Does Packback Detect AI? (What It Actually Checks)

The short answer

Yes. Packback says its platform detects AI-generated content from any commercially available model — it names ChatGPT, Claude and Gemini specifically — and it checks your writing before you submit rather than only after. On its own FAQ Packback states it tuned its model to a 0.005% false positive rate, and its Originality product page claims "the lowest false positive rate of any commercially available tool (<0.01%)."

Two things that answer is missing, and they matter more than the number. First, that figure is a vendor claim — there is no published methodology or independent audit behind it, and a false-positive rate on its own is only half of an accuracy picture. Second, the detector is not the only thing watching: Packback also builds an originality report and a record of how a piece of writing came together. This page walks through all of it — what Packback says, what independent research says about AI detectors as a category, what happens procedurally when something is flagged, and what to do if you're flagged and you didn't use AI. It is not a guide to getting past a detector, and you should be suspicious of any page that offers one.

What Packback says its AI detection does

Packback isn't a plain message board with a plagiarism scanner bolted on. Every post runs through an automated feedback layer as you write it, and integrity checking is part of that same layer rather than a separate step at the end.

Reading Packback's own materials, here is what the company claims:

• It targets commercial models generally, not one product. The Originality page says it "detects AI generated content from any commercially available AI model, including ChatGPT, Claude, Gemini, etc." So "does Packback detect ChatGPT" and "does Packback detect AI" have the same answer — the claim isn't model-specific.
• It warns before submission, not just after. Packback's FAQ says it incorporates "proactive flagging for students before they submit work on both standard plagiarism detection and AI-generated text," and that if you copy-paste such content in, "we will both warn and coach them before submission." You are meant to see a preview of the report while you can still do something about it.
• It also does conventional similarity checking. Packback describes comparing submissions against a live web search, a repository of "over 100M saved web-based documents," and "over 50M pieces of student-submitted content."
• The published accuracy claim is a false-positive rate. 0.005% on the FAQ — Packback frames this as roughly 1 in 20,000 submissions at risk of being incorrectly identified — and "<0.01%" on the product page, with a footnote comparing it to a 1% rate it attributes to other academic-integrity vendors.

How much should you trust that 0.005% number?

This is where most articles on this topic stop, and it's where the useful part starts. Take the claim seriously — and then notice what it doesn't say.

A false-positive rate alone is not accuracy. Any classifier can drive its false-positive rate to nearly zero by simply flagging almost nothing. The number you'd need alongside it is the false-negative rate — how much real AI writing sails through — and Packback does not publish one. A tool tuned to almost never accuse an innocent student is a defensible design choice, arguably the right one, but it is a choice that trades away sensitivity, and you can't evaluate one half of that trade in isolation.

It is a vendor claim, not an independent finding. Packback has not published a peer-reviewed evaluation, a test corpus, or a third-party audit that anyone outside the company can check. Compare it to what it's being compared against: Vanderbilt University's teaching-technology group put Turnitin's own 1% figure in context by noting the university had submitted roughly 75,000 papers in a year, which at 1% would mean about 750 students wrongly flagged. Vanderbilt's conclusion was blunt — "we do not believe that AI detection software is an effective tool that should be used" — and they disabled the detector rather than rely on the vendor's number.

The number can be true and still not describe your situation. A rate measured across all submissions says nothing about which submissions carry the risk. Detection research (below) consistently finds the errors are not evenly distributed — they cluster on writing that is plain, structurally regular, or produced by someone writing in a second language.

None of that makes Packback's claim false. It makes it unverifiable from the outside, which is a different thing, and it's the honest position to hold.

The detector is not the only thing Packback records

Here's the part almost nobody searching "packback ai detector" knows about, and it's arguably more consequential than the classifier.

Packback's Insights and Originality materials describe two things beyond a simple flag. There's an "Originality Fingerprint" report, which students can preview before submitting and instructors see inside the normal grading workflow — including a side-by-side "Match Comparison" showing additions, deletions and replacements, plus sentence-level indications of possible AI use. And there's a "Writing Process Report," which Packback describes as "a transparent, comprehensive trail of engagement across a student's entire writing lifecycle," built on Zimmerman's self-regulated learning framework.

Read that plainly: the platform records how the writing happened, not only what it looks like when finished. A post that appears in the box fully formed in one paste has a different process signature from one that was drafted, revised, and cited over twenty minutes — and no classifier confidence score is involved in noticing that. If you are weighing risk, this is the signal to think about, not the detector.

The flip side is worth saying out loud too, because it's genuinely good news for honest students: a process record is evidence that can work in your favour. If you actually wrote the post, the trail showing you wrote it is on the same screen your instructor is looking at.

What independent research says about AI detectors generally

Two published studies are worth knowing, and it's important to be precise about scope: neither tested Packback. They tested other AI-text detectors. What generalises is not a verdict on Packback's accuracy but an understanding of what this class of tool is actually measuring.

Weber-Wulff et al. (2023), published in the International Journal for Educational Integrity, tested 14 detection systems — 12 public tools plus Turnitin and PlagiarismCheck. Their conclusion: "the available detection tools are neither accurate nor reliable and have a main bias towards classifying the output as human-written rather than detecting AI-generated text." They also found that content obfuscation techniques "significantly worsen the performance of tools."

Liang et al. (2023), published in Patterns, ran seven widely-used detectors over 91 human-written TOEFL essays and 88 US eighth-grade essays. The detectors were near-perfect on the eighth-grade essays. On the TOEFL essays — written by non-native English speakers, with no AI involved at all — the average false-positive rate was 61.22%. All seven detectors unanimously flagged 18 of the 91 essays, and 89 of the 91 were flagged by at least one.

