No. SafeAssign is a similarity checker, and there is no SafeAssign AI checker to turn on. It matches your text against four databases and returns a matching percentage, which says nothing about who wrote the words. Below: what those four databases are, why a matching engine cannot catch ChatGPT even in principle, how SafeAssign differs from Turnitin, and where an AI percentage in Blackboard actually comes from.
Does SafeAssign detect AI?
No. SafeAssign is a similarity checker and produces no AI score. It reports how much of your text matches sources it already holds. Anthology, which owns Blackboard, tested AI detection and decided not to put it in SafeAssign.
Checked against Anthology SafeAssign documentation Anthology AI white paper Turnitin guides
Two different questions get asked with the same words, and separating them resolves nearly all the confusion.
Does Blackboard check for plagiarism? Yes. SafeAssign is built in, at no extra cost, and many institutions have it on by default. Does it check for AI writing? No. SafeAssign returns a matching score against existing sources, and Anthology chose not to add an AI score.
Because Blackboard genuinely does ship a built-in checker, people reasonably assume the checker covers AI. It does not, and it structurally cannot. Anthology’s own documentation describes SafeAssign as a tool that compares submitted assignments against a set of academic papers to identify areas of overlap between the submitted assignment and existing works. Overlap with existing works is the entire job. Nothing in that sentence is about authorship.
So if you have seen a SafeAssign percentage and wondered whether it was an AI verdict: it was not. It was a measure of how much of your text matches text that already existed somewhere. For the wider platform picture, including third-party tools and proctoring, see does Blackboard detect AI.
Knowing which collection is which explains most surprising scores, including the ones that feel unfair.
Always checked
Every SafeAssign submission is compared against these three, wherever your institution is.
Only if your institution enabled it
One collection is opt-in, and being asked about it is a reliable sign SafeAssign is running.
Not checked, ever
These are the things people assume SafeAssign does. It does none of them.
Reading the resulting score. SafeAssign returns an Originality Report with an overall matching percentage and a per-source breakdown. A match is not automatically plagiarism: correctly quoted and properly cited material matches too, as do reference lists, standard phrases and assignment prompts you were told to restate. A high percentage on a paper built from many short quotations can be entirely legitimate, and a low percentage does not certify anything. The per-source list matters far more than the headline number.
This is the part worth being precise about, because it is the reason no amount of SafeAssign improvement would change the answer.
A similarity checker works by matching. It finds strings of your text that also appear in a source it holds. Its entire capability rests on the copied passage existing somewhere else first.
Text generated by a language model does not exist anywhere else. It is assembled fresh for your prompt. There is no prior document to match it against, so a matching engine has nothing to find. An essay written entirely by ChatGPT can legitimately return a very low SafeAssign score, and that is the system working exactly as designed, not failing.
AI detection works on a completely different principle. Instead of matching, it scores statistical properties of the writing itself: how predictable each word is given the ones before it, how evenly distributed the sentence rhythms are. That is a probability estimate about style, not a lookup, which is why it carries a false positive rate in a way a matching engine does not.
The two failure modes are worth holding side by side. A matching engine is nearly never wrong about what it found, and routinely silent about what it missed. A statistical detector always has something to say and can be confidently wrong. Neither is a verdict on its own. We explain the mechanism in how AI detectors work and how detector accuracy actually works, and the failure modes in can AI detectors be wrong and AI detector false positives.
Most LMS vendors have stayed quiet on AI detection. Anthology, which owns Blackboard, has not, and its position is unusually direct. It did not simply decide against adding AI detection to SafeAssign. It ran a test first and published the numbers.
Anthology’s white paper AI, Academic Integrity, and Authentic Assessment describes a beta test run across May and June 2023 with market-leading AI detection tools, explicitly to evaluate putting AI detection inside SafeAssign. 65 client institutions submitted more than 1,000 texts, some authentic and some AI-generated, then reported how accurately each was assessed.
The same paper sets out the outside research it was testing against, which is unusually candid for a vendor document. It cites Sadasivan and colleagues at the University of Maryland, who asked whether AI-generated text can be reliably detected and concluded it cannot, with simple paraphrasing enough to evade detection. It cites Weber-Wulff and colleagues, whose study of 14 detectors across six countries found accuracy ranging from just 33 to 81 percent depending on provider and method. And it cites the Stanford work on bias against non-native English writers. Anthology therefore decided not to make AI detection available natively in Blackboard Learn.
What Anthology recommends instead is pedagogical rather than technical: authentic assessment designed around critical thinking and personal reflection, assignments an AI cannot easily answer, peer assessment and group work that adds accountability, and personalised tasks where an instructor already knows a student’s voice. Its own AI product, the AI Design Assistant, helps instructors build rubrics, questions and course structure. It generates material for teachers. It does not judge students.
