Most of the detectors on this blog started as plagiarism checkers and bolted on AI detection later. Pangram is different: it launched as a dedicated AI-text classifier, built by a small team with an academic-research pedigree, and its whole pitch is doing that one job unusually well. If your school, journal, or employer has started mentioning "Pangram" specifically instead of just "the AI checker," here's what it actually does and what to make of its accuracy claims.
What Pangram is and who uses it
Pangram is a standalone AI-content detection service, not a plagiarism tool. It reads a block of text and classifies it as human-written, AI-assisted, or fully AI-generated, and highlights the specific sentences that drove the verdict rather than returning one flat percentage. It's sold to universities, publishers, and businesses that want an AI check on its own, sometimes as an add-on inside a learning-management system and sometimes as a standalone check an instructor runs by hand. If your assignment goes through Turnitin instead, that platform's own indicator is covered in does Turnitin detect ChatGPT.
The headline claim: a very low false-positive rate
Pangram's core marketing angle isn't "we catch the most AI text" — it's "we rarely accuse a real human of using AI." That's a meaningfully different design goal than older detectors, which have historically leaned toward flagging more aggressively and accepting some false accusations as the cost of catching more AI text. Independent researchers, including academic teams unaffiliated with Pangram, have reported testing that found its false-positive rate stayed extremely low on longer, unedited passages, ticking up only slightly on short ones. We're deliberately not repeating a specific percentage here, for the same reason we don't quote exact numbers for any detector on this blog: a single benchmark reflects one dataset and one set of AI models, and goes stale fast. For the general mechanics, see how AI detectors work.
The claim is disputed, not settled
Here's the part worth being upfront about: a rival detector company publicly challenged the methodology behind one widely cited academic study of Pangram's accuracy, arguing the test setup queried the tool incorrectly, and published its own re-run with different numbers. Neither side is neutral — one is defending a study that favors it, the other its own competing product. That back-and-forth is actually healthy; it means the claims get scrutinized rather than accepted at face value. The honest takeaway: "lowest false-positive rate" is a real, actively contested claim, not a settled fact, and it can shift as both companies keep updating their models and tests.
Sentence-level highlighting, and its limits
One useful feature: Pangram doesn't just hand back a single score, it highlights which sentences it believes are AI-written. That's more actionable for an instructor deciding whether to have a conversation, and more informative for a writer trying to understand a flag, than a bare percentage. But highlighting a sentence as "AI-like" is still a statistical guess about that sentence's patterns, not a record of who typed it. A tidy, formal, low-variety sentence you wrote yourself can still get highlighted, for the same reasons covered in why is my writing flagged as AI.
How it fits next to detectors you've heard of
If your institution already uses GPTZero, Copyleaks, or Originality.ai, Pangram is aiming at the same job with a different design bet: fewer false accusations, at the cost of being newer and less battle-tested. None of these tools agree with each other perfectly — the same document can score differently across detectors, which is itself evidence none of them measures ground truth, only a statistical estimate of it. For the wider reliability picture, see do AI detectors actually work.
Using AI the honest way
None of this is a roadmap for beating a specific detector — that's not a game worth playing. Read your institution's or employer's actual policy on AI use first; the rules vary by school and assignment, and following them is on you. Where AI is allowed as an assistant, it's genuinely useful for brainstorming and getting past a blank page — but the work you submit should be yours: your ideas, your edits, your understanding of it. If your own honest draft reads stiff or overly uniform, our free AI humanizer can help with that specific problem — it's not a bypass tool, and it makes no promise about any detector's score, Pangram included.
The bottom line
Does Pangram detect AI writing? Yes — that's its entire purpose, and it's increasingly used directly by schools and publishers rather than as a bolt-on feature. Its headline strength is a low false-positive rate that independent academic testing has generally supported, though a competing detector has publicly disputed the exact numbers behind that claim. Treat a Pangram result the way you should treat any detector's result: a useful, imperfect signal — never a verdict, and never something worth gaming instead of just writing honestly.