The mechanism behind that gap is the useful takeaway. These detectors respond to how statistically predictable a piece of text is, sentence by sentence — not to who typed it. Plain, regular, carefully-constructed prose looks predictable. So does AI output. That is why a careful multilingual student and a language model can land in the same bucket, and why a flag is a statement about the shape of your sentences, not a finding about your conduct.

What actually happens if a post is flagged

This is the consequence question underneath every version of this search, and the answer has two halves.

Before you submit: by Packback's account, this is where most flags are meant to land. You get a warning and coaching, you revise, and nothing ever reaches your instructor. That design choice is real and it's the most student-friendly thing about the platform — it treats a flag as a prompt rather than a verdict.

After you submit: what happens is not decided by the software. Packback surfaces the report to your instructor; your instructor decides what to do with it under your course's and your institution's academic-integrity policy. That can mean nothing at all, a conversation, a request to revise, a grade penalty, or a formal integrity referral, and the range is wide because institutions genuinely differ. A flag on its own is not a failing grade and is not a finding of misconduct.

It's worth knowing that even detector vendors say their score shouldn't stand alone. Vanderbilt's write-up notes Turnitin launched its detector claiming a 1% false-positive rate while giving limited transparency about how it worked — and Vanderbilt turned it off anyway. If your institution treats an automated score as sufficient evidence by itself, that is a policy problem worth naming.

If you're flagged and you didn't use AI

False positives are real, and the research above shows exactly who absorbs most of them: students who write plainly, students working from a tight outline, and students writing in a second language. If it happens to you, the situation is recoverable. What you need is a record, not an argument.

• Preserve your process, ideally before you ever need it. Draft in a document with version history turned on. Google Docs and Microsoft Word both keep one automatically. That history is the single most persuasive thing you can produce, and it can't be reconstructed after the fact.
• Keep the sources you actually read. Open tabs, library records, a note with page numbers. Someone who read the material can talk about it; that conversation usually ends the matter.
• Ask what the evidence is, specifically. "Which sentences, and what score?" is a fair question. So is asking whether the finding rests on the automated report alone.
• Read your institution's procedure, not just the syllabus. Most schools have a written academic-integrity process with defined steps, a right to respond, and often an appeal. Knowing it exists changes the conversation.
• Offer to talk it through. Explaining your own argument out loud is the oldest and best test there is, and it's one an instructor can apply immediately.

What not to do: don't rewrite the post to "look less like AI." It doesn't address the accusation, it destroys the timestamps that would have helped you, and — per the Liang findings — the stylistic changes people reach for are exactly the ones these tools respond to unpredictably.

Is using AI on Packback against academic integrity?

It depends on your course, and that isn't a dodge — the variation is genuinely enormous. Packback's own published stance is that not all generative-AI use is a bad thing, provided there is transparency, attribution and citation. But Packback doesn't set your rules. Your instructor does, and some permit AI for brainstorming with disclosure, some ban it for discussion work entirely, and many haven't written anything down, which is its own trap.

The International Center for Academic Integrity frames integrity as "a commitment, even in the face of adversity, to six fundamental values: honesty, trust, fairness, respect, responsibility, and courage." Note that none of those name a technology. They describe how you conduct yourself, which is why the workable rule is behavioural rather than tool-based:

• Use AI only where your course policy permits it, and disclose when the policy asks you to.
• Read both documents — your syllabus sets course rules, your institution's honor code sets institutional ones, and they are not the same document.
• When it isn't written down, ask, in writing, before you submit.
• Never submit something you couldn't explain out loud.

We go deeper on where that line sits in is using an AI homework bot cheating? — and if what you actually want is help getting a post written, what to do when you want someone to write your discussion post covers the honest options and the risky ones.

Writing Packback posts that hold up

The reassuring thing about Packback specifically is that the behaviour which earns a strong Curiosity Score and the behaviour which produces a clean originality report are the same behaviour. You don't have to optimise for two audiences.

• Read the source before you open the post box. Specific references — a line, a figure, a claim you're reacting to — are what graders reward, and they only come from having done the reading.
• Lead with a question you actually want answered. Packback's entire model is built around open-ended inquiry; a sincere question outperforms a polished summary.
• Cite something real and link it. Credibility is a scored dimension, and a live link is the cheapest point on the board.
• Draft somewhere with version history. Not for the detector — for you. It's how you keep a record, and it's how you stop losing work.
• Write in your own register. A concrete example, a position you'd defend, your actual phrasing. That's what makes the post yours in the only sense that matters.

For structure, sourcing and tone from a blank page, see how to write a discussion post.

Where Silent Student fits

Full disclosure, because this is our product: Silent Student is a desktop app for macOS and Windows — not a browser extension and not an answer site. It connects to your own courses, sorts the work by due date, and can write Packback questions and replies in a source-grounded style so you aren't staring at an empty box at 11pm.

One thing to be straight about, because it's the difference that matters on this page: Packback posts are not held for your approval. Silent Student's Draft Review Mode covers the essays, discussion posts and file uploads you submit in Canvas or Brightspace — those wait in your dashboard until you read and approve them. Packback is external courseware reached through your LMS, and the solver writes and posts there directly. If you want to review or revise a Packback post, you do it in Packback, and given everything above about the writing-process record, doing exactly that is the sensible habit.

What we will not tell you is that any of this produces a particular result from Packback's detector, or from any detector. Nobody can honestly promise that — a classifier's output isn't something an outside tool controls, and as the section above shows, the process record sitting next to it isn't either. Whether using it fits your course is a decision only you and your syllabus can make. See how it works, or the FAQ.