01Chapter
What a detector actually measures
Not authorship. Patterns.
An AI detector is a classifier. It reads a text and estimates how closely its statistical patterns resemble the machine-generated writing it was trained on. It cannot see who typed the words, which tools were open, or how long the writing took.
The signals most detectors lean on
- Predictability
- Language models write by choosing likely next words, so their output tends to be statistically predictable. Some detectors measure this directly; it is often called perplexity.
- Uniformity
- Human writing usually varies more in sentence length and structure from one line to the next. Very even text can look machine-made. This is often called burstiness.
- Learned patterns
- Many detectors are themselves trained models. They learn whatever separates their human examples from their AI examples, so they inherit the strengths and the gaps of that training data.
What the number means
A result shown as a percentage usually describes the model’s confidence that the text resembles AI-generated writing. It does not mean that share of the text was written by AI, and it is not the probability that a particular person broke a rule. Where a tool highlights sentences, the highlights show where the model’s signal is strongest, not which sentences a machine typed.
02Chapter
Why scores disagree
Run one paragraph through three detectors and you may get three different answers. That is expected, not a malfunction.
- Different training data
- Each detector learned from different examples of human and AI writing. A text can resemble one tool’s training set and not another’s.
- Different thresholds
- Every tool chooses where to draw the line between “human” and “AI”. The same underlying signal can be reported as a pass by one tool and a flag by another.
- Length
- Short texts give a classifier little to go on. A score on a single paragraph is far noisier than a score on a full article.
- Genre
- Legal boilerplate, technical documentation, recipes, product specifications and listicles are predictable by design, whoever writes them.
- Mixed authorship
- Human drafts edited with AI, and AI drafts edited by people, are now the norm. Detectors are asked to give one answer about text with two kinds of origin.
- Versions
- Detectors are retrained and updated. The same text can score differently next month, with no change to the text.
Calibrate before you rely on a tool
Before you use any detector to make decisions, run it on texts whose origin you know: pieces your team wrote by hand and pieces you know were AI-drafted, in the genres you actually publish. Watch how it behaves on your own kind of writing, especially your human writing. That small test teaches you more about a tool than any accuracy claim on its homepage.
03Chapter
False positives, and who they hurt
A detector that is right most of the time can still be wrong about the person in front of you.
A false positive is human writing labeled as AI-generated. It is the error with the highest cost, because it lands on a real person: a freelancer who loses a client, a student facing a misconduct process, an employee whose honest work is doubted.
Writing that is more likely to be misread
- Writing by non-native English speakers, which often relies on common vocabulary and standard structures.
- Formulaic genres: legal, regulatory, technical and scientific writing, product specifications.
- Strict house styles that standardize structure and phrasing across writers.
- Short texts, where a handful of predictable sentences dominate the score.
- Heavily edited writing, where smoothing by editors or grammar tools removes natural variation.
This is documented, not hypothetical
In 2023, Stanford University researchers (Liang and colleagues, published in the journal Patterns) reported that several widely used detectors misclassified a large share of essays written by non-native English speakers as AI-generated, while judging essays by native-speaking students far more accurately. The same year, OpenAI withdrew its own AI text classifier, citing its low rate of accuracy.
Why small error rates become big numbers
Take a purely hypothetical detector that wrongly flags 1 in every 100 human-written texts. That sounds accurate. Now imagine an agency checking 2,000 human-written articles a year with it. On average, about 20 honest pieces would be flagged as AI, each one a potential dispute with a writer or a client. At volume, even a low error rate produces a steady stream of wrong answers.
04Chapter
What a clean score does not prove
Detectors fail in both directions, so a pass certifies nothing either.
A false negative is AI-generated text labeled as human. Light editing, unusual topics and newer models can all lower a detector’s confidence. That matters for anyone tempted to treat a low score as a certificate of authorship: it is not one.
The useful conclusion is not that detectors are worthless. It is that they are one signal among several, strongest when combined with the things a score cannot see: drafts, notes, version history, sources and a conversation with the writer.
It also points to a better question. “Was this written by AI?” is often unanswerable from the text alone. “Is this accurate, useful and right for the reader?” can always be answered, and it is the question your audience actually cares about.
05Chapter
A fair review protocol
What to do, in order, when a piece of writing is flagged.
Check the conditions
How long is the text? What genre is it? Is the writer working in a second language? Is it built on a template? Discount scores on short, formulaic or templated writing.
Read the highlights, not the headline
If the tool marks passages, look at which ones. Definitions, boilerplate, lists and standard phrasing are flagged often and prove little.
Review it as an editor would
Does the piece have the problems that actually matter: vagueness, filler, unsupported claims, factual errors? Those are worth fixing whoever wrote the text.
Ask about process, not for a confession
Drafts, outlines, notes, sources and version history show how a piece was made. Ask for them the way you would ask any writer about their research.
Get a second reading if you need one
Another reviewer, or another tool, can help. Remember that two tools are two different measurements, not a vote.
Decide on the writing, and record why
If the piece is accurate and meets the brief, a score alone should not sink it. If it falls short, revise it, whoever wrote it. Write down the decision and the reasons.
06Chapter
Talking to clients about detection
For agencies: set expectations in writing, before a score sets them for you.
Most detection disputes between agencies and clients start with an expectation nobody wrote down. Settle it at the start of the engagement, in plain words.
Put in writing
- Which detector the client uses, if any, and why.
- That detector scores are probability estimates that vary between tools and over time.
- What you deliver: accurate, reviewed, human-quality writing based on the approved brief or draft.
- What happens if a piece is flagged: a review using an agreed protocol, and a revision if the writing needs one.
Never promise
- A specific score on a specific tool.
- “100% human” results.
- That a passing score today will still pass after the next detector update.
- That AI assistance needs no disclosure where the client’s rules say it does.
Wording you can adapt
We deliver original, fact-checked writing reviewed by our editors. AI detection tools produce probability estimates that vary between tools and over time, so we do not guarantee specific detector scores. If a piece is flagged, we will review it with you against an agreed checklist and revise it wherever the writing can be improved.
07Chapter
Where HumanBlend fits
A read-back before you publish, and a mode built for the strictest reviews.
- AI Detector
- Read a draft the way a detector does, before an editor or a client does. Use it as an inspection step: find the passages that still read as machine-made, then edit them. Open the AI Detector.
- Ultra
- The humanization mode HumanBlend builds and positions against Pangram 4.1, at 3.5 credits per input word. Standard is not positioned against Pangram 4.1.
- Repair
- A rewrite that reads as human but says something different is worse than no rewrite. Repair compares the result with your original for quality and intent.
08Chapter
Questions people ask
01Can any detector prove that a text was written by AI?
No. Detectors estimate how closely a text resembles AI-generated writing. Evidence of how a piece was made comes from process: drafts, notes and version history, not a score.
02Why was my own writing flagged?
Short, formulaic or highly polished writing can look statistically predictable. It is a known limitation of detection, and it is why a score should start a review rather than end one.
03Does a low AI score prove a person wrote it?
No. Detectors miss in both directions. A low score is one signal, not a certificate.
04Should I add typos or awkward phrasing to look more human?
No. Deliberate errors make writing worse, and readers and clients notice them long before a detector does.
05Which detector is the most accurate?
We do not publish rankings of detectors. Accuracy depends on the genre, the length of the text and the version of the tool. Test any detector on writing whose origin you know, in the genres you publish, before relying on it.
06Is it wrong to use AI to help write?
That depends on the rules of your client, publication, platform or institution. Where disclosure is required, disclose. Where it is not, the standard that matters is the one your readers apply: is it accurate, useful and honest?