Our pick · Verified 18 August 2026
Pangram — the lowest published false-positive rate, with LMS integration
Free tier · Individual $20/mo
For classroom use the number that matters is how often a detector wrongly flags a student who wrote their own work. Pangram publishes the strongest claim in the category — roughly 1 in 10,000 — and points to third-party evaluation by University of Chicago and University of Maryland researchers rather than only its own benchmarks. It integrates with Canvas, Moodle, Google Classroom and Brightspace, and the free tier gives 2,000 words a day, which is enough to test it against your own past student work before adopting anything.
Try Pangram free Opens on the vendor’s site · CASRAI referral link
Want sentence-level highlighting for the conversation? → — GPTZero is built around the classroom discussion rather than the score, which some instructors find more useful pedagogically.
Editorial disclosure: CASRAI has commercial referral arrangements with some of the vendors named on this page, and may earn a commission if you subscribe to them. We name them here regardless of whether a link is present. We only recommend tools our editorial team has independently researched. Read our full disclosure policy.
In summary
- False positive rate matters more than headline accuracy. At 500 essays a term, a 1% rate wrongly flags five students.
- Pangram: free tier at 2,000 words/day; Individual $20/mo. LMS integration for Canvas, Moodle, Google Classroom, Brightspace. Verified 18 August 2026.
- Non-native English writers are disproportionately flagged. Any classroom use must account for this explicitly.
- A detection score is never proof. It is a reason to open a conversation and look at other evidence.
- Test any detector against your own past student work before adopting it — that number is the only one that matters.
Detectors worth considering for classroom use
Ranked for teaching rather than for editorial screening. Test whichever you shortlist against your own students’ past writing before you adopt it.
#1 Pangram — our winner
The strongest false-positive evidence, with the LMS coverage classroom use needs.
Best for: Instructors and integrity offices who may have to defend a result to a student or an appeals panel
Price: Free tier · Individual $20/mo
Claims 99.98% accuracy and roughly a 1-in-10,000 false positive rate, with third-party evaluation by University of Chicago and University of Maryland researchers — independent evaluation being rare in this category. Integrates with Canvas, Moodle, Google Classroom and Brightspace, plus a Google Docs integration that is useful for looking at how a document developed rather than only at the finished text.
Strengths
- Lowest published false positive rate in the category
- Third-party evaluation rather than vendor benchmarks alone
- Broad LMS coverage — Canvas, Moodle, Google Classroom, Brightspace
- Free tier at 2,000 words/day is enough for a real evaluation
Trade-offs
- Newer entrant with a shorter track record than incumbents
- Independent evaluation is a snapshot; models shift constantly
- Institutional licensing is custom-quoted rather than transparent
Try Pangram free CASRAI referral link — disclosed on this page
#2 GPTZero
Built around the conversation with the student rather than the score.
Best for: Instructors who want something specific to discuss rather than an opaque percentage
Price: Free tier · Individual from ~$9.99/mo
Sentence-level highlighting is a pedagogical feature more than a technical one: it gives you particular passages to ask about instead of a single number, which makes for a far better conversation and a fairer process. LMS integration across Canvas, Moodle and Blackboard means it appears where marking already happens — which is what determines whether a detector actually gets used consistently.
Strengths
- Sentence-level highlighting supports a real discussion
- Appears inside the marking workflow
- Free tier available
Trade-offs
- Reported accuracy varies widely between studies (roughly 87–99.5%)
- Less suited to high-volume screening
Try GPTZero free CASRAI referral link — disclosed on this page
#3 Originality.ai
Better suited to volume auditing than to a single classroom.
Best for: Departments or integrity offices screening large numbers of submissions
Price: From $14.95/mo (2,000 credits)
Credit-based pricing suits bursty high-volume screening, with plagiarism and readability checks bundled in and unlimited team seats with per-member activity logs. The absence of LMS integration is a real constraint for classroom use — it means copying text into a separate tool, which is exactly the friction that causes inconsistent use.
