
Grammar Feedback Without Grammar Penalties
Should grammar mistakes cost students points? What research says, how to give grammar feedback that helps, and when conventions should count.
· 12 mins
Read the article →How to use AI marking for GCSE English mocks with AQA and Edexcel mark schemes, what Ofqual and JCQ allow, and where teacher judgement must lead.
It's mock season. You've got 120 Year 11 Language Paper 1 scripts, and every one ends with a 40-mark piece of writing. Each takes ten minutes to mark well, and that's before you write any feedback. Your department wants marks in by Friday.
AI marking tools promise to help, and plenty of English teachers are curious. But you're probably also asking the right questions. Is it allowed? Can it really mark against a levels-based mark scheme? And will the marks be good enough to talk to students and parents about?
This guide answers those questions with what the rules and the research actually say. Then it walks through how the AQA and Edexcel mark schemes work, and a safe way to use AI marking for mocks and classwork.

The short answer: yes, you can use AI to help mark GCSE English mocks and classwork, as long as a teacher reads the work and decides the mark. Ofqual says AI can't be the sole marker in regulated qualifications, and JCQ says the same for non-exam assessment. But Ofqual now openly recognises a role for AI in lower-stakes, classroom marking. Used well, AI can draft a level, point to the evidence for it, and handle the first pass of feedback. Your job is to check, adjust and own the mark.
The rules are clearer than many teachers think.
GCSE English Language and Literature are mostly assessed by exam, so for most English teachers the practical question is about mocks, homework and classwork. That's exactly the lower-stakes space where Ofqual sees a role for AI, with a teacher in charge.
The exam boards themselves are cautious. AQA has researched AI as a way to check its examiners' marking, not to replace them. In an October 2025 research summary, it describes "investigating how AI can support expert judgement" while keeping "human oversight at the core."
The evidence has moved fast, and it's mixed. Here's where it stands.
First, English is hard to mark even for experts. Ofqual studied how often the grade a student receives matches the grade a senior examiner would give. For maths, it's around 96% of the time. For English language and literature, it's around 55 to 60%. As Ofqual puts it, for essays "there are often other plausible marks that are equally correct." That's not a failing of English teachers. It's what happens when you judge extended writing against levels.
AI has caught up quickly on some writing tasks. No More Marking, which runs large comparative-judgement writing assessments in UK schools, tested AI against human judges on about 70,000 scripts from Years 7 to 9. The AI agreed with 83% of human decisions, "similar to the typical human-human agreement." Just two years earlier, the same team had found GPT-3's results "not great". Even now, they recommend humans check a share of the judgements.
But it isn't reliable everywhere. A 2026 Durham University study of AI marking physics essays found the AI's marks barely tracked the quality of the work, and that "adding a mark scheme does not improve discrimination". And in a 2024 study comparing feedback on students' essays, teachers' feedback was better than ChatGPT's in most areas, though the gaps were often small.
Ofqual's January 2026 principles for AI in marking sum up where things stand: "Current evidence does not support an overall or general case for the validity of AI marking." And: "agreement with human marks alone is insufficient."
What that means for you: AI can give a useful first judgement on GCSE writing, especially when it shows its reasoning. But you need to be able to see why it chose a level, check that against the student's work, and change it. A number without a reason isn't good enough for English.
Before you use any tool, it helps to be clear about what you're asking it to do. Here's how the two biggest boards structure GCSE English.
Two papers, each 1 hour 45 minutes, 80 marks and 50% of the GCSE (AQA specification).

The writing marks map onto two assessment objectives. AO5 is about communicating clearly, effectively and imaginatively, and organising ideas. AO6 is about vocabulary, sentence structures, spelling and punctuation. AQA's specification says AO6 "must constitute 20% of the marks."

The papers are weighted differently from AQA's (Edexcel specification):

Some schools teach Edexcel's newer English Language 2.0 (1EN2) instead, with two 80-mark papers at 50% each. Its writing sections use the same 24 + 16 split. Students can't be entered for both.
In GCSE English Literature, spelling, punctuation and grammar (AO4) are worth 8 of the 160 marks, and only on certain questions:
That matters for AI marking. A good tool should keep these marks separate, and never let spelling slips quietly pull down a student's mark for analysis.
Most GCSE English questions are marked in levels, not by ticking points. Examiners decide which level an answer fits, then the mark within that level. AQA's Language mark scheme tells examiners to "start at the lowest level of the mark scheme and use it as a ladder," and adds: "Once you have assigned a level you need to decide on the mark."
Two more lines from AQA's mark schemes are worth pinning above any marking desk:
This is exactly why GCSE English is hard to mark, for humans and AI alike. A level isn't a count of features. It's a judgement about quality, and it needs reasons.
AI is useful for some parts of GCSE English marking and risky for others. Here's a practical split.

