For teachers, lecturers and schools
See which parts of a submission read as AI and why, with an honest “Uncertain” when it cannot tell. It runs in your browser, so student work is never uploaded, stored or used for training.
500 words free · no account · student work never leaves this browser
Paste a submission or drop a .docx or .pdf. It is analysed in this browser and never uploaded.
Open the checker: 500 words free, nothing uploadedExample report
Illustrative example · invented text
We wrote this in a student’s voice and pasted in one chatbot-style paragraph. Below is Ghost Act’s real, unedited output: the overall score sits in the “Uncertain” band, and the mixed-authorship check splits the essay where the writing changes.
Calibrated ensemble
48% AI likelihood for the whole text. 10 passages marked, each with the reason.
What to take from itOn its own the score is no basis for action, and Ghost Act says so. But the report shows where to look: every flag sits in the middle paragraph, one of them is harmless, and the voice changes. That is a conversation to have, not a verdict: ask the student to talk you through Article 48.
“What exactly did the software find?”
A parent has emailed. Their daughter’s coursework was flagged for AI, and they want to know which tool you used, what it found, and where her essay was sent. If this goes further, an appeal panel will ask the same questions.
A percentage will not answer them. JCQ says detection must be part of a holistic approach and never the sole marker. The student ombudsman has upheld complaints about AI findings and said decision-makers must understand a detector’s strengths and limitations.
With Ghost Act you can reply with the passages that read differently and the reason for each, say plainly whether the result was “Uncertain”, explain that you have looked at her drafts and talked to her, and confirm that her essay was never uploaded anywhere.
The problem
Misconduct cases involving AI have tripled, ombudsmen are overturning findings that leaned on a detector, and schools are now told student work must not train AI without permission. The tool you use has to stand up to all three.
proven cases of AI cheating at UK universities in 2023–24, or 5.1 per 1,000 students, up from 1.6 the year before and described by experts as the tip of the iceberg.
Guardian FOI, via E&T (2025)published AI-misconduct complaints were upheld or partly upheld by the UK student ombudsman; three involved second-language students.
Times Higher Ed (2025)of human writing with permitted, light AI editing was flagged as AI by commercial detectors, against 9–15% of unedited human abstracts.
Karr et al., arXiv (2026)is the University of Southampton’s last year of Turnitin, after terms that would let student submissions be used to train AI.
Times Higher Ed (Aug 2026)How Ghost Act helps
“Where does the essay go?”
The DfE says schools must not let students’ work train AI without permission. Ghost Act’s language model runs in your browser: submissions are never uploaded, stored on a server or used for training, and saved reports stay in that browser.
“Use more than one tool”
JCQ notes that a second detection tool can add evidence. Ghost Act works differently from your institutional detector: 46 calibrated signals plus a language model, tested on 24 AI models, including current releases.
“What exactly did it find?”
Each flagged passage says why: LLM vocabulary, stock phrases, stacked transitions, uniform rhythm, repeated openers. You can put the text in front of a student or a panel, not just a percentage.
“Some AI use was allowed”
The mixed-authorship check shows where the writing changes within one submission, so a pasted paragraph stands out from the student’s own voice, and permitted help on one section is not mistaken for a generated essay.
“How often is it wrong?”
81% of AI text caught at 3% false flags, 82% on models it never trained on. By genre: 0% of school essays and 7% of university essays wrongly flagged; 1% of essays by English-language learners against 2% by native writers.
“We have no budget for a site licence”
Pro is £100 a month for 500,000 words, with batch reports for a whole class or cohort. No site licence and no per-seat pricing. Individual teachers can buy word packs from £4.99 for 5,000 words and use them whenever they need.
