Key Pointers

• The binding constraint on your 2027–28 purchase isn’t accuracy. It’s whether the tool can be used compliantly as one input to a human decision, because that’s what the new state AI statutes and institutional policies now require.
• Oklahoma’s SB 1734 requires every school board to adopt an AI policy before the 2027–28 school year, and bars AI from being the primary basis for grading, discipline, placement, promotion or retention. Maryland, Idaho and Virginia passed district-policy mandates the same year.
• Only two of the five tools below publish their prices. Turnitin, GPTZero and Copyleaks are quote-only or unretrievable, which makes budget comparison genuinely hard.
• Not one of the five publishes a data retention period. All say you can request deletion. None states a term. That’s the single sharpest question to put in writing during procurement.
• Washington State University cancelled its Turnitin AI detection contract mid-semester in February 2026, estimating roughly 1,485 false positives across 148,547 Fall 2024 assessments, and noting that 33% of its AI integrity cases from 2023 to 2025 ended in a “not responsible” finding where detection was the only evidence.
• Disclosure before you read further: Quetext publishes this guide and appears in the comparison. We’ve applied the same checklist to ourselves, including where we don’t currently publish something.

What changed between 2023 and now

The first wave of AI detection was bought on accuracy claims. The second wave is being bought, or dropped, on defensibility.

Three things moved at once.

Institutions started turning detection off. Vanderbilt disabled Turnitin’s AI detector in August 2023. By August 2026, Inside Higher Ed reported that Yale, Johns Hopkins and Indiana had policies banning or discouraging detectors as sole evidence, and that at least a dozen universities including Northwestern, Georgetown and NYU had disabled Turnitin’s AI detection outright. We’ve tracked that pattern in universities restricting AI detectors.

States started mandating district AI policies. In 2026 alone, 87 AI-in-education bills were introduced across 28 states and 13 were enacted. Four states now require districts to adopt AI policies aligned to state guidance: Idaho, Maryland, Oklahoma and Virginia.

The licence terms became the story. The most significant 2026 development wasn’t about false positives at all. Several UK universities, including Southampton and Newcastle, announced they would end their Turnitin relationships over proposed terms that would have allowed student submissions to be used in developing AI services. Turnitin postponed those terms to September 2027 after sector objections.

Put together, that’s a different buying question than the one most procurement checklists were written for.

The question that actually governs your purchase

Not “which detector is most accurate?” but “can this tool be used, lawfully and defensibly, as one input to a decision a human makes?”

Three separate sources converge on that framing.

Oklahoma’s SB 1734 states that responsibility for AI-informed decisions cannot be delegated to an AI system or vendor, requires an educator to review AI outputs before use in assessment or decisions, and says AI cannot be the primary basis for high-stakes outcomes.

Washington State University’s policy says it will not allow a detector to be the sole support for a misconduct case, and requires the instructor to meet the student and find misconduct more likely than not.

Turnitin’s own position, as quoted by education counsel, is that the AI score is not proof of misconduct and should not be the sole basis for a finding.

A guide organised around benchmark scores will age badly. One organised around that question won’t.

Six things to evaluate

Evidence depth, not just a score

A single document-level percentage is the hardest thing to defend in an appeal. What you want is the tool showing its work: which passages triggered the result, and ideally what they matched.

This varies more than you’d expect. GPTZero documents scanning at document, sentence and word level with highlighted passages. Turnitin describes scoring each sentence and averaging into a document-level prediction. Copyleaks returns segment offsets through its API, which a client can render as highlights, though its marketing implies a richer view than its technical docs describe. Pangram accounts for errors at segment level.

Ask to see a real report in the demo, not a screenshot.

Student data retention and model training

This is where procurement has the most power to shape terms, and where every vendor is currently vague.

None of the five publishes a retention period. All offer deletion on request. Only Pangram states clearly that submissions are not used to train its models. GPTZero and Copyleaks are silent on the question. Turnitin’s position is that it trains a classifier rather than a generative model, while telling Times Higher Education it may use anonymised student submissions to improve its tools.

Three separate legal sources converge on this one question. California’s Education Code 49073.1 requires a contract certification that records will not be retained by or available to the vendor after the contract ends. The amended COPPA Rule, in full effect since April 2026, requires a written retention policy with a specific deletion timeframe and expressly prohibits indefinite retention, and treats AI model training as outside the “integral to the service” category, meaning it needs separate verifiable parental consent. Oklahoma’s SB 1734 adds a data minimisation duty.

So the question to put in writing is concrete: does student writing submitted to this tool persist, for how long, and is it ever used to develop your models?

