SafeAssign vs Turnitin: What Each One Actually Checks

SafeAssign and Turnitin both check papers for copied text, but only one of them looks for AI. SafeAssign, Blackboard's built-in tool, does similarity checking and — as of this writing — has no AI writing detection at all. Turnitin does both: similarity plus an AI indicator it added in April 2023. That gap matters less than students think, for reasons worth spelling out.

If you've landed here, you're probably asking one of two questions. Either you're comparing the tools on their merits — an instructor or administrator weighing what each catches — or you're a student doing risk arithmetic: my course runs SafeAssign, not Turnitin, so does anyone see AI use? This post answers both honestly, including the part the second group won't love: the inference "SafeAssign can't detect AI, therefore AI won't be detected" fails often enough that betting an academic record on it is a genuinely bad trade.

What SafeAssign is

SafeAssign is the plagiarism-checking service built into Blackboard Learn, the LMS now owned by Anthology. When an instructor enables it on an assignment, every submission gets compared against a set of sources, and the instructor receives an Originality Report showing what percentage of the paper matches existing text, with the suspected source for each matching passage.

Per Anthology's documentation, SafeAssign uses "a unique text matching algorithm capable of detecting exact and inexact matching between a paper and source material" — meaning it catches light paraphrase and word-swapping, not just copy-paste — and compares submissions against several databases of articles dating from the 1990s onward. Anthology's current documentation lists three sources: the internet, your institution's own archive of past submissions, and the Global Reference Database — a cross-institutional pool Anthology describes as "over 15 million papers volunteered by students from Blackboard client institutions." Students can exclude their work from the institutional and global databases at submission. Older write-ups often add the ProQuest ABI/Inform database to this list; it no longer appears in Anthology's documentation, and we could not verify it as current.

Two design details are worth noticing because they differ from Turnitin's defaults. First, the Global Reference Database is opt-in at the student level: you choose whether your paper joins the cross-school pool. Second, Anthology is explicit that the Originality Report doesn't judge anything by itself — its own help text says the report "does not state whether a phrase that matches a source is properly referenced. Your instructor must read the report and determine if you used proper citations." A 30% match on a properly quoted literature review is fine; an 8% match on a stolen paragraph isn't. The number was never the verdict.

What Turnitin is

Turnitin is the incumbent. Founded in the late 1990s, licensed by thousands of institutions worldwide, and integrated into every major LMS — including Blackboard itself, which matters later — it built its reputation on similarity checking against a very large corpus: web content, published academic articles, and a repository of student papers accumulated over two decades. Turnitin's media kit claims "Over 91 billion web pages" and "Over 1.9 billion student papers," while its own product pages say the crawler "has indexed 54 billion web pages to date" — the company publishes both figures without reconciling them; the precise numbers matter less than the structural point that its student-paper repository is opt-out-by-configuration and enormous, where SafeAssign's cross-institutional pool is opt-in and smaller.

Then, on April 4, 2023, Turnitin added the thing this comparison usually turns on: an AI writing indicator, delivered inside the existing Similarity Report. It estimates what percentage of a document's prose was AI-generated, based on a classifier that measures how statistically machine-typical the text is — the perplexity-and-burstiness family of methods we unpack in how AI detectors work. Turnitin claims 98% accuracy with under 1% false positives, but both figures apply only to documents where more than 20% of the text is flagged; scores of 1–19% display as an asterisk rather than a number, Turnitin's own concession that low scores are unreliable. In its first year the indicator screened over 200 million papers; about 11% showed at least 20% AI writing. The full picture — thresholds, false positives, the Vanderbilt walk-back — lives in our Turnitin AI detection hub.

Does SafeAssign detect AI? (Verified: no)

This is the load-bearing question, so let's be precise about the current status rather than repeating 2023-era blog posts.

As of this writing, SafeAssign has no AI writing detection feature. Anthology's own product documentation describes SafeAssign purely as an originality/text-matching service, with no AI-detection component. More interesting: Anthology has taken a public position against the detector approach altogether. Its guidance on protecting against AI plagiarism argues that "the best defense against plagiarism is to provide instructors with opportunities to use AI-powered technologies to support authentic assessment" — designing assignments around critical thinking, personal perspective, self-reflection, and observed student work — rather than relying on detection tools whose false-positive problems are well documented.

Give Anthology its due: that's a defensible, even principled position. The research on detector false positives is genuinely ugly — Liang et al. in Patterns (2023) found seven detectors falsely flagged an average of 61.22% of human-written TOEFL essays by non-native English speakers, and OpenAI retired its own classifier in July 2023 after it caught just 26% of AI text. Vanderbilt disabled Turnitin's AI indicator in August 2023 over exactly this math. Anthology looked at that terrain and declined to ship a detector. Whether that's wisdom or a competitive rationalization, the practical fact stands: SafeAssign itself will not flag AI-generated text. A fully machine-written essay with no copied sources will typically sail through SafeAssign with a low originality match, because it isn't copied — it's freshly generated.

If that status ever changes, it will change loudly — an AI-detection launch is a press release, not a quiet patch — but verify against Anthology's current documentation before relying on any blog post, including this one.

