The promise and the pitfall
Large language models can now evaluate student writing with surprising accuracy. But accuracy on a benchmark does not mean reliability in a classroom. Even the best models sometimes invent feedback. The model tells a student their answer is wrong, with full confidence, when the answer is correct.
For that reason, a teacher must review, edit, and approve every piece of feedback Ren produces before a student sees it.
Where models fail
We tested thousands of student responses. Three failure modes recur.
1. Context blindness
A model does not know what the class learned last week. A student can use unconventional terminology that came from the teacher's own explanation. The model penalises the language, although the student shows correct understanding.
2. Criteria drift
A model applies its own implicit standards unless you ground it in the teacher's marking criteria. This causes the most trouble in subjects such as English Literature. There, the marking criteria differ substantially between exam boards.
3. Confidence without calibration
A model rarely says "I am not sure". It assigns a mark in the same tone whether it is certain or whether it guesses. This false confidence misleads a teacher who has no time to check every response.
Our approach
Ren treats the AI as a first-pass assistant, not a final authority. In practice, this means three things:
- Ren presents every piece of feedback as a draft for teacher review.
- The teacher can edit, approve, or reject each feedback point.
- We build for explainability. For a structured question, Ren shows which part of the answer scheme it applied.
The goal is not to replace teacher judgement. The goal is to give the teacher a head start, so they can spend their time on the feedback that matters most.
"But does review not add more work?"
Teachers ask a fair question. If a teacher must review every piece of AI feedback, does the AI save time, or does it only move the bottleneck?
The software industry asks the same question. Anthropic's research on AI productivity gains found large time savings on tasks such as code, documentation, and data manipulation. The study also states a limit. It does not fully account for the time a human spends to review, edit, and validate the AI output.
In software engineering, the AI did not remove the need for human judgement. It moved where that judgement applies. Developers write less boilerplate and spend more time on code review, architecture decisions, and edge cases. The role changes from creator to curator.
The same shift applies to teaching
A teacher who marks 30 scripts from scratch spends most of that time on mechanical work. They read the script, they score it against the criteria, and they write similar feedback for the tenth time. The high-value work fits into whatever time is left. That work identifies misconceptions, personalises guidance, and plans the next intervention.
The AI reverses this ratio. Ren produces the first pass. The teacher's time moves to review and refinement, which is the work that changes student outcomes.
What this looks like in practice
We measured this with a partner school. One teacher split 30 submissions into two groups of 15.
- With Ren: the AI marked in 9m 32s. The teacher then reviewed at about 2m 52s per script. Total: 52m 32s.
- Without Ren: the teacher marked from scratch at about 10m 23s per script. Total: 155m 45s.
That is a reduction of more than 65%. The teacher also spent that time on review and refinement instead of mechanical scoring.
The parallel to software is clear. A developer with AI writes less code and reviews more. A teacher with Ren marks less from scratch and refines more feedback. Total hours fall, and the share of time on high-value work rises.
What comes next
We continue to invest in three areas: better calibration, marking-criteria grounding, and explainability features that show a teacher why the model gave a mark. We develop each of these with teachers, so the AI supports human expertise instead of replacing it.
Do you want to see how a teacher reviews feedback in Ren? Get in touch to book a demo.