←  Artificial MindsThe fourth project · September 2026

Did AI Write This?

Two passages, and an AI model wrote one of them. Can you tell which? Then see what AI detectors, watermarks and editing histories can and cannot show.

Guess which of these two passages was written by a machine

How well people guess often depends on how much they use these AI models

In a 2021 study, researchers gave a mix of short stories, news articles and recipes to readers with no training in spotting AI writing. People had written some of those pieces, and OpenAI GPT-3 had written the rest. The readers were right about half the time, the same as guessing. After a short training session, some readers' accuracy rose to about 55%.

In a 2025 study, the researchers collected 300 long articles in English, all of them nonfiction. People wrote some of the articles, and three models from OpenAI and Anthropic wrote the rest. The readers in this test used those three models every day. A group of five of these readers voted on each article, and the group's answer was right 299 times out of 300.

The two tests used different writing, different lengths and different readers, so their results cannot be compared directly. Putting the results side by side does not show that people got better at this between 2021 and 2025. In the 2021 test, untrained readers did no better than guessing on short pieces. In the 2025 test, the readers used AI models every day. They did far better than guessing on long articles of a kind they knew well.

The readers who did well in 2025 were asked what they had been looking for. Part of their answer was a list of words and phrases that AI models use too often. The rest was harder for them to put into words. They said they could tell whether the writer cared about the subject. They also recognised a formality in the models' writing that they could not fully explain.

The idea behind AI detectors is nearly 140 years old

Neither of those conditions holds when the suspect is an AI model. There is no short list of suspects. Every version of every model writes differently from the last. A model asked to write in someone else's style will do it. The idea of counting habits a writer cannot see survived anyway. Every AI detector today is descended from it.

There are three broad ways to check whether an AI model wrote a text.

1. AI detectors. Programs that read the finished text and score how AI-written it looks.
   • 1a. An AI trained to mimic another AI. It runs a stand-in model and checks how predictable each word was. Free research tools such as Binoculars work this way.
   • 1b. An AI trained to catch another AI. It learned the difference from millions of matched human and AI documents. Pangram works this way.
2. Watermarks. Hidden patterns planted in the text by the AI model itself.
3. Editing history. The writing software's log of how the text got onto the page.

None of the three works well on a paragraph that a person and an AI model wrote together. AI detectors also face humanizers. These services rewrite AI text until an AI detector scores it as human-written.

1AI detectorsPrograms that read the finished text and score how AI-written it looks.

All an AI detector has is the words of the text. There are two designs. The first scores how predictable each word would be to an AI model. The second is trained on millions of examples of AI and human writing.

1aAn AI trained to mimic another AI

Inside the first design sits an AI model of its own, used as a simulator. The simulator plays the part of the AI model that might have written the text. The AI detector feeds the text to the simulator one word at a time. At each point the simulator lists the words most likely to come next. The AI detector then checks where the real next word sits on that list. A word near the top of the list is easy to guess. A word far down is a surprise. An AI model writes by picking a likely next word, again and again. Its sentences are built from words the simulator puts near the top. A person surprises the simulator more often.

The simulator made one guess in each of two sentences, and the results sit side by side.

one guess from each of two sentences

The weather today was cold and grey, so we stayed inside.

the simulator's expectations for the marked word

grey		31%
overcast		22%
wet		14%
dark		9%
damp		6%
sugary		0.01%

A match. "Grey" was the simulator's first choice. On its own this word is a small sign that an AI model wrote the sentence. A person could easily have written that sentence too.

The weather sulked all morning, so Ruth stayed in.

the simulator's expectations for the marked word

was		40%
today		18%
is		12%
has		6%
here		3%
sulked		0.001%

A surprise. "Sulked" sat thousands of places down the simulator's list. On its own this word is a sign that a person wrote the sentence.

The simulator has to resemble the AI model that might have written the text. An AI detector tested against one AI model does worse on text from an AI model it has not seen. Every new version of an AI model changes its habits. An AI detector tuned last year checks for habits that may already be gone.

1bAn AI trained to catch another AI

The second design is trained on examples. Someone collects millions of documents that people wrote and has AI models write a twin of each one, on the same subject and at the same length. A program is then trained to tell each document from its twin. Because both documents cover the same subject, the program has to look for differences in how they are written.

The program is then run over a large pile of human writing from before 2022. Every human passage it wrongly flags goes back into the training pile, and the program is trained again on its own mistakes. Pangram makes one of these programs and says this loop cuts its rate of wrong flags by a factor of between 100 and 1,000. The figure is the company's own.