How the similarity checking itself differs

Set AI aside for a moment; the two tools aren't identical even at their shared job.

Database reach. Turnitin's decisive advantage is its student-paper repository, built over twenty-plus years across thousands of institutions with storage as the default in most configurations. Recycled essays — this year's student borrowing from a prior cohort, papers bought from mills that resell, self-plagiarism across courses — get caught by Turnitin far more often, simply because the prior paper is more likely to be in the database. SafeAssign sees your institution's own archive plus only those outside papers students volunteered into the Global Reference Database. For web and published content, both are credible; for the gray market of previously submitted work, Turnitin's net is much wider.

Matching and reporting. Both catch inexact matching, not just verbatim copying. The reports differ in maturity: Turnitin's Similarity Report offers granular filters (exclude quotes, bibliography, small matches), source-by-source drill-down, and — since 2023 — the AI panel alongside. SafeAssign's Originality Report is serviceable and simpler: overall match percentage, per-passage suspected sources, and instructor-side controls like draft submissions that check text without storing it.

Cost and ecosystem. SafeAssign is bundled with Blackboard Learn at no extra licensing cost, which is often the honest reason a school "uses SafeAssign": it's already paid for. Turnitin is a separate, substantial license. That budget line, not a considered stance on AI detection, frequently decides which report your instructor sees.

The comparison at a glance

"My school uses SafeAssign, so AI won't be detected" — the four ways this bet fails

Here's the risk arithmetic done honestly. Each failure mode is independent; you'd need to win all four.

1. Instructors run secondary tools, constantly. Nothing stops an instructor from pasting suspicious work into GPTZero, Copyleaks, or any of a dozen public detectors — and in SafeAssign-only schools this is more likely, not less, because the official tool leaves the AI question unanswered. An accusation built on an unofficial consumer tool stands on shakier procedural ground, true — we cover that dynamic in Turnitin vs GPTZero — but shaky ground still starts a process: a meeting, a viva-style questioning about your argument, a request for drafts. If the work isn't yours, that conversation goes badly regardless of which detector prompted it.

2. "Blackboard school" doesn't mean "SafeAssign-only school." Turnitin integrates directly into Blackboard Learn, and Anthology's own administrator documentation lists it among available plagiarism tools. Plenty of institutions run Blackboard as the LMS while licensing Turnitin for some or all courses; individual departments sometimes hold their own licenses. The tool named in your syllabus this term is not a reliable guide to what your submission actually passes through — or what next term's does.

3. The policy doesn't care what the software catches. Academic integrity rules prohibit unauthorized AI use, full stop — not "unauthorized AI use where detection software is installed." If your course bans AI assistance and you use it, you've committed the violation at the moment of submission. Detection only changes the probability of the consequence, and violations have long tails: tools improve, policies tighten, papers get re-examined when a pattern emerges, and a graded essay sits in an archive for years. Vanderbilt turning off Turnitin's AI indicator didn't legalize AI-written essays at Vanderbilt; Anthology declining to build a detector doesn't authorize anything either.

4. Humans detect what software doesn't. Most integrity cases still start with an instructor's read, not a scanner: an essay that doesn't sound like your discussion posts, a sudden register shift mid-paper, a confident survey of literature you can't discuss when asked, a citation that doesn't exist. Hallucinated references — a signature failure of AI-generated academic writing — get caught by a human checking the bibliography, no detector required. The instructor who's read four hundred sophomore essays has a trained ear that doesn't appear on any feature list.

None of this is a scare tactic; it's just the actual failure surface. The honest conclusion cuts both ways. If AI is banned in your course, the absence of a detector is not permission, and gambling on tooling gaps is a bad bet with your transcript as the stake. If AI is permitted with disclosure, then the tool question is moot — disclose and document. And if you're using AI legitimately as a drafting aid where allowed, the durable protection isn't beating a scanner; it's work you can defend out loud, with drafts and version history behind it. We've written about what that looks like for essay writing specifically.

One note for the worried-honest rather than the tempted: if you write in a SafeAssign school and an instructor's side-tool scan flags your genuine work, the false-positive research above is your context, and process evidence — version history, notes, drafts — is your answer. If you want to see your own writing the way a detector sees it before anyone else does, HumanFlow's free detector shows a sentence-level readout of what pattern-matches machine text. It won't tell you what any instructor's tool will say — it doesn't promise to predict or beat any detector, because nobody can honestly promise that — but knowing which passages read as statistically flat, and why, beats guessing.

For instructors: what each report is actually evidence of

A brief word to the other audience of this comparison, because misreading reports harms students in both directions.

A SafeAssign Originality Report is evidence about text overlap with known sources, nothing more. A low score does not mean the student wrote it — AI-generated text scores low by nature — and a high score does not mean misconduct, since properly cited quotation inflates matches. Anthology says exactly this in its documentation.

