Free integrity tool · Sentence-level evidence

Free AI detector for teachers & UK schools

Paste student writing, choose EAL or younger-writer context when it fits, and review highlighted sentences with plain-English guidance — built for fair conversations, not opaque accusations.

EAL & younger-writer context Sentence-level highlights Assistive signal, not proof

1. Paste student writing

0 words · 0 characters Min 25 words · fair-use caps apply
Example output (crawlable, no JavaScript)

Sample report: GCSE History mixed-AI essay

62%

Some AI-like patterns · Medium confidence

Overall AI-like score is 62%. There are mixed signals — review the highlighted sentences and discuss with the student if needed.

38%
Flagged coverage
68%
Model score
54%
Statistics score

Why this score

  • Burstiness / rhythm: Score 71% (weight 18%)
  • Discourse markers: Score 64% (weight 14%)
  • First-person classroom detail: Opening refers to a named teacher and a specific lesson task.

Human counter-signals

  • First-person classroom detail: Opening refers to a named teacher and a specific lesson task.

Highlighted submission

In Miss Carter's Year 11 History lesson last Tuesday we compared two cartoons about the Treaty of Versailles. I thought the German one was angrier because the artist drew the Allies as giants crushing a tiny figure. My homework draft starts here. Furthermore, it is important to note that the Treaty of Versailles reshaped Europe. The settlement served as a testament to the desire for lasting peace after the Great War. In essence, navigating the complexities of interwar diplomacy requires a holistic approach. Moreover, the reparations clauses fostered a deeper understanding of economic grievance in Germany. Consequently, historians debate whether the treaty made a second conflict more likely. When I compared the cartoons again I noticed the British one used humour more than anger. That matches what we said in class about propaganda styles. Overall I think the homework middle section sounds less like me than the opening.

Verified quotes

  • “Furthermore, it is important to note that the Treaty of Versailles” — Formulaic AI discourse opener with even sentence rhythm.
  • “navigating the complexities of interwar diplomacy” — Stock generative phrasing uncommon in Year 11 classroom writing.

This score is an assistive signal only—not proof of misconduct. Use it alongside your judgement, the student's process, and other evidence.

Step-by-step instructions

How to check student writing for AI fairly (and why writer context matters)

A fair check starts with context, then evidence, then a conversation — never a score alone.

1

Paste enough text & set context

Aim for 120+ words. Choose EAL or Younger writer when it matches the learner so formulaic or evenly paced prose is less likely to inflate the score.

2

Read highlights, not just the %

Compare flagged coverage with the overall score. Inspect hover reasons and verified quotes. Note human counter-signals such as first-person classroom detail.

3

Hold an evidence-led conversation

Ask the student to talk through planning and drafting. Pair with Authorship Fingerprints or Citation Checking when appropriate.

UK guidance: EAL, PEEL/PETAL frames, and JCQ authenticity

EAL false-positive reduction

EAL learners often reuse taught connectives and safer vocabulary. Writer context lowers the weight of those patterns so detection stays fairer.

Taught PEEL / PETAL frames

Scaffolded essay structures can look “template-like” to statistical detectors. Treat uniform PEEL paragraphs as pedagogy first; use sentence highlights and process evidence before raising concerns.

JCQ AI-use coursework guidance

JCQ expects authentic coursework and clear AI-use policies. Use this free detector as a prompt for discussion, then record outcomes in the full Feedback Flows AI Writing Detector when your school needs an audit trail. Read our JCQ AI guidance for NEA and coursework.

Need ZIP uploads, student history, and PDF reports?

Open a free Feedback Flows account to run the full AI Writing Detector on DOCX/PDF/ZIP batches, keep student timelines, and download evidence packs — still with a human in control of every outcome.

Create free account →
How it works

How the hybrid AI detector works (model + stylometric signals)

Built for UK teachers who need inspectable evidence, not a black-box percentage.

Sentence-level model pass

The language-model component scores each sentence for generative likelihood and returns verified quotes you can show in a pastoral meeting — never a mysterious score without the text behind it.

Statistical writing signals

Alongside the model, we weigh stylometric features common in LLM prose and authentic classroom writing:

  • Burstiness / rhythm — even sentence lengths vs natural variation
  • Discourse markers — stacked “Furthermore / Moreover / In essence”
  • Essay-template rigidity — formulaic PEEL-like blocks without local detail
  • Lexical uniformity — narrow, polished vocabulary across paragraphs
  • Hedging density — generic academic hedges without subject-specific voice
  • Human counter-signals — named teachers, lesson tasks, first-person classroom detail
Interpretation guide

How to read score bands vs flagged coverage

Always read the overall band together with how much of the text is highlighted.

0–34% · Low AI signal

Writing looks consistent with human drafting for the chosen writer context. Still skim highlights if any appear; a short polished insert can hide inside a mostly authentic piece.

35–69% · Mixed / uncertain

Common on drafts with a human opening and a later AI-assisted section. Compare flagged coverage with the overall score and open the highlighted sentences before any conversation.

70–100% · Strong AI-like signal

Many sentences look generative. This is still not proof — gather draft history, ask for an oral explanation, and consider Authorship Fingerprints before any formal integrity step.

