How Teachers Read AI Likelihood Scores on EAL Student Work

AI detection for EAL students is one of the most uncomfortable parts of running a likelihood check on a class set. A Year 10 EAL student who learned English at 11 and has worked relentlessly to write in clean, simple, well-structured sentences is exactly the kind of student a probabilistic detector will flag — not because the work is AI-generated, but because the writing pattern looks more uniform than that of a native speaker still figuring out their voice.

This post is for the teacher who has just run a GradeOrbit AI detection check on a Year 9 or Year 10 piece of writing, seen a 73% likelihood score next to an EAL student's name, and has thirty minutes before next lesson to decide what to do. It walks through why EAL writing can spike a score, what a high number actually evidences, the conversation to have with the student before any accusation lands, and how to keep your school's academic integrity policy applied consistently across native and EAL pupils.

Why EAL Writing Can Spike a Likelihood Score

AI detection models work by spotting statistical patterns associated with machine-generated text — low burstiness (sentence-length variation), low perplexity (predictable word choices), and uniform syntactic structures. A confident native English speaker in Year 10 will write with varied sentence lengths, slang, half-finished idioms, and occasional grammar shortcuts. An EAL student who has spent four years drilling formal sentence structures in EAL support sessions will often produce work that is grammatically tidier and structurally more uniform than their native peers.

That uniformity is not a sign of AI use. It is a sign that the student has learned English in a structured, formal setting and has not yet developed the messy, idiosyncratic voice that native speakers absorb through years of reading novels, watching TV, and arguing in the playground. The detector cannot tell the difference. It sees uniform syntax and flags it.

This is the single most important thing to understand before you act on a score: a high likelihood number on EAL work tells you about the writing's surface features, not about the student's honesty. Treat the score as a prompt for further investigation, never as a verdict.

What a High Score Actually Evidences

GradeOrbit's AI detection tool returns a likelihood percentage from 0 to 100% and offers two model tiers — a 1-credit quick check and a 3-credit deeper analysis. The number it returns is the model's confidence that the writing pattern resembles AI-generated text. It is not a confession, not a confidence interval on guilt, and not evidence admissible at a malpractice panel.

What it does give you is a starting point. A high score on an EAL student's work tells you: this writing has the surface features the model associates with AI output. Your job as the marker is to layer your own professional judgment on top — the context you have on this specific student's writing history, their behaviour in lessons, their drafting process, and any patterns you have seen across other pieces of their work this year.

If the EAL student you are marking has handed in five pieces of writing this term, all of which scored 60-80% on AI detection, all of which were written in class under supervision, the consistent score tells you the score is reflecting the student's writing style — not their use of AI. If a single piece spikes from 30% to 95% with no other change in circumstance, that is a different conversation.

Running the Conversation Fairly

If you do decide to follow up on a high score, the conversation must start from curiosity, not accusation. EAL students are already navigating a school system in their second language; a clumsy "did you use ChatGPT" question landed in front of them in front of peers can damage trust for the rest of the year and may breach equality duties depending on how it lands.

Borrow the framing from our wider guide on talking to students about high AI detection scores — open with questions about how they planned the piece, what notes they used, what bits they found hard, and which sentences they are most proud of. A student who genuinely wrote the work will be able to walk you through their drafting decisions; a student who pasted from an AI tool usually cannot.

For EAL students specifically, add one more question: ask which sentences felt natural to write and which they struggled with. Their answers will tell you more than any percentage. A student who can point at the paragraph they spent forty-five minutes redrafting because they could not figure out how to use the word "however" is not the student who outsourced the piece.

Aligning With Your School's Academic Integrity Policy

Most UK secondary academic integrity policies were written before generative AI was widely available and now sit in an uncomfortable middle ground. Some schools require any "evidence of AI use" to trigger a malpractice referral; others leave the decision with the class teacher. Whichever policy your school has, the rule of thumb is: a likelihood score from any detector — GradeOrbit's included — should never be the sole basis for a disciplinary outcome.

This matters extra for EAL students because false-positive rates on EAL writing are well-documented in detection research, and any disciplinary process that disproportionately catches EAL pupils will eventually be challenged by parents or by the school's own equality monitoring. Schools that have updated their integrity policy in 2026 increasingly require: a likelihood score, plus a process-evidence check (drafts, notes, planning), plus a conversation with the student, before any sanction lands.

If your school's policy does not yet make that distinction, this is worth raising at the next safeguarding or teaching-and-learning meeting. The detection tool is useful; the disciplinary process needs to be calibrated to deal with what it tells you.

What to Do When You Are Still Unsure

If after the conversation, the planning check, and reviewing the student's writing history you still cannot tell whether the work is theirs, the answer is usually: mark the piece, give the feedback the student needs to improve, and set the next task in lesson under supervision. A single supervised in-class write of a comparable task will tell you in twenty minutes whether the original was genuine. It is more diagnostically powerful than re-running the detector with a different model.

If the in-class write looks dramatically weaker than the home piece, you now have evidence of a process gap — and that is a conversation to have with the student, the head of department, and (if your school requires it) the safeguarding lead. If the in-class write is consistent with the home piece, you have your answer and the AI detection score was a false positive. Move on without making a fuss.

For a broader walk-through of false-positive handling, see our guide to AI detection false positives — the EAL angle is one of several patterns worth knowing before you act on any score.

Try GradeOrbit AI Detection on Your Next Class Set

GradeOrbit's AI detection tool is designed to be the start of a teacher's investigation, never the end. The 0-100% likelihood score is a prompt for professional judgment, not a verdict. The 1-credit and 3-credit model tiers let you choose between a quick sweep across a class set and a deeper analysis on the pieces that warrant it.

The product is built around UK teachers making fair, consistent decisions about real students — including the EAL student whose tidy, careful sentences would otherwise be misread by a less context-aware tool. Head to the homepage to sign up and run your first detection check today.