Why this matters if you have been flagged. If you are a student at a Blackboard institution facing an AI accusation, this is genuinely useful to know. The company that sold your institution its learning platform has published that AI detectors are too unreliable and too biased for high-stakes use. That does not settle a case by itself, but it is a documented vendor position rather than a student’s opinion, and it belongs in a calm written response. The full treatment of Anthology’s position, the third-party tools and the proctoring layer is on does Blackboard detect AI.
They overlap on similarity checking and diverge completely on AI. If you have been given a percentage, which tool produced it changes what the number means.
| SafeAssign | Turnitin | |
|---|---|---|
| How you get it | Built into Blackboard at no extra cost. Many institutions have it on by default. | A separately licensed product. Attached to Blackboard, Canvas or Moodle over LTI by an administrator. |
| Similarity checking | Yes. Four collections, reported as an Originality Report with a per-source breakdown. | Yes. Its own repositories, reported as a Similarity Report. |
| AI writing score | No. None. Anthology tested and decided against it. | Yes. A separate AI writing indicator, usually visible to instructors only. |
| Published error rate for AI | Not applicable: there is no AI score to be wrong. | A document-level false positive rate under 1 percent for documents assessed at 20 percent or more AI, no percentage shown below that threshold, and approximately 4 percent across the full distribution of documents. |
| What a low score proves | That the text is not in SafeAssign’s sources. Nothing about authorship. | Similarity and AI are separate numbers. A low similarity score says nothing about the AI indicator, and neither is proof. |
The practical consequence: if your Blackboard course uses only SafeAssign, no automated AI judgement is being formed about your writing at all. If it also has Turnitin attached, one is, and it is worth knowing which number you are being shown.
We read Turnitin’s own published figures clause by clause in is Turnitin’s AI detector accurate, compare the tools for teaching staff in Turnitin vs TextSight for teachers, and set out the honest options in Turnitin alternatives. Copyleaks, the other tool commonly attached to Blackboard, is covered in TextSight vs Copyleaks, and the full vendor table is on the comparison page.
You can still see an AI percentage in a Blackboard course. If you do, it did not come from Blackboard and it did not come from SafeAssign. There are only two other places it can have come from.
Which makes the first question about any AI number in Blackboard simply: which product generated this, and what does that product publish about its own error rate? If it came from Turnitin, there are published figures to discuss. If it came from a free web tool, there may be no published rate at all. Those are very different pieces of evidence and they should not be treated as interchangeable.
“SafeAssign flagged it for AI” is not a coherent statement, and saying so politely is a reasonable thing to do. SafeAssign does not produce AI scores. Asking which tool did is not a challenge to your instructor’s judgement, it is the first step to reviewing evidence at all.
Plenty of instructors at Blackboard institutions run SafeAssign for similarity and consult a separate AI detector when something about a submission does not sit right. That is a defensible workflow if the two are kept in their own lanes, and a poor one if their outputs get merged into a single verdict.
What the pairing genuinely gives you is two independent signals about different things. SafeAssign tells you whether the words already existed somewhere, with a source list you can open and read. An AI detector tells you how closely the prose resembles model output statistically. Agreement between them is not corroboration, because they are not measuring the same thing.
More for teaching staff in our guidance for teachers, AI detector for professors, sentence-level AI detection and For Educators.
Blackboard comes in two interfaces, Ultra and Original. The signals look slightly different but the logic is the same in both.
What you cannot tell from the student side is whether a third-party tool’s AI component is switched on, because its output is generally instructor-only. So ask. A course that uses AI detection should be able to say so.
TextSight is a text AI detector. It is not connected to Blackboard, it does not proctor, and it has no view of your screen or camera. There are two honest uses for it around a SafeAssign course.
Students can run a draft through it before submitting, to see whether writing they did themselves reads as AI and which sentences carry that risk. That is a self-check, not a guarantee: our score is a probability, and we cannot tell you what a third-party tool at your institution would say about the same text. What it can do is show you the specific lines that read as flat, so you can look at them rather than guess.
Instructors can use it for a second opinion on a passage before starting a conversation, with sentence-level highlights showing what drove the score. A second model agreeing is still not proof, only one more input to a judgement that stays yours.
Our detector is English-only, and we would rather say so here than let you discover it from a bad result. Try it on the AI detector page, or use the document detector if your submission is a file. We are explicit about what it cannot do in AI detection limitations and about how we test in our accuracy methodology. Plans are on the pricing page; see also For students and AI detection at college level.
The thing worth doing regardless of any tool: keep draft history on, in Google Docs or Word. It is the strongest authorship evidence you can hold, and it costs nothing until the day you need it. See how to prove you did not use AI.
This section is about process. We will not tell you how to change your writing to avoid a detector, and we do not publish that anywhere on this site. What settles a meeting is your ability to show and discuss your work.