Mendenhall counted one habit by hand, for two writers he had already picked out. This program counts thousands of habits at once.

The known limits of AI detectors were found in the first design. The second design was built to defeat those limits.

Run Pangram on both passages

Each passage from the quiz has an empty dial. Run Pangram and the dial fills with its score.

With scripts off the dials show Pangram’s scores already.

Passage A · the AI model

Rain hammered the tent until each drop seemed to strike inside my skull. I sat with my knees drawn up, watching the thin walls shudder and bow beneath the wind. Water gathered along a seam, swelled into bright beads, then fell onto the corner of my sleeping bag. The air smelled of wet nylon, mud, and the wool of my socks. Whenever lightning flashed, the scattered boots and cooking pots appeared stark against the fabric, then vanished into gray. Thunder rolled through the ground beneath me. I pressed my palms around a cooling mug and counted the spaces between gusts. Gradually, the canvas stopped snapping. The rain softened to a steady patter, and somewhere beyond the entrance, water trickled over stones. I loosened my grip and stretched my cramped legs.

130 words

Run Pangram

AI Generated

Passage B · Cherry-Garrard, 1922

One summer blizzard is much like another. The temperature, never very low, rises, and you are not cold in the tent. Sometimes a blizzard is a very welcome rest: after weeks of hard pulling, dragging yourself awake each morning, feeling as though you had only just gone to sleep, with the mental strain perhaps which work among crevasses entails, it is most pleasant to be put to bed for two or three days. You may sleep dreamlessly nearly all the time, rousing out for meals, or waking occasionally to hear from the soft warmth of your reindeer bag the deep boom of the tent flapping in the wind, or drowsily you may visit other parts of the world, while the drifting snow purrs against the green tent at your head.

130 words

Run Pangram

Human Written

Pangram was right on both passages. Passage B has been on Project Gutenberg for years and may have been in Pangram’s training data. A correct result tells us less if the detector has already seen the passage.

What an AI detector cannot do

1. Guarantee that human writing never gets flagged. In 2023 several AI detectors flagged the opening of the United States Constitution. Edward Tian built the AI detector GPTZero. He explained that AI models learn to write by reading an enormous pile of text, and the Constitution sits in that pile. An AI detector's simulator was therefore trained on the Constitution and would have picked much the same words. Pangram keeps a test on every release. The test blocks the release if the Declaration of Independence scores as AI-written.

   A person writing plainly in a second language can be flagged as well. Plain, common words are the easiest words to guess, and a writer working in a second language tends to reach for them. In 2023 researchers ran seven AI detectors over 91 essays written for TOEFL. TOEFL is an English test taken by people whose first language is not English. The AI detectors flagged 61% of those essays. The same AI detectors flagged fewer than 10% of essays by American eighth-graders, who are about fourteen years old. When an AI detector goes wrong more often for one group of writers than for another, the pattern is called a bias.

   A 2026 study asked whether the same bias appeared in Czech. The researchers looked at essays written in Czech, some by people who grew up speaking the language and some by people who did not. They found no such bias in the AI detectors they tested. In another 2026 study, researchers tested 16 AI detectors on English student essays. They found the bias in some of the AI detectors and none in the others. So the bias depends on the AI detector and on the language. A single flag on a second-language writer's essay is weak evidence. A teacher would need to ask which AI detector raised the flag and which essays that AI detector had been tested on.

   Both failures came from AI detectors of the first design. Makers of the second design report no such bias in their own testing. Nobody outside those companies has checked the claim.

2. Give the chance that a flagged text was written by a person. An AI detector can be wrong in two ways. It can miss an AI-written text (a false negative), and it can flag a human-written text (a false positive). Each error rate describes the detector’s performance on a particular set of texts. Real rates differ by AI detector, by language and by the length of the text. Take a made-up AI detector that misses 10% of AI-written texts and flags 1% of human-written ones, and put it to work on the essays of a school.

   sample rates for a made-up AI detector

   	flagged	not flagged
   AI-written	90% caught	10% missed: the false negative rate
   student-written	1% wrongly flagged: the false positive rate	99% cleared

   A teacher holding a flagged essay wants to know the chance that this essay was student-written. The false positive rate cannot give that chance. The chance depends on how many students cheated. Say 1 student in 100 cheats. Out of 10,000 essays, the AI detector flags 90 of the 100 AI-written ones and 99 of the 9,900 student-written ones. Just over half of those flags are wrong. If half the students cheat, the same AI detector's flags are nearly all right. The chart shows every share from none to 1 student in 5.

   chance that a flagged essay was student-written

   Nobody has published a count of how many essays in a school are AI-written. The nearest figure comes from Turnitin. Many universities use Turnitin to check student work for plagiarism and AI writing. In May 2023, Turnitin reported scanning 38.5 million papers. In 3.5% of those papers, its AI detector flagged at least 80% of the words as AI-written. That count comes from the AI detector itself, so the true share of AI-written papers could be higher or lower. If the true share were 3.5%, about 1 flag in every 4 from the made-up AI detector would land on a student-written essay.