Here's the opening of a made-up Year 11 mock response to a descriptive writing task: Describe a busy market, as suggested by this picture.
The market woke before the town did. Canvas roofs snapped in the wind like sails, and the air smelt of oranges, diesel and something frying. A man with a voice like gravel sold fish he swore were caught that morning. Nobody beleived him, everybody bought them anyway. Further down, under a striped awning, an old woman arranged her lemons into perfect pyramids, as if the whole day depended on it.
Say an AI tool suggests the following for the whole piece. This isn't an official mark, just an illustration of the kind of reasoning you should expect to see:
Now your job starts. You read the full response with the level ladder beside you. Maybe you think the ending pulls the piece together better than the tool noticed, and you move content up to 18. Maybe you agree on technical accuracy. Either way, you can see exactly what the tool based its suggestion on, so you're checking reasons rather than starting from a blank page, and the final mark is yours.
That's the test for any AI marking tool: can you see why it chose a level, and can you change it in one step? If not, it isn't saving you anything, because you'll have to re-mark the script anyway. In DfE research on AI marking, one teacher put it simply: "I think I would end up going back over and reading their essays anyway to see if the AI was correct" (DfE user research, 2024). Showing the evidence is what makes that checking quick.
Here's a way to use AI marking for a set of GCSE English mocks that keeps you in charge at every step.

GradingPal is an AI marking tool built around a simple rule: it suggests, you decide. Here's how it handles a set of GCSE English mocks.
A few things GradingPal doesn't do. It isn't affiliated with or endorsed by AQA or Pearson, and it doesn't contain or reproduce their mark schemes. It doesn't predict GCSE grades. And it's never the sole marker: nothing reaches a student until you've reviewed it. Student work is never used to train AI models. For data protection details, see privacy and security and our guide to UK GDPR for schools using AI tools.
More on how GradingPal works with exam board papers is on the exam boards page. You can try it on the free plan, which includes 50 submissions a month.
AI can help with GCSE English marking, but it can't own it. The rules say a teacher decides. The research says AI is getting better at judging writing but still needs checking. And English is hard enough to mark that reasons matter more than numbers.
So use AI for the first pass. Set up each question with its own levels, keep accuracy separate from content, and insist on seeing the evidence behind every suggested level. Then read, adjust and decide. The marking stays yours.
Want to try it on your next set of mocks? Start free with GradingPal and mark one question with your department's levels.
AI can suggest marks for GCSE English mocks and classwork, and recent UK trials show it can agree with human judges about as often as humans agree with each other on some writing tasks. But a teacher should always review the work and decide the final mark. English is hard to mark even for experts.
Ofqual says AI may not be the sole marker in any regulated qualification. It also recognises a role for AI in supporting lower-stakes and classroom-based marking, with humans in charge.
Only if the school allows it, and only with a human in charge. JCQ guidance says an AI tool "cannot be the sole marker," and the assessor must review all the work and decide the mark.
It's worth 40 marks: 24 for content and organisation, and 16 for technical accuracy. Examiners use a levels-based mark scheme, first deciding the level and then the mark within it.
AQA has two equally weighted papers of 80 marks. Edexcel's 1EN0 has a 64-mark Paper 1 (40%) and a 96-mark Paper 2 (60%). Both split each 40-mark writing task into 24 marks for content and 16 for technical accuracy.
For both AQA and Edexcel, 8 of the 160 marks. On AQA, that's 4 on the Shakespeare question and 4 on the modern text question. On Edexcel, it's on the post-1914 text question.
No tool can do this reliably, because grade boundaries change every exam series. Use mock marks to talk about levels and next steps, not predicted grades.
One email when we publish something new on grading, feedback and standards. No newsletter filler.
See our privacy policy.
Reading about it only goes so far. Bring a class set you have graded and compare.
Free plan, no credit card. 50 submissions a month, any assignment type.