A Turnitin AI score is evidence about statistical machine-typicality, nothing more. It doesn't know who typed. Its own vendor displays 1–19% as an asterisk because those numbers aren't trustworthy, restricts its accuracy claims to documents over the 20% line, and — per its chief product officer — deliberately leaves roughly 15% of AI text unflagged to limit false accusations (BestColleges, April 2023). Treating either number as a verdict, rather than as the start of a human conversation with the student, is precisely the misuse both vendors warn against — and the false-positive burden lands hardest on non-native English writers, per the Patterns study above.

The strongest integrity practice looks the same whichever tool your LMS ships: assignments that require visible process, conversations before accusations, and scores used as prompts rather than proof.

FAQ

Does SafeAssign detect AI writing like ChatGPT? No. As of this writing, SafeAssign is a similarity checker only — it matches submitted text against databases of existing sources. Anthology, Blackboard's owner, has publicly favored authentic-assessment design over AI detectors and has not added AI detection to SafeAssign. Freshly generated AI text typically shows a low match score because it isn't copied.

Is Turnitin better than SafeAssign? Turnitin does more: mature similarity checking against a much larger student-paper repository, plus an AI writing indicator since April 2023. SafeAssign is a competent similarity checker bundled free with Blackboard. "Better" depends on whether an institution wants AI detection at all — Vanderbilt disabled Turnitin's indicator over false-positive concerns, so the extra capability is itself contested.

Can my school use both SafeAssign and Turnitin? Yes. Turnitin integrates into Blackboard Learn, and Anthology's administrator documentation lists it among available plagiarism tools. Some institutions run SafeAssign by default while licensing Turnitin for particular departments or courses. The tool on one assignment doesn't tell you what every course uses.

Will AI-written work pass SafeAssign? It will usually pass SafeAssign's scan, since generated text rarely matches database sources. It will not necessarily pass the course: instructors run secondary detectors, spot voice shifts and hallucinated citations themselves, and integrity policies prohibit unauthorized AI use regardless of what software catches. Passing a scanner and passing scrutiny are different things.

Does a low SafeAssign score prove a paper is original? No. A low score means little matching text was found — which is equally consistent with genuinely original writing and with AI-generated text. Anthology's own documentation adds that even matches don't judge citation quality; instructors must read the report. The score is an input, not a conclusion.

Do students get to see their SafeAssign or Turnitin reports? It depends on instructor and institution settings. SafeAssign lets instructors share Originality Reports with students, and draft-check configurations exist; Turnitin's AI score is instructor-facing and often never shown to students. If you can't see a report you're being judged on, you're entitled to ask what it says.

What's the Global Reference Database? It's SafeAssign's cross-institutional pool of student papers, and it's opt-in: students choose at submission whether to contribute their paper so future submissions elsewhere can be checked against it. This contrasts with Turnitin, where storage in its repository is typically the default under institutional settings.

If SafeAssign can't see AI, is using AI in a SafeAssign course safe? Not if your course prohibits it. The violation is the unauthorized use, not the detection; instructors use secondary tools and their own judgment, Turnitin may be licensed nearby, and archived papers can be revisited. If AI use is permitted with disclosure, disclose. The tooling gap is not a permission slip.

Key facts

• SafeAssign has no AI writing detection as of this writing; Anthology's documentation describes it solely as a text-matching/originality service (Anthology help documentation).
• Anthology's stated position: the best defense against AI plagiarism is authentic assessment design, not detection tools (Anthology, "Protecting Against AI Plagiarism").
• Turnitin's AI indicator launched April 4, 2023, inside the Similarity Report; claims of 98% accuracy / <1% false positives apply only above the 20% flagged threshold, and 1–19% displays as an asterisk (Turnitin AI writing FAQ).
• Turnitin screened 200M+ papers in the indicator's first year; ~11% showed ≥20% AI writing, ~3% were ≥80% AI (Turnitin, April 2024).
• Turnitin integrates with Blackboard Learn — a Blackboard school is not necessarily a SafeAssign-only school (Anthology administrator documentation).
• Liang et al., Patterns (2023): seven detectors falsely flagged an average of 61.22% of 91 human-written TOEFL essays from non-native English speakers.
• Vanderbilt disabled Turnitin's AI indicator in August 2023, publishing its false-positive reasoning; OpenAI retired its own classifier in July 2023 at 26% detection (Vanderbilt; OpenAI).

Sources

1. Anthology / Blackboard help documentation — SafeAssign overview, matching algorithm, databases, Originality Report interpretation.
2. Anthology help documentation — "Protecting Against AI Plagiarism" (stance on AI detectors; authentic assessment; Turnitin listed among available plagiarism tool integrations).
3. Turnitin — AI writing detection FAQ and transparency documentation (launch date, accuracy conditions, asterisk policy).
4. Turnitin — first-anniversary data release, April 2024 (200M+ papers screened; prevalence figures).
5. Liang, W. et al., "GPT detectors are biased against non-native English writers," Patterns (Cell Press), 2023.
6. Vanderbilt University — announcement disabling Turnitin's AI detector, August 2023.
7. OpenAI — announcement retiring the AI Text Classifier, July 2023.
8. BestColleges — "We Tested Turnitin's New AI Detector," April 21, 2023 (Annie Chechitelli on the 85%/15% trade-off).