False positives, and who absorbs them

Every vendor claims a low false positive rate. The useful move is to multiply whatever rate they quote by your actual submission volume.

Washington State did exactly that. At Turnitin’s self-reported 1%, the roughly 148,547 assessments it analysed in Fall 2024 implied about 1,485 wrongly flagged papers. That arithmetic, applied to a vendor’s own published figure, is what ended the contract.

There’s a documented equity dimension too. A 2023 Stanford-led study found detectors falsely flagged just over 61% of genuine TOEFL essays written by non-native English speakers. A February 2026 replication found that figure had fallen to 23.1% for one commercial detector on the same essays. Better, and still roughly one in four. We’ve covered this in detector bias against ESL students.

Ask what the rate is, on what corpus, measured when. Then ask what happens to the students in that percentage.

One contract or two

Plagiarism checking and AI detection are separable products at most vendors, and the separation is where budget leaks.

Turnitin licenses them separately enough that Washington State cancelled AI detection while keeping plagiarism checking. GPTZero sells plagiarism checking as a premium add-on. Copyleaks sells three distinct SKUs: AI alone, plagiarism alone, and a bundle. Pangram includes similarity detection on paid plans, though that’s narrower than a full plagiarism product. Quetext includes both in every tier, including the free one.

Two contracts means two renewals, two DPAs and two vendor reviews.

Pricing you can actually compare

Here’s an awkward finding. Of the five tools in this guide, only Pangram and Quetext publish prices.

Turnitin publishes nothing; pricing is quote-only. Inside Higher Ed reported disclosed figures of £87,621 for Southampton across one year and £50,892 for UEA, which gives a sense of institutional scale but won’t help you build a budget line. GPTZero’s and Copyleaks’s public pricing pages did not render retrievable figures when we checked on 30 October 2026.

Pricing units differ enough to make like-for-like comparison hard anyway: per seat, per credit where one credit equals 250 words, per word allowance, or per institution. Normalise to your own expected annual volume before comparing anything.

LMS integration, verified rather than implied

“Integrates with all major LMS platforms” is a sentence that should trigger a follow-up question.

Turnitin publishes an explicit LTI 1.3 list including Canvas, Blackboard, Moodle and D2L, with Google Classroom and Schoology absent. GPTZero documents Canvas and Google Classroom only. Copyleaks names no LMS on its education page, though a Moodle plugin and a Canvas partnership exist. Pangram names no LMS at all. Quetext currently integrates with Anthology, lists Canvas and Moodle as coming soon, and supports embedding through its developer API.

Get the list in writing, with versions. If Canvas is your environment, whether Canvas detects AI writing covers what the platform does and doesn’t do natively.

The comparison table

Status as of 30 October 2026, from each vendor’s own public pages. Where a vendor doesn’t publish something, that’s recorded as not published rather than guessed.

	Turnitin	GPTZero	Pangram	Copyleaks	Quetext
Price published?	No, quote only	Not retrievable	Yes	Not retrievable	Yes
Entry price	Not published	Not published	$20/mo individual; $20/seat team	Not published	$14.99/mo Essential
Education pricing	Quote	Quote	Not published	Enterprise, quote	Quote
Plagiarism + AI together	Separable	Plagiarism is an add-on	Similarity included	Sold as three SKUs	Both in every tier
Evidence depth	Sentence scored, averaged	Document, sentence, word	Segment-level	API offsets; UI unclear	Line-by-line AI analysis
Published accuracy method	No (figure dates to 2023)	Partial, self re-scored	Yes, detailed corpus	No	No
Retention period	Not published	Not published	Not published	Not published	Not published
Trains on student work?	Classifier, terms in flux	Not stated	Stated: no	Not stated	Not stated
FERPA stated	Yes	Yes	No	No	Not stated
COPPA stated	Not found	Yes	No	No	Not stated
SOC 2	Yes	Yes, Type II	Yes, Type 2	Yes, plus SOC 3	Not stated
LMS, documented	Canvas, Blackboard, Moodle, D2L +	Canvas, Google Classroom	None named	Moodle, Canvas (indirect)	Anthology; Canvas/Moodle coming
Bulk scanning	Via LMS submission	250 files per batch	Not documented	Not documented	20–100 files by plan

Read the bottom half of that table before the top half. The rows where most columns say “not published” are the rows your district counsel will care about.

The evidence asymmetry nobody advertises

Every last one of the five vendors guarantees high accuracy, but only one can provide sufficient transparency for verification. Pangram not only publishes an entire corpus, sample sizes, per domain breakdowns, and clearly defined scope limitations, but it also claims to have configured the algorithm for higher false negative than false positive rate. Regardless of what you think of those numbers, it is an entirely different level of transparency.