Why coverage matters: a high model score with low flagged coverage usually means one polished paragraph (or a pasted conclusion), not a wholly AI-written essay. A moderate score with high coverage across most sentences is a stronger prompt for a calm check-in. See how teachers spot ChatGPT in student essays and evidence over AI detection scores.

UK coursework playbook

GCSE, NEA, EPQ, BTEC & T Level: fair investigation steps

Practical steps that sit beside your school or college integrity policy.

1

Note authentication stage

Before a JCQ authentication declaration, centres usually handle concerns internally. After signing, suspected malpractice may need reporting to the awarding organisation — check current JCQ wording for your qualification.

2

Protect EAL & ND learners

Set EAL or Younger-writer context first. Formulaic connectives from scaffolding are pedagogy, not automatically misconduct.

3

5-minute viva / draft review

Ask the learner to explain one flagged paragraph, name sources, and show earlier drafts or notes. Record what you heard.

4

Add second signals

Pair with authorship and citation checks when style or bibliography also looks off.

For department policy language, see our school AI policy guide for UK leaders and MATs. For JCQ coursework stages in more depth, read JCQ AI guidance for NEA and coursework.

Authority guides

Further reading for UK teachers and leaders

Practical articles on AI detection, JCQ coursework rules, school policy and fair investigation.

Integrity · 11 min

JCQ AI guidance for NEA & coursework

Authentication stages, acceptable AI use, and evidence that stands up in a centre investigation.
Integrity · 10 min

How teachers spot ChatGPT in student essays

Linguistic tells, EAL fairness, and a non-accusatory check-in script for UK classrooms.
Governance · 10 min

School AI policy guide for leaders & MATs

Red/Amber/Green boundaries, DPO checks, and a fair department protocol for suspected AI misuse.
Integrity · 7 min

Evidence over AI detection scores

Why a lone percentage is weak evidence and what stronger integrity looks like.
Integrity · 12 min

Writing fingerprints & authorship checks

Compare new work with a learner’s own baseline under JCQ and UK GDPR expectations.
Integrity · 10 min

How to spot fake AI-generated references

Hallucinated DOIs, year mismatches, and bibliography checks that complement AI detection.
FAQ

Free AI detector FAQ for UK teachers

No. AI detection scores are assistive signals only — never proof of misconduct. Use the highlighted sentences, process evidence (drafts, notes, viva), and your professional judgement before any formal conversation.

Learners with English as an additional language and younger secondary writers often produce more formulaic or evenly paced prose. Selecting the matching context recalibrates the hybrid model so those patterns are less likely to raise false positives.

Flagged coverage is the share of analysed sentences highlighted as AI-like. A high overall score with low coverage often means a short stretch of polished text — review that passage with the student rather than treating the whole essay as AI-generated.

Pasted text and results are stored privately for up to 30 days so we can prevent abuse and improve the tools, then deleted. They are never published, shared, or used to train AI. We never create public shareable result links.

This free tool checks one pasted text with fair-use rate limits. The full product supports DOCX/PDF/ZIP batches, student timeline history, downloadable PDF reports, Authorship Fingerprints, and recorded review outcomes for your department.

JCQ emphasises authentic coursework and process evidence. Treat detection as one prompt for a conversation: ask how the student planned, drafted and revised, and pair it with Authorship Fingerprints or Citation Checking when appropriate.

No. Under JCQ guidance and UK GDPR Article 22 protections against solely automated decision-making, a single detector percentage is not proof of malpractice. Centres need human judgement plus process evidence before any formal sanction.

Many detectors reward unpredictable vocabulary and uneven sentence rhythm. EAL writers and neurodiverse learners who rely on taught frames often write more evenly, which can look “AI-like”. Writer context and classroom process evidence reduce unfair flags.

The free tool needs at least 25 words to run, but 120+ words (ideally a full paragraph or short essay section) produce more stable scores. Very short answers suppress stylistic variation and should not drive formal concerns.

Paraphrasers can blur some model signals, which is why we never treat a score alone as proof. Pair detection with Authorship Fingerprints, citation checks, draft history and a short viva so process evidence still carries the investigation.

Light grammar tools rarely dominate a score. Taught PEEL/PETAL scaffolds can look template-like statistically — treat pedagogy as the first explanation, use Younger-writer or EAL context where appropriate, and read highlighted sentences before escalating.

AI detection looks for generative writing patterns in one text. Authorship verification compares a submission with that learner’s earlier authentic samples. Collusion checking compares two learners for shared passages. Use the matching free tool for each question.

Keep the meeting exploratory, not accusatory. Show the highlighted quotes, ask how they planned and drafted, and invite them to explain key terms orally. Record the outcome and next steps rather than leading with a raw percentage.

Yes. Feedback Flows Ltd is ICO registered (ZC132812). Pasted free-tool text is retained privately for abuse prevention then deleted, never used to train AI, and never published. Schools remain the data controller under our DPA for account use.
Get started

Batch-check classes with evidence you can defend

Upgrade for DOCX/PDF/ZIP uploads, Authorship Fingerprints, Citation Checking, and recorded review outcomes.

Full AI Writing Detector Guide: Evidence over scores