   Vanderbilt University sent 75,000 papers to Turnitin in 2022. Turnitin says its AI detector wrongly flags fewer than 1 student-written document in 100. One in a hundred of 75,000 is 750, so up to 750 student-written papers a year could be wrongly labelled. Vanderbilt published that calculation and turned the AI detector off in August 2023.

3. Turn a percentage score into the chance that a student cheated. Turnitin's percentage is the share of the words in one document that its AI detector scored as AI-written. The percentage counts words and carries no estimate of the chance that the student cheated. Turnitin says the number is less reliable below 20%. It also says its AI detector does not score documents shorter than 300 words.

   Turnitin scores have started to turn up in court. In January 2026 a university in New York found a first-year student guilty of cheating. The university built its case on a Turnitin score of 100%. A New York court threw the finding out. The court said the evidence was thin. The court also found the hearing unfair: the university would not let anyone speak for the student, and the same official who found him guilty also decided his appeal. The court left open whether the score itself was right.

2WatermarksHidden patterns planted in the text by the AI model itself.

An AI model gives every word it could write next a probability. Suppose it has written "The weather today was cold and". It might put "grey" at 31%, "overcast" at 22%, "wet" at 14%, "dark" at 9% and "damp" at 6%, with the rest of the 100% spread thinly over every other word it could write. It then picks one at random in proportion to those numbers, so a word at 30% comes out about 3 times in 10. Where several words sit close together, the pick is nearly a coin toss. A reader will not care which of those words appears. A watermark changes the odds in picks like those.

How the watermark is planted

The company that makes the AI model keeps a secret number called a key. Before choosing each next word, the model uses the key and the preceding text to work out a list of favoured words, called the green list. The probabilities of the words on that list go up a little and the probabilities of all other words go down a little. The model still picks at random, so any word can come out. Over a page of writing, though, the favoured words come out more often than they otherwise would. Each word has different words in front of it, so the green list is worked out again at every step.

The key steers a pick when the model has a real choice between words. In a sentence with one obvious next word, like "the capital of France is Paris", there is nothing for the key to steer.

the sentence so far

How a checker finds watermarks

Later the company can run a checker over the text. The checker holds the same key. At each word, the checker uses the key and the preceding text to recreate the green list, then checks whether that word was on the list. In text without a watermark, nothing steered the picks, so about half the words land on the green list by chance. In watermarked text, far more words land on it.

The checker needs a good stretch of text before its count is reliable. In one case worked out in the green-list paper, a checker needs about 135 words to report a watermark that puts 65 of every 100 choices on the green list.

two passages, checked with the same key

Two real designs

The green list came from John Kirchenbauer and colleagues at the University of Maryland, who published it in January 2023. It is the simpler of the two. The other design, SynthID-Text, came from Google DeepMind and was published in the science journal Nature in October 2024. It works like a tournament. The model produces several possible next words and uses the key to pick the winner of each pair of words, until one word is left to write. Google was already using this design in Gemini when the Nature paper came out. Anthropic announced on 14 August 2026 that new Claude models would carry a version of this design, with older models to follow in the months after. Every candidate word in the tournament is one the model could have written anyway, so the sentence reads no differently.

Who watermarks text today

What a watermark cannot do

1. Most text carries no watermark at all. OpenAI built a text watermark in 2022 and has never switched it on, so text from ChatGPT carries no watermark today. Some companies publish everything needed to run their models, so anyone can download a model and run it on their own computer. The program that writes the text adds the watermark, and whoever runs that program can leave the watermark switched off.

2. Editing weakens the watermark and does not reliably remove it. In one study of the green-list design, people rewrote watermarked texts heavily by hand. A checker could still find the watermark as long as it had about 800 tokens, or roughly 600 words of English, to work with. The 800-token figure is an average from that study. The study used one watermark design and a particular set of texts and human editors.