Compared
Cloud detectors differ, and several publish strong accuracy figures on their own tests. We do not compare accuracy; we compare what you can check and explain.
| Feature | Ghost Act | Typical cloud detector |
|---|---|---|
| Where student work goes | Stays in the teacher’s browser: never uploaded, stored on a server or used for training. | Uploaded for scoring; often kept in a repository, and training use depends on the contract. |
| Why a passage was flagged | A named reason on every flagged passage. | Usually a score and highlighted sentences. |
| When it cannot tell | An explicit “Uncertain” band (29%–78%). | Usually a percentage either way. |
| Published error rates | By genre, for English-language learners, and on AI models it never saw in training. | Usually one headline false-positive rate on the vendor’s own tests. |
| Watermarks and EU AI Act Art. 50 | Tests open watermark schemes with your key and explains when “no watermark found” means nothing. | Rarely covered. Most tools neither test open watermark schemes nor explain what “no watermark found” means. |
| Price | £100 a month per department, or packs from £4.99. | Site licence or per-seat contract, usually annual. |
Word, PDF or plain text, one essay or a whole set on Pro. It is analysed in your browser and never uploaded.
“Uncertain” means talk, not act. Look at which passages were flagged, why, and whether the voice changes part-way through.
Print the report and read it beside drafts, in-class work and what the student can explain. Never decide on the score alone.
Every figure on this page comes from Ghost Act’s published validation. How we measure
Pro for departments
£100/month500,000 words a month: enough for about 250 two-thousand-word essays. Batch reports for whole cohorts, printable reports for the file, and an MCP server to run checks from your own tools. Checks sent through the MCP server run on Ghost Act’s server in Western Europe instead of your browser: analysed in memory, never stored or used for training.
Questions
Look for a change in voice from the student’s other work, vocabulary they do not normally use, and an argument they cannot explain when asked. A detector adds a second view: Ghost Act highlights the passages that read as AI, gives the reason for each and shows where the writing changes within a submission. Then talk to the student and look at their drafts before deciding anything.
Ghost Act checks 500 words free with no account. After that, word packs start at £4.99 for 5,000 words, and Pro for a whole department is £100 a month for 500,000 words. Every check gets the full report.
JCQ’s AI Use in Assessments guidance (revised April 2025) says detection should be part of a holistic approach that considers all available information, that it may help to use more than one tool, that accuracy varies with a student’s English-language competency, and that teachers must not use AI tools as the sole marker. It names example tools and endorses none, including Ghost Act.
It varies by tool and by genre, which is why we publish ours by genre. On held-out data Ghost Act wrongly flags 3% of human writing while catching 81% of AI text. By genre: 0% of school essays, 7% of university essays and 14% of encyclopaedia-style prose. Results on independent 2026 datasets are lower (AUROC 0.83 on one), and they are published on our methods page too.
Jisc advises running a DPIA on any detector, and the DfE recommends that personal data is not put into generative AI tools. Ghost Act is designed so student work is never processed by a third party: the analysis runs in the browser, nothing is uploaded or stored on a server, nothing is used for training, and saved reports stay on the device. Your data-protection officer should still assess it. Pro’s assistant connection (MCP) is the exception: text sent through an assistant goes to that assistant’s provider and to Ghost Act’s server, where it is analysed in memory and never stored, so include that route in the assessment if you use it.
On Pro, yes, but the essay then goes to the assistant’s provider as well as to Ghost Act’s server, where it is analysed in memory and never stored. For student work, the browser check is simpler for data protection: nothing leaves your device.
Research on older detectors found they flagged non-native writers far more often, and several upheld ombudsman complaints involved second-language students. We measure it: on independent essays by English-language learners, Ghost Act flagged 1%; on essays by native writers, 2%. That is on the PERSUADE corpus of school essays; other writing may differ.
Ghost Act is tested on 24 AI models, including GPT, Claude and Gemini releases, and catches 82% of text from models it was never trained on. It cannot read Claude’s or Gemini’s keyed watermarks; only their makers can. Text rewritten by another AI model is caught much less often, and we publish by how much.
For AI writing, it can be a second source alongside your existing tools, or a private option if your institution is moving away from uploading student work. It is not a plagiarism or similarity checker: it does not compare submissions against the web or other students’ work.
Check it in your browser, read the reasons, and walk into the conversation knowing what the software found and what it could not tell. 500 words free, and student work never leaves the device.
Check a submission free