likely when the AI is less certain of its answer, so by taking the “confidence score” approach described above we can further mitigate the risk of hallucinations appearing, and make it more likely that teachers will correct them when they do occur.We do periodic random sampling of all marks in our assessments and we get an expert to automark those samples and flag any hallucinations. In the most recent 2,000 expert-marked questions we have discovered no hallucinations.   
‍
‍Embedded Bias
‍Our approach to automarking involves no passing of personally identifiable information to our marking engine, so no bias can be derived from knowledge of a student’s characteristics. Moreover, we use a prescriptive mark scheme that the AI applies directly, rather than allowing the model to generate open-ended interpretations, which could in theory be subject to bias of some form. 

False Positives
We evaluated false positives using a confusion matrix. While present, they were uncommon. Interestingly, we observed more false negatives than false positives.

We’re currently trialing using AI to mark scanned in paper assessments.

This is because no personally identifiable information is passed to the AI system, and no automated decision-making occurs.

However, if participating schools would like to notify parents, we provide wording that can be used in parent-school communications.