If you wrote it and you have been flagged at a Blackboard institution, do these things, in this order.
More detail: accused of using AI in college, my essay was flagged as AI but I wrote it, and how to write an AI detection appeal letter.
If your question is not below, email support@textsight.ai. Replies go to a human inside our team, not a routing bot.
No. SafeAssign is a similarity checker, not an AI detector. It compares your submission against the open internet, the ProQuest ABI/Inform database, your institution’s own document archive and the Global Reference Database, then reports how much of your text matches those sources. It produces a matching percentage and makes no claim at all about who or what wrote the words. Anthology, which owns Blackboard, tested AI detection and decided not to build it into SafeAssign.
No. There is no SafeAssign AI checker and no AI setting inside SafeAssign to switch on. SafeAssign has one function, which is matching submitted text against sources it holds. If your Blackboard course shows an AI percentage, it came from a third-party tool such as Turnitin or Copyleaks that your institution licensed and attached over LTI, or from an instructor pasting your text into a consumer detector outside the LMS entirely.
No. There is no SafeAssign AI detector, and the two tools work on completely different principles. A similarity checker like SafeAssign looks up whether your exact words already exist in a source it holds. An AI detector scores statistical properties of the writing itself, such as how predictable each word is given the ones before it. SafeAssign only ever does the lookup, which is why Anthology’s own documentation describes it as identifying areas of overlap between a submission and existing works.
Yes, and that is the system working as designed rather than failing. A matching engine can only find text that already exists in its sources. Language-model output is assembled fresh for your prompt, so there is no prior document to match it against. An essay written entirely by ChatGPT can legitimately return a very low SafeAssign percentage. A low score means the text does not appear in SafeAssign’s sources. It does not mean a human wrote it.
Four collections. The open internet. The ProQuest ABI/Inform database, a licensed scholarly collection covering more than 1,100 publication titles and roughly 2.6 million articles from 1990 onward, updated weekly. Your institution’s own archive of everything previously submitted to SafeAssign there, which is what catches recycled coursework including your own from an earlier term. And the Global Reference Database, a cross-institutional collection reported at more than 15 million papers that students contribute to voluntarily.
SafeAssign is built into Blackboard at no extra cost and does similarity checking only. Turnitin is a separately licensed product that does similarity checking and also sells an AI writing indicator. So the practical difference for an AI question is that Turnitin can produce an AI percentage and SafeAssign cannot. Turnitin reports a document-level false positive rate under 1 percent for documents assessed at 20 percent or more AI, shows no percentage below that threshold, and reports approximately 4 percent across the full distribution of documents.
That a large share of your text matches sources SafeAssign holds. It is not by itself a finding of plagiarism. Correct quotations, properly cited passages, reference lists, standard phrases and an assignment prompt you were told to restate all match. A paper built from many short quotations can score high and be entirely legitimate, and a low percentage certifies nothing. The per-source breakdown in the Originality Report is what tells you whether a match is a problem, so read that rather than the headline number.
No, and no amount of improvement to SafeAssign would change that. Detecting ChatGPT is not a harder version of what SafeAssign does, it is a different task. Matching requires a prior copy of the text to exist somewhere. ChatGPT writes new text for each prompt, so a matching engine has nothing to compare against. Only a statistical AI detector attempts that question, and every one of them returns a probability rather than proof.
Sometimes, and less reliably than it matches verbatim text. SafeAssign works on textual overlap, so heavy rewording reduces what it can match even when the underlying idea came from a source. That is a limitation of matching engines generally, not a SafeAssign quirk. It is also why unattributed paraphrase remains an academic integrity issue that a low matching score does not clear you of.
If you wrote the work and cited your sources, a SafeAssign report is a routine check rather than an accusation. What it flags are overlaps, and the instructor sees which source each overlap came from. The useful habit is citing as you draft and keeping version history in Google Docs or Word, because draft history is the strongest authorship evidence you can hold if any question is ever raised.
Primary documentation, not other people’s blog posts. Check them yourself.
TextSight is an AI-detection and writing-trust tool. We help you check and improve your own work, and we do not help anyone disguise AI writing or evade an institution’s detection. Vendor statements on this page were checked against Anthology’s and Turnitin’s published documentation on 29 September 2026.
The platform view: Anthology’s published decision, third-party LTI tools, and the proctoring layer.
No built-in detector, but Canvas does ship the integration layer that lets Turnitin post scores back.
Moodle ships neither a checker nor a detector. Integration status, stated plainly.
Originality reports, what they compare against, and why that is not AI detection.
Whether a native TextSight integration for Blackboard exists yet. It does not, and we say so.
Who gets falsely flagged, why plain prose scores higher, and what to do about it.
Run your own writing through a detector before you submit, or get a second opinion on a passage before you raise it with a student. English-only, and the score is a probability, not a verdict.