Strengths
- Designed for volume rather than one-off checks
- Plagiarism checking bundled in
- Activity logging across a team
Trade-offs
- No LMS integration — significant for classroom workflows
- Accuracy figures are vendor-reported
- Credit model needs watching so you do not run dry mid-term
Try Originality.ai CASRAI referral link — disclosed on this page
Why false positives matter more than accuracy
Every detector advertises accuracy in the high nineties. That figure combines two errors that are not remotely equivalent for a teacher.
A false negative means a student who used AI is not caught. That is a bad outcome. A false positive means a student who wrote their own essay is accused of cheating and has to prove a negative — which is close to impossible, and which can affect their record, their progression, their visa status and their wellbeing. The asymmetry is enormous.
Put a class size to it. Suppose you mark 500 pieces of work in a term:
- At a 1% false positive rate, five students are wrongly flagged.
- At 0.1%, one student every two terms.
- At 0.01% — Pangram’s published claim — one student every twenty terms.
Five wrongful accusations a term, sustained across a department, is not a tolerable process. So when you evaluate a detector, ask for the false positive rate specifically, on text resembling your actual students’ writing — and treat an inability to give one as an answer in itself.
Some students are more likely to be flagged
This is the part that most classroom-facing content omits, and it is the part with the clearest potential to cause harm.
Research has repeatedly found that text written by non-native English speakers is disproportionately flagged by AI detectors. The reason is structural rather than incidental: the features detectors associate with machine generation — limited vocabulary variation, conventional sentence construction, predictable phrasing — overlap substantially with the characteristics of competent second-language academic writing. The tool is not detecting AI in these cases; it is detecting a writing register.
The same mechanism affects other groups. Autistic students and others whose writing is more formally structured. Students using assistive writing technology, including grammar and predictive-text tools. Students explicitly taught a formulaic structure — which is most students who have been drilled for a standardised exam.
The consequence is direct: a department that adopts detection without accounting for this will produce disproportionate accusations against international students and disabled students, whatever the aggregate false positive rate looks like. That is a discrimination risk as much as an accuracy problem.
The mitigations are practical. Test the detector against writing from your own student population before adopting it. Never let a score alone trigger a process. Weight draft and version history heavily, because it is far better evidence than any classifier. And be conscious that the students least able to contest an unfair process are frequently the ones most likely to be flagged by it.
A fair process, step by step
The tool is the easy part. This is the part that determines whether your use of it is defensible.
- Treat the flag as a prompt to look, not a finding. Nothing has been established at this point. Read the work properly yourself.
- Gather other evidence before speaking to anyone. Draft and version history in Google Docs or Word. Comparison with the student’s previously submitted work. Whether the content matches what was actually taught. Whether cited sources exist and say what the essay claims — fabricated citations are far stronger evidence than any detector score.
- Open a conversation, not a case. Ask the student to talk you through their argument, their sources and how they approached it. A student who wrote the work can almost always do this. Frame it as a discussion about their process, because at this stage that is genuinely all it is.
- Tell them what you observed and let them respond. Withholding the basis of a concern makes it impossible to answer and is the fastest way to lose an appeal.
- Escalate formally only with evidence beyond the score. If the detection result is all you have, you do not have enough.
- Document consistently — what was flagged, what else you looked at, what was discussed, what was decided. Inconsistent documentation across a department is what appeals panels find hardest to defend.
If your institution has no written policy on how detection results may be used, that gap is a more urgent problem than which detector you buy. Without one, individual instructors make ad-hoc decisions, outcomes vary by department, appeals succeed inconsistently, and the students harmed are the ones least equipped to challenge it.
Assessment design beats detection
Detection is a rearguard action against a capability that is improving faster than the classifiers chasing it, and every teacher relying on it is on a treadmill.
