PangramFor teachers
Pangram's AI Assisted Label Marks Machine-Edited Sentences, Not a Verdict
Pangram's AI-Assisted label marks a stretch of text, not the whole paper. The model card tags every segment Human, AI-Assisted or AI-Generated, and the document verdict is arithmetic on characters: Human at 90 percent human or more, AI at 80 percent AI-generated or more, Mixed in between. Pangram's 3.0 release (December 2025) split assistance in two: light, meaning grammar and phrasing edits, and moderate, where AI may have rewritten or added passages. So a paper reading 30 percent AI is Mixed. What that costs a student is the school's call.
Olive Staff & OliveUpdated
What Does Pangram's AI Assisted Label Mark?
AI-Assisted is a label the model puts on a stretch of text, never on the student. Pangram 4 reads a document at the token level and decodes one of three labels for each segment: Human, AI-Assisted or AI-Generated 1. An assisted segment is one whose wording reads to the model as machine-edited. The document verdict comes afterwards, from a count of characters.
Pangram publishes that count. It calls a document Human when at least 90 percent of its characters are classified as human, AI when at least 80 percent are classified as AI-generated, and Mixed for everything in between 1.
So a paper reading 30 percent AI is Mixed by definition. The Mixed box runs from just over 10 percent non-human characters to just under 80 percent AI-generated, wide enough to hold very different papers. One has two polished paragraphs. Another is half machine-drafted.
Pangram 3.0, released on 11 December 2025, introduced a four-way scale: fully human-written, lightly AI-assisted, moderately AI-assisted and fully AI-generated 2. The same announcement defines the two middle bands. Light assistance covers spelling and grammar fixes and updated phrasing, plus translation and readability changes. Moderate assistance covers changes where AI may have rewritten significant portions of the text or added content of its own 2. Moderately AI-assisted is the reading that sends a teacher looking for a policy. Nothing on that scale names a grade.
Who Sets the Line a Mixed Reading Has to Cross?
The school does, where it has set one at all. Neither Pangram's model card nor its page for teachers names a cutoff at which a percentage becomes a penalty 13. Its page for teachers calls the product a decision-support tool rather than a disciplinary weapon, and points a grader at the sentence-level highlighting instead 3. The percentage is evidence about text. What that evidence licenses is a policy decision.
A syllabus that names no number leaves the grader to choose one. What percentage of AI is acceptable is a policy question, and declining to answer it with one figure is defensible.
Human, tuned for college-level academic writing, publishes where its own detector raises a flag, which is a different line from a penalty. By Human's own in-house measurement as of September 2026, not an outside audit, accusation-safe flags a paper above 15% machine content, standard above 6% (the default), sensitive above 2%. "This is an estimate from our detector. A flag should be reason to look closer and not verified proof of cheating." No setting turns an AI verdict into a Human one; an AI verdict needs at least 80% of the document to read as machine-written. A reader setting policy finds the measurements behind those points on Human's research page.
Each product publishes an 80 percent line for its AI verdict, counted in its own way: Pangram on characters classified AI-generated 1, Human on the share of the document that reads as machine-written. Neither draws the line where a paper loses marks. That line is an academic judgment, and neither company is in the room when it's made.
Read the Highlighted Sentences Before the Percentage
The highlighting is the part of the result that can be checked against the paper. Pangram's teachers page says a mixed result means some sections appear AI-generated while others are human-written. It tells a grader to use the sentence-level highlighting to see exactly which parts were flagged 3. Reading those sentences takes a couple of minutes, and it changes what the conversation afterwards can be about.
Three things about that view are worth knowing before it's quoted at a student.
- The High, Medium or Low confidence label beside a segment measures the peakedness of the model's posterior and is not a calibrated probability, by the model card's own statement 1.
- A paragraph judged human on its own can switch to AI once it sits inside a larger segment, which is why a paragraph rechecked on its own can read differently 4.
- A 100 percent AI reading does not mean every word was machine-written, according to Max Spero, a Pangram co-founder 4.
None of that shows up in the percentage.
Pangram also lists writing that is outside its model's primary scope or more prone to error: source code, tables of contents and reference sections, along with templated writing and text dominated by mathematical notation 1. Its September 2026 post on false positives also advises against screening short bullet lists, outlines and single-sentence answers 5. So a reading on a methods section thick with equations deserves less weight.
One more piece of the vendor's own advice gets skipped a lot. Strip the headers, footers and assignment instructions out of a document before checking it 1. A prompt pasted at the top of a file is exactly that material: text the student did not write 1. A reference list is the same trap in a different place: formatted text carried over from a style guide, which the model card already puts outside the model's primary scope 1. Neither belongs in the count a teacher acts on.
Which Innocent Readings Fit a Mixed Result?
Several, and Pangram's own figures describe the first of them. At the Pangram 3.0 launch in December 2025 the company published rates of about 1 human-written essay in 700 coming back lightly AI-assisted and about 1 in 7,500 coming back moderately AI-assisted 2. If those launch-day rates held on a course's human-written papers, which is an assumption (Pangram names no test set for them and they describe the 3.0 model), 120 such papers would give an expected 0.17 lightly assisted readings.
At the 3.0 launch the free checker rounded the scale off. The Pangram 3.0 announcement says free users see lightly AI-assisted text as Human and moderately AI-assisted text as AI-generated 2. If that still holds, a teacher on the free tier and a colleague on a paid seat could report different labels for the same paper. The piece on what the free Pangram checker shows covers where the rounding happens.
Editing tools can sit in the same band. A draft put through a grammar assistant or a translation round trip has been machine-edited in the plain sense of the phrase, and so has one run through a phone keyboard's rewrite button. Pangram's light band is defined to cover edits of that sort 2.
