AI grading tools for teachers: what works in 2026

"AI marking" spans everything from a chatbot with a rubric prompt to deterministic criterion checkers calibrated to one specification. Some of it is genuinely useful; some of it will hand you an undefendable grade. Here's how to tell which is which — as a department choosing, and as a teacher protecting your judgement.

What AI marking genuinely can do in 2026

• Read evidence against published criteria. Checking a submission against a fixed grid — was this criterion addressed, at the verb's depth, with the brief's data — is a reading task, and machines read well.
• Cite the evidence. The good tools show you the exact passages behind each verdict, so you confirm rather than trust.
• Produce the paperwork. Records, feedback sheets, verification logs — the administrative exhaust of assessment, generated from verdicts that already exist.
• Screen whole batches instantly for gaps, thin evidence and authenticity signals, before human attention gets spent.

What it cannot do — and any vendor claiming it is selling you a problem

• Award the grade. Grade decisions sit with the assessor and the centre's quality process. A tool that outputs "this is a Merit" as its product has mispositioned itself into malpractice territory.
• Replace authentication. Tools can flag AI-writing indicators; only you can authenticate the learner's work, through process evidence and conversation. See the detection guide.
• Know your learner. The "this is below her usual standard" judgement that changes how you read evidence — that's a teacher input no model owns.

The eight questions to ask any vendor

1. Which specification, which issue? "Aligned to BTEC" is not an answer. "L3 Business Issue 4" is. Where do their criteria come from, and how do they track spec updates?
2. Show me the evidence behind a verdict. If the product can't cite the passage, it's asking you to trust a black box with your assessment records.
3. Is it deterministic? Same evidence, same brief, same config — same verdict? Non-deterministic tools make re-checks and appeals unwinnable. (checkb.tech replays identical runs for exactly this reason.)
4. Who sees the verdict first? Advisory-to-teacher tools and grade-emitting-to-learner tools are different products with different compliance implications.
5. Where does learner data go, and is it used for training? GDPR, data residency, retention, and written confirmation that submissions are never used to train models. Get it in writing.
6. How does it handle brief versions? V1/V2 contexts differ; criteria don't. A tool that mixes them up will read evidence against the wrong scenario's data (PSA guide).
7. What does the output look like to an IV? Ask for a sample export. If it's a score, it's decoration; if it's criterion → verdict → cited evidence, it's paperwork.
8. What happens when it's confidently wrong? Every tool misreads sometimes. The answer you want: you overrule in one click, and the record shows your decision.

Red flags

• Marketing that shows grades being emitted to students.
• "97% accurate" with no definition of what "accurate" means, against whose judgements.
• No data protection page, or one that won't rule out training on learner submissions.
• A rubric pasted into a chatbot, repackaged as an assessment platform — fine for your own drafting, not for assessment records.

Where checkb.tech sits

Deliberately narrow: one specification family (Pearson BTEC International L2/L3 Business — with IT and Sport units live too), criteria verbatim from the spec, advisory per-criterion verdicts with cited evidence, deterministic runs, no grade-emission, and a UK/EU GDPR posture with no training on student data (details). If you assess other boards, expect a general tool instead — but for BTEC evidence, purpose-built beats general-purpose. Free tier to test on a real script.

FAQ

Are AI grading tools allowed by JCQ?

Assistive use with teacher accountability is the compliant pattern; tools that automate the decision are not. The teacher must review tool output — JCQ's guidance protects the integrity of decisions, and advisory tools that keep the decision with the assessor sit on the right side of it.

Will using a tool undermine my professional judgement?

The opposite, for a well-designed one: it removes the mechanical reading so your judgement gets more of the attention — see where marking time goes.