Assessment design is the more durable response, and it has the advantage of improving teaching rather than only policing it. Approaches that hold up: process-based assessment that marks proposal, draft and reflection alongside the final piece, making the work visible as it develops; in-class and oral components, including a short viva on a submitted essay, which is fast and remarkably informative; tasks anchored to specifics a general model cannot know — this week’s seminar discussion, a particular local dataset, the student’s own placement or fieldwork; personal reflection and application that requires the student’s own experience; and explicit permitted-use policies that tell students what is allowed and require disclosure, which converts an integrity problem into a transparency one.
That last point is worth dwelling on. A great deal of what currently gets flagged is students using tools in ways nobody told them were prohibited, because the rules were never stated. Clear guidance plus a disclosure requirement resolves more cases than any detector, and it does so before rather than after the harm.
Test it on your own students’ work first
Do not adopt on a published figure. Free tiers make a proper evaluation cost nothing.
- Collect known-human work from your own courses — ideally submitted before generative AI was widely available, so provenance is certain. A hundred pieces is enough.
- Deliberately include work by non-native English speakers in proportion to your actual cohort. This is the group most at risk, and the check that matters most.
- Run them and count the false positives. That number, on your students’ writing, is the only one your appeals process will care about.
- Then test detection using AI text that has been genuinely edited, not raw model output — because edited output is what actually gets submitted.
- Test the workflow. If it does not appear where you already mark, it will be used inconsistently and abandoned within a term.
- Repeat annually. Performance shifts as models change.
Run the false-positive test before you adopt anything
The free tier gives 2,000 words a day — enough to work through a batch of your own students’ past writing, including non-native English speakers, and get the one number that determines whether using it would be fair.
Free tier · Individual $20/mo
Try Pangram free Opens on the vendor’s site · CASRAI referral link
Frequently asked questions
What is the best AI detector for teachers?
For classroom use the deciding metric is the false positive rate, and Pangram publishes the strongest claim — roughly 1 in 10,000 — with third-party evaluation by University of Chicago and University of Maryland researchers, plus integration with Canvas, Moodle, Google Classroom and Brightspace. GPTZero is a strong alternative if you want sentence-level highlighting to structure the conversation.
Can I fail a student based on an AI detector result?
No. A detection score is a probabilistic signal about textual features, not evidence of what a student did. It should prompt you to look more closely and gather other evidence — draft history, comparison with prior work, whether cited sources exist and say what is claimed, and a discussion of the content. If the score is all you have, you do not have enough.
Do AI detectors discriminate against international students?
Research has repeatedly found that text by non-native English writers is disproportionately flagged, because the features detectors associate with machine generation overlap with those of competent second-language writing. Autistic students, assistive-technology users and students taught a formulaic register are similarly affected. Any adoption must account for this or it will produce disproportionate accusations.
How many students would a 1% false positive rate wrongly flag?
Across 500 pieces of work in a term, five students. At 0.1% it is one student every two terms; at 0.01% it is one every twenty terms. That difference is what separates a workable process from one that generates wrongful accusations routinely.
How should I raise a concern with a student?
Gather other evidence first, then open a conversation rather than a case. Ask them to talk you through their argument, sources and process — a student who wrote the work almost always can. Tell them what you observed and let them respond; withholding the basis of a concern makes it unanswerable and is the fastest way to lose an appeal.
Is there a better approach than detection?
Assessment design, and it improves teaching rather than only policing it. Process-based assessment marking drafts alongside the final piece, short oral components, tasks anchored to this week’s seminar or the student’s own placement, and explicit permitted-use policies with a disclosure requirement all resolve more cases than a detector — and they do so before the harm rather than after.
How do I evaluate a detector before adopting it?
Collect around a hundred pieces of known-human work from your own courses, deliberately including writing by non-native English speakers in proportion to your cohort, run them through the free tier and count the false positives. Then test detection against realistically edited AI text, test the LMS workflow, and repeat annually.
