humanflow

SafeAssign, reviewed

SafeAssign does not detect AI. That is not a gap someone forgot to fill — the company that makes it tested the idea across 65 institutions, decided against shipping it, and published its reasoning. It is the most interesting document any detector vendor has put out, and it argues against the product.

Last reviewed 7 August 2026 · The HumanFlow team

What it actually does

SafeAssign is Blackboard’s built-in plagiarism checker, owned by Anthology. It runs what its documentation calls “a unique text matching algorithm capable of detecting exact and inexact matching” between your submission and an existing corpus: academic papers, publications, articles and internet sources. It returns a similarity percentage and highlights the passages that matched, with the sources beside them.

That design is the whole answer to the AI question. Anthology’s own description of the model is blunt: the traditional anti-plagiarism approach “is based on building a large database of existing texts to compare with the learner’s submission, rather than on the ability to protect against texts created by generative AI.” A language model writes text that exists nowhere else. There is nothing to match.

The test Anthology ran, and published

Most vendors in this category publish research showing their product works. Anthology published research concluding that a product it was considering does not, and then did not ship it. The white paper says so directly: “both Anthology and participating clients concluded that AI detection is not currently fit for purpose in education.”

The test ran across May and June 2023 as part of evaluating whether to put AI detection inside SafeAssign. Sixty-five client institutions submitted more than 1,000 texts — some authentic, some AI-generated — then answered how accurately the detector had assessed them. Asked how often the tool correctly identified whether text was written by AI:

Almost always5%
Often15%
Sometimes37%
Rarely21%
Almost never22%

Anthology’s summary: 80% of respondents felt detectors were, at best, only able to “sometimes” identify texts correctly. The written feedback is sharper than the chart. One participant: “I used ChatGPT to generate five different texts. Four introduction paragraphs and two entire essays, and the AI detector completely misidentified all of them as being human generated.” Another: “It is easy to manipulate the tool into thinking the text was written entirely by humans.”

One caveat we would apply to our own evidence, so we will apply it here: this is a survey of what institutions concluded, not a controlled measurement. It records perceived accuracy after a hands-on trial, which is a real and useful signal about whether practitioners will trust a tool — but it is not the same kind of evidence as a peer-reviewed error rate, and it should not be quoted as one.

What this does not mean

Two mistakes are easy here, and both appear on pages currently ranking for this question.

It does not mean your Blackboard course has no AI detection. An institution can run a third-party detector alongside SafeAssign, through the LMS or outside it. SafeAssign having no AI detection is a fact about SafeAssign, not about your university.

It does not mean AI text will never be flagged by SafeAssign. It can be — for similarity. If a model reproduces phrasing that exists online, that overlap matches exactly like any other overlap. The score still means “this wording appears elsewhere”, which is a different accusation with a different defence.

Where your paper goes

Two separate stores, and the distinction matters. Your institution’s own database receives submissions by default. The Global Reference Database is separate and voluntary — papers there are checked against submissions from other institutions, which is what makes it useful and also what makes it permanent.

The commitment is stated plainly in Anthology’s student guidance: “When you submit your papers to the database voluntarily, you agree not to delete papers in the future.” To Anthology’s credit the same page also says “Blackboard LMS doesn’t claim ownership of submitted papers,” and that you are free to check a paper without contributing it.

The practical trap is timing: exclusion has to be set before submission. Work already submitted cannot be pulled back out afterwards. If you intend to publish a dissertation later, that is worth knowing on the day you upload it rather than the day you submit it to a journal.

What Anthology recommends instead

Its guidance to administrators points away from detection entirely and toward assessment design: tasks built on critical thinking, personal perspective and self-reflection; peer assessment and collaboration; and personalised assessments that let an instructor become familiar with how a particular student writes.

We think that is right, and it is worth noticing who is saying it. Anthology sells to the institutions that would have bought this feature. Declining to sell it, and publishing the reasoning, is a more credible signal than any vendor accuracy claim on this site or anyone else’s.

Where we stand

We sell a detector and a humanizer, so we have an obvious interest in how this question is answered, and you should weigh the page accordingly. We have published no measured accuracy figure for our own detector, and will not before our benchmark produces one under a method published in advance.

The honest summary for a student: a SafeAssign score is not an AI accusation, and if it is being treated as one, that is a factual error you can correct with the vendor’s own documentation. If a separate AI detector has been run on your work, that is a different conversation — start with what to do if you have been accused and how to prove you wrote it.

Compared against

  • SafeAssign vs Copyleaks SafeAssign against Copyleaks on what independent testing found, who each is sold to, and what happens to the text you submit.

Related

Common questions

Does SafeAssign detect AI or ChatGPT?
No. SafeAssign is a text-matching tool: it compares your submission against a database of academic papers, publications, articles and internet sources, and reports overlap. Generative AI produces original text that matches nothing, so there is nothing for SafeAssign to match it against. A SafeAssign score is a similarity score, and it is not evidence about AI use in either direction.
Will SafeAssign flag my essay if I used ChatGPT?
Not for being AI-written, because it does not look for that. It can still flag AI-written text for an unrelated reason: if the model reproduced phrasing that exists elsewhere online, that overlap will match like any other. So a high score means 'this wording appears somewhere else', not 'a machine wrote this' — and a low score is not a clearance either.
Is Blackboard adding AI detection to SafeAssign?
Anthology, which owns Blackboard, evaluated exactly that and decided against it. Its white paper states that after a beta test with market-leading AI detection tools, 'both Anthology and participating clients concluded that AI detection is not currently fit for purpose in education'. Its guidance to institutions recommends authentic assessment design instead of detection. An institution can still bolt on a third-party detector separately — so 'SafeAssign has no AI detection' does not mean 'this course has none'.
What was in Anthology's beta test?
It ran across May and June 2023 with 65 client institutions submitting more than 1,000 texts, some authentic and some AI-generated, then surveying how accurately the detector assessed them. 80% of respondents said the tool was at best only 'sometimes' correct. One participant reported generating six ChatGPT texts of which the detector 'completely misidentified all of them as being human generated'. Read it for what it is: a survey of what institutions concluded, not a controlled measurement of accuracy.
Does SafeAssign keep my paper?
Your institution's copy goes to its Institutional Database by default. There is a second, separate store — the Global Reference Database — which is voluntary, and the commitment attached to it is unusually firm: Anthology's own student guidance says 'when you submit your papers to the database voluntarily, you agree not to delete papers in the future.' It also states that 'Blackboard LMS doesn't claim ownership of submitted papers.' If an instructor wants submissions excluded, they must set that before students submit; it cannot be applied retroactively.
How accurate is SafeAssign?
As a plagiarism checker, the question has no single published answer, and Anthology does not advertise one. As an AI detector the answer is simpler: it is not one, so it has no AI accuracy rate. Beware any page quoting a percentage for 'SafeAssign AI detection accuracy' — there is no such figure to quote, and its existence on a page tells you what that page is worth.

Sources

  1. 1.AI, Academic Integrity, and Authentic Assessment: An Ethical Path Forward for Education Anthology (white paper), 2023
  2. 2.SafeAssign — student guidance, including the Global Reference Database Anthology / Blackboard Help, 2026
  3. 3.SafeAssign — instructor guidance Anthology / Blackboard Help, 2026
  4. 4.Protecting against AI plagiarism Anthology / Blackboard Help (administrator), 2026
  5. 5.Testing of detection tools for AI-generated text Weber-Wulff et al. — International Journal for Educational Integrity 19:26, 2023
  6. 6.GPT detectors are biased against non-native English writers Liang, Yuksekgonul, Mao, Wu & Zou — Patterns (Cell Press), 2023