A student who ran one paragraph through a rewrite button and a student who had a model rewrite most of the essay can both land in the middle of the scale. On its own the label settles nothing.
Boilerplate is the quiet one. A rubric pasted into the file or a department's lab template is the kind of templated writing the model card lists outside the model's primary scope, and the card asks for extraneous material to be removed before checking 1. The student composed none of that text.
The label describes the words on the page. It names no tool and no session, and it says nothing about intent.
Open the Conversation With the Draft, Not the Score
A mixed result is a reason to ask about the writing, and the strongest questions are about how the paper got written. Pangram's chief executive told The Atlantic in May 2026 that the tool should never be the ending arbiter, and should instead be a starting point for a more thorough investigation 6. A grader who opens with the percentage has already made the percentage the subject. The draft is the subject, and the score is one witness to it.
Ask what the student read. Where did the second argument come from? How did the third paragraph get rewritten? Then look at the version history together, and have the student walk through the draft. What they can and can't show belongs in the note. The number doesn't.
At least one system-wide policy requires that meeting whenever feasible. The City University of New York's Academic Integrity Policy, revised June 2024, says a faculty member who suspects a violation shall review the facts and circumstances with the student whenever feasible, before a report is filed 7. Where local policy says the same, a note of what was said belongs in the file.
The mechanics are set out in how to talk to a student about suspected AI use, and the order of the first hour in what to do when a paper is flagged for AI.
Whatever the reading says, a grade change needs evidence a person can point to.
Common questions
Does AI-Assisted mean the student used ChatGPT?
No. AI-Assisted is a label on a segment of text, and it reports that the wording reads to the model as machine-edited 1. It doesn't name a product or an account, or say when anything happened. A grammar assistant and a chatbot rewrite of the student's own paragraph can land in the same box, and so can a translation pass. Pangram's chief executive has described the model's inner workings as pretty uninterpretable 6.
What percentage of AI is too much on a paper?
Neither Pangram's model card nor its page for teachers names such a number 13, and a detector vendor isn't the party that could set one for somebody else's course. The percentage reports how much of a document reads as machine-written. The consequence attached to it comes from a syllabus, a department or an integrity policy. A course that wants a threshold has to write one down in advance, in language a student can read before submitting and an appeal panel can read afterwards.
Why did the same paper read differently on the free checker?
One documented reason: at the Pangram 3.0 launch in December 2025, the announcement said free users saw lightly AI-assisted text as Human and moderately AI-assisted text as AI-generated 2, so the middle of the scale disappeared in both directions. If that still holds, two readers looking at one essay on different plans can honestly report different labels. Before quoting a result to a student, note which tier produced it and when.
Is a mixed result enough to open an academic integrity case?
On its own, no. Pangram's own model card acknowledges a non-zero error rate and the serious harm a false accusation causes 1. Its teachers page presents the product as decision support rather than a disciplinary instrument 3. A case rests on what a person can show: drafts, version history, the student's account of their own process, and the record of a conversation held before anything was filed.
Does a short assignment give a reliable reading?
Not the shortest ones. The model card sets a floor of 50 words in complete sentences for the writing it's built to read 1. The company's false-positive post also advises against screening single-sentence answers, outlines and short bullet lists 5. A discussion-board post of a sentence or two falls squarely in that group 5.
References
- 1.Pangram 4 Model Card Pangram Labs, 2026. pangram.comThe three segment labels, the character arithmetic behind the Human, AI and Mixed document verdicts, the statement that the confidence label is not a calibrated probability, the 50-word minimum, the writing listed as outside the model's scope, the advice to strip headers and instructions first, and the acknowledgement of non-zero error and the harm of false accusations.
- 2.Introducing Pangram 3.0 with AI assistance detection Pangram Labs, 2025. pangram.comThe four-way scale released on 11 December 2025, its definitions of light and moderate assistance, the rates published at that launch for human-written essays landing in each band, and the statement that free users see lightly assisted text as Human and moderately assisted text as AI-generated.
- 3.AI Detector for Teachers: Free AI Checker Pangram Labs, 2026. pangram.comThe page's description of a mixed result, its instruction to use sentence-level highlighting to see which parts were flagged, and its framing of AI detection as a decision-support tool rather than a disciplinary weapon.
- 4.AI-detection tools have made huge leaps forward — how good are they? Nature (Naddaf and Van Noorden), 2026. nature.comMax Spero, a Pangram co-founder, on why a 100 percent AI reading does not mean every word was machine-written, and on how chunking lets a paragraph judged human alone switch to AI inside a larger segment.
- 5.All About False Positives in AI Detectors Pangram Labs, 2026. pangram.comThe September 2026 post's list of writing the company advises against screening, including short bullet lists, outlines and single-sentence answers.
- 6.America Has a Pangram Problem The Atlantic (Matteo Wong), 2026. theatlantic.comMax Spero's statement that Pangram should never be the ending arbiter but a starting point for a more thorough investigation, and his description of the model as pretty uninterpretable.
- 7.Academic Integrity Policy The City University of New York, 2024. cuny.eduThe requirement that a faculty member who suspects a violation review the facts and circumstances with the student wherever feasible before filing a report.
7 sources, numbered by first appearance.
General guidance for teachers, administrators and students. What holds at one institution, on one assignment, may not transfer to another.
Human checks whether a piece of writing can be verified as written by a person, and shows the sentences behind the result. It does not report a probability that a person used AI, it does not check for plagiarism, and no number it produces stands for a student's honesty. A flag should be reason to look closer and not verified proof of cheating.