GPTZero does have a real methodology employed in the process, but also a grandiloquent figure that is achieved by re-scoring a certain external evaluation report following questioning the way the initial authors called their API. It is worth reading, and attributing properly.

The “less than 1%” figure by Turnitin is related only to documents flagged with over 20% AI and no methodology of how that number was computed and comes from a blog published in June 2023. The false positive rate on the other hand has been around 4%.

Copyleaks claims 99% accuracy and 0.03% false positive rate without any available methodology for both numbers.

Quetext along with this guide does not disclose any accuracy methodology of test corpus either, and we would prefer this assertion to be made directly instead of letting the comparison table to imply otherwise.

In case that accuracy is one of the criteria scoring for your RFP, penalize the evidence, not the statement.

Questions to ask every vendor

Take these into the demo. Each one is tied to something that will be asked of you later. If you’re evaluating for a single department rather than a district, our roundup of AI detectors for teachers covers the classroom-level view instead.

1. How long does your organization keep data, and will you include this in the agreement? (Amended COPPA Rule creates a deadline for deletion.)
2. Does the writing submitted by a student to you help improve the models you have? Will you guarantee this in writing? (Oklahoma has a data minimization obligation, and COPPA considers training a non-core activity.)
3. Will you assure us that after the end of the agreement there will not be any student records kept by you or access to it?
4. Can you fall into the category of “school official” under FERPA, working under our control and covered by our annual notice?
5. Which false positive rates does your organization provide?
6. What other information does the report provide, and can this evidence be used for an appeal?
7. How can it be set to enable the closure of a case only as an addition?
8. Which LMS integrations are offered for now and which LTI versions do they cover?
9. Are you willing to sign a Data Processing Agreement in the Student Data Privacy Consortium repository?
10. What is the pricing system, and what do we need to pay for a year?

Questions 1, 2 and 3 are the ones that most often change a shortlist.

A sample AI detection policy

Adapt freely. This is a starting point, not legal advice, and your counsel should review it against your state’s requirements.

1. Purpose. This policy governs the use of AI writing detection tools in [District]. It exists to support academic integrity while protecting students from unfounded accusations.
2. Status of a detection score. An AI detection score is a signal, not a finding. No score, by itself, constitutes evidence of academic misconduct. No grade, disciplinary outcome, placement, promotion or retention decision may rest primarily on the output of an AI system.
3. Required corroboration. Before an integrity concern is filed, the instructor must review at least one additional form of evidence, which may include draft history, document version history, in-class writing samples, or a conversation with the student about their work.
4. Student conversation. The student will be told of the concern and given an opportunity to respond before any formal report is filed.
5. Appeal. A student may appeal any integrity finding to [body]. The appeal file will include the full detection report, the corroborating evidence, and a record of the student conversation.
6. Equity. Staff will be informed that detection tools have documented higher error rates for writing by multilingual students and some neurodivergent students. A flag on such a student’s work carries no additional weight.
7. Disclosure. [District] will publish annually the AI tools in use, their vendors, the categories of student data collected and the educational purpose, and will provide a written opt-out pathway with no academic penalty.
8. Vendor conditions. No AI detection tool will be procured without a stated data retention period, a written warranty on model training, and a signed data privacy agreement.
9. Review. This policy will be reviewed annually before budget season.

Clause 7 tracks Oklahoma’s annual parental disclosure and opt-out requirements. If you’re in Oklahoma it’s mandatory; elsewhere it’s defensible practice that costs little.

Where Quetext fits

We publish this guide, so treat this section as what it is.

What’s verifiable: plagiarism detection and AI detection are included in every Quetext tier including the free one, so there’s no second contract. Self-serve pricing is published, starting at $14.99 per month for Essential and $29.98 for Professional, with bulk uploads of 20 and 100 files respectively. Institutional plans go through our education team, and the AI content detector provides line-by-line analysis rather than a bare score. Teachers running whole classes can use bulk scan.

A number of things that we do not publish at present and that we should include a data retention period, model training guarantee, statements regarding FERPA and COPPA, and accuracy methodology. 4 out of 5 vendors mentioned above can be found missing in terms of all of those. However, please remember that having those gaps merely serves as an explanation; it is by no means an excuse. Therefore, if you are evaluating us, you should basically ask questions 1-5 above and treat us the same way as you would treat everybody else.

What we will not say is that the true accuracy of our solutions matches that of any particular competitor. The truth is that nobody has published enough content in this category to be able to make such comparisons, including us.

What to do before budget season

Make sure that the compliance rows are listed first instead of the accuracy claims. The only relevant row will be Retention, training, FERPA, DPA. Those would pass legal scrutiny.

Start a pilot in one college using the outputs of one department and analyze 10 reports fully. You need to decide if a professor would be able to argue the conclusions drawn in these reports if asked to explain them in case of a complaint.

The policy should be ready before you sign the agreement. The experience Washington state had shows that the tool is not bad, however, there was no evidence standard for using it.

Also reckon with human factor that should account for every statute that is written differently.

Frequently asked questions

What should a school look for in an AI detector in 2027?

Place the emphasis on defensibility compared to accuracy in the press. Check if the application provides passage-level proof instead of just the general score and ask the vendor to indicate the data retention time and whether or not its models are trained on student essays. Make it clear that the application can be set up in such a way that the score cannot be used to make a conclusion.

• The depth of evidence matters more than a single percentage of results.
• The answers to retention and training questions should be in the contract.
• The score must only be one of the pieces of information, but not the final judgment.

Can an AI detector score be used as proof of cheating?

No, and vendors admit that. According to Turnitin, its score does not indicate wrongdoing, and that it should not be the sole factor in making a determination. Washington State University prohibits detection being the only basis for making a determination. Oklahoma prohibits the use of AI detection being the sole basis for grades or a punishment. Evidence is required in the form of drafts, version history or a conversation with the student.

• Vendors say it is not sufficient proof
• Now at least one state has made it illegal
• Introduce evidence as an essential aspect of the policy

How much does an AI detector cost for a school?

It has turned out to be way harder to answer than it ought to be. Pangram and Quetext are the only tools that provide pricing on their solutions, and five others checked out on 30 October 2026. For Turnitin, only quote-based pricing has been given, and estimated UK institutional figures varied around £50,000 to £88,000 annually. Sometimes the quote may include other pricing terms, like price per seat, a specific number of words allowed, or credits of 250 words-so normalize using your own annual submission volume. The costs are deal-specific and vary by number of seats, word allowance, or token allowances of 250-word units, so do pre-analysis for each annual submission volume requested. Pricing presently checks are Pangram, Quetext, and Turnitin. Prices will have a peculiar sense of taking depending on everything, more like a deal, up to £50,000 to £88,000 from UK institutions.

Mostly quoted; units do not always compare directly; model your real volume before requesting quotes.

• Most vendors quote only
• Units differ
• Low volume request for quote estaciones

Do AI detectors violate student privacy laws?

Not necessarily, but procurement needs to perform some actual work. The vendor must meet the FERPA definition of an official of the school that is under the district’s control. The amended COPPA Rule specifies when to delete data and recognizes model training, requiring parental consent. California requires certification about the non-retention of data after a service is no longer being offered. Among the five mentioned vendors, there is no period of retention published, so it is recommended to inquire.

• ERPA school official status must be determined
• COPPA prevents unlimited retention for children under 13 years of age
• Get confirmation regarding retention and training in writing

Are AI detectors biased against English language learners?

There is proof that this is true, although the situation has improved. A 2023 study led by Stanford showed that 61% of TOEFL essays written by non-natives were wrongly detected as being fraudulent. However, a February 2026 study showed that this number fell to only 23.1% of essays published on the same papers by the same company. In that sense, one out of four honest papers continues to be a serious fairness issue for any school district.

• Error rates differ for multilingual authors
• While the gap may have been reduced, it hasn’t been closed.
• Staff instructions should reflect this.

Why are universities dropping AI detection?

The two different reasons that were given for different conclusions, which often were confused for each other. One group, such as Washington State in February 2026, referred to volume of false positives and clear evidence showing that one third of their AI integrity cases ended with an acquittal where there was only one kind of proof – detection. Another group, including several universities in the UK in 2026, acted on the basis of the license conditions allowing institutions to use submissions for the purpose of developing AI services. The second reason is something new.

• First reasons are connected with accuracy and volume.
• Data and license conditions led to events of 2026.
• Both of the reasons may be viewed as procurement issues rather than pedagogical ones.

Should we buy plagiarism checking and AI detection separately?

Typically, not if the bundled option works fine for you. Using separate products involves two renewals, two privacy agreements, two vendor security checks, and there are administrative costs involved that are often overlooked. Some vendors sell separate solutions, which was how the state of Washington managed to cancel AI detection while keeping the minute checking feature. Thus, separability has at least some positive sides.

• Advantages of packing
• Advantages of separability
• What type of contract is preferable for you