3. A watermark gives no measure of how much of the text a person wrote. Anthropic says a watermark "cannot distinguish 'Claude wrote this' from 'Claude heavily edited this.'"

4. No reader can run the check. Anthropic's checking tool is in a private preview. Approved organisations alone can use it: government bodies, news organisations, fact-checkers, researchers, schools and the like. Google's public SynthID checker and Gemini's built-in check accept images, video and audio. Neither accepts text.

5. A watermark can be faked. In a 2024 study, researchers picked a model that plants watermarks. They sent it question after question and roughly worked out what its key was doing. They then added that pattern to text the model had never written. Added to a person's own writing, the pattern would make that writing look AI-written. The researchers put the cost at under 50 dollars. They did this under test conditions, on the particular watermark designs they chose.

3Editing historyThe writing software's log of how the text got onto the page.

The third kind of evidence is a record of the writing as it happens. Google Docs keeps a version history, a log of every edit made to a document. Grammarly's Authorship feature, released in August 2024, sits inside the document while the writer works, and it marks each stretch of text as typed, pasted, or generated by AI.

What an editing history looks like

what the writing software recorded for two essays

Essay A looks like honest work and Essay B looks like cheating. Both readings can be wrong.

What an editing history cannot do

1. An editing history shows how the words arrived and leaves out who thought of them. A student who pasted an essay in three pieces may have been moving text out of her own notes. A student who built an essay a few words at a time may have been reading an AI model's answer on a phone and typing it out by hand.

2. The record covers one document. Google Docs and Grammarly log what happens inside the document they run in. A draft written somewhere else and pasted in looks like a paste, whoever wrote it. An essay written on paper leaves no editing history at all.

A teacher who wants to know whether the student did the thinking still has to ask the student.

What if a person and a machine both wrote it?

All three techniques have limits when a person and an AI model write together. Much writing now has two authors, a person and an AI model, mixed inside the same paragraph. That makes the question harder.

Next, the sentences of Passage A and Passage B were shuffled together into one paragraph of 260 words, with 130 words from each writer.

Run Pangram on the mixed paragraph

Pale green marks Cherry-Garrard’s sentences and pale red marks the AI model’s. Run Pangram to see how it scored each sentence.

With scripts off Pangram’s scores show already, in the second card.

One summer blizzard is much like another. Rain hammered the tent until each drop seemed to strike inside my skull. I sat with my knees drawn up, watching the thin walls shudder and bow beneath the wind. Water gathered along a seam, swelled into bright beads, then fell onto the corner of my sleeping bag. The temperature, never very low, rises, and you are not cold in the tent. Sometimes a blizzard is a very welcome rest: after weeks of hard pulling, dragging yourself awake each morning, feeling as though you had only just gone to sleep, with the mental strain perhaps which work among crevasses entails, it is most pleasant to be put to bed for two or three days. The air smelled of wet nylon, mud, and the wool of my socks. Whenever lightning flashed, the scattered boots and cooking pots appeared stark against the fabric, then vanished into gray. You may sleep dreamlessly nearly all the time, rousing out for meals, or waking occasionally to hear from the soft warmth of your reindeer bag the deep boom of the tent flapping in the wind, or drowsily you may visit other parts of the world, while the drifting snow purrs against the green tent at your head. Thunder rolled through the ground beneath me. I pressed my palms around a cooling mug and counted the spaces between gusts. Gradually, the canvas stopped snapping. The rain softened to a steady patter, and somewhere beyond the entrance, water trickled over stones. I loosened my grip and stretched my cramped legs.

Cherry-Garrardthe AI model

260 words

Run Pangram

what Pangram scored

One summer blizzard is much like another.

wrong

Rain hammered the tent until each drop seemed to strike inside my skull. I sat with my knees drawn up, watching the thin walls shudder and bow beneath the wind. Water gathered along a seam, swelled into bright beads, then fell onto the corner of my sleeping bag. The temperature, never very low, rises, and you are not cold in the tent. Sometimes a blizzard is a very welcome rest: after weeks of hard pulling, dragging yourself awake each morning, feeling as though you had only just gone to sleep, with the mental strain perhaps which work among crevasses entails, it is most pleasant to be put to bed for two or three days.

right

The air smelled of wet nylon, mud, and the wool of my socks. Whenever lightning flashed, the scattered boots and cooking pots appeared stark against the fabric, then vanished into gray.

wrong

You may sleep dreamlessly nearly all the time, rousing out for meals, or waking occasionally to hear from the soft warmth of your reindeer bag the deep boom of the tent flapping in the wind, or drowsily you may visit other parts of the world, while the drifting snow purrs against the green tent at your head. Thunder rolled through the ground beneath me. I pressed my palms around a cooling mug and counted the spaces between gusts. Gradually, the canvas stopped snapping. The rain softened to a steady patter, and somewhere beyond the entrance, water trickled over stones. I loosened my grip and stretched my cramped legs.

right

scored as humanscored as AIright or wrong: against the true writer

Mixed

AI 42%human 58%

Pangram’s note: “AI-generated content appears at the start and end”

Pangram got 11 of the 14 sentences right and correctly labelled the paragraph Mixed. This example combines unchanged sentences from a person and an AI model; it does not test what happens when they revise the same sentences.

Humanizers: making AI text read as human-written

Humanizers rewrite AI text with the aim of getting it past detectors. Undetectable AI claims more than 24 million users and charges between 10 and 42 dollars a month. One popular free humanizer is the rule list from blader/humanizer, a GitHub project. The list holds 25 editing rules, and more than 48,000 people have marked it as useful. That free list was assembled from "Signs of AI writing", a Wikipedia page written to help people spot AI writing.

On 24 September 2026 an AI model applied the blader/humanizer rule list to Passage A and the mixed paragraph. The AI model took out long words and split long sentences.

Humanize, then run Pangram

Humanize each text to see its blader/humanizer rewrite, then run Pangram on that rewrite.

With scripts off each rewrite and its score show already.

Passage A · the AI model

Rain hammered the tent until each drop seemed to strike inside my skull. I sat with my knees drawn up, watching the thin walls shudder and bow beneath the wind. Water gathered along a seam, swelled into bright beads, then fell onto the corner of my sleeping bag. The air smelled of wet nylon, mud, and the wool of my socks. Whenever lightning flashed, the scattered boots and cooking pots appeared stark against the fabric, then vanished into gray. Thunder rolled through the ground beneath me. I pressed my palms around a cooling mug and counted the spaces between gusts. Gradually, the canvas stopped snapping. The rain softened to a steady patter, and somewhere beyond the entrance, water trickled over stones. I loosened my grip and stretched my cramped legs.

130 words

Humanize

Humanized with blader/humanizer

Rain hit the tent hard enough that I felt each drop in my head. I sat with my knees pulled up to my chest. The thin walls shuddered and bowed in the wind. Water gathered along a seam above me and hung there in beads. One after another they dropped onto the corner of my sleeping bag. The air smelled of wet nylon and mud. Under that was the wool of my socks. When lightning flashed I saw the boots and the cooking pots lying where I had left them, black against the wall. Then the gray came back. Thunder came up through the ground under me. I held a mug that had gone cold and counted the gaps between the gusts. The snapping of the nylon slowed, then stopped. The rain settled into a steady patter. Outside the entrance I could hear water running over stones. I loosened my hands and stretched out my legs. They had cramped from sitting.

161 words

Run Pangram

AI Generated

The mixed paragraph

One summer blizzard is much like another. Rain hammered the tent until each drop seemed to strike inside my skull. I sat with my knees drawn up, watching the thin walls shudder and bow beneath the wind. Water gathered along a seam, swelled into bright beads, then fell onto the corner of my sleeping bag. The temperature, never very low, rises, and you are not cold in the tent. Sometimes a blizzard is a very welcome rest: after weeks of hard pulling, dragging yourself awake each morning, feeling as though you had only just gone to sleep, with the mental strain perhaps which work among crevasses entails, it is most pleasant to be put to bed for two or three days. The air smelled of wet nylon, mud, and the wool of my socks. Whenever lightning flashed, the scattered boots and cooking pots appeared stark against the fabric, then vanished into gray. You may sleep dreamlessly nearly all the time, rousing out for meals, or waking occasionally to hear from the soft warmth of your reindeer bag the deep boom of the tent flapping in the wind, or drowsily you may visit other parts of the world, while the drifting snow purrs against the green tent at your head. Thunder rolled through the ground beneath me. I pressed my palms around a cooling mug and counted the spaces between gusts. Gradually, the canvas stopped snapping. The rain softened to a steady patter, and somewhere beyond the entrance, water trickled over stones. I loosened my grip and stretched my cramped legs.

260 words

Humanize

Humanized with blader/humanizer

One summer storm is much like another. Rain hammered the tent until each drop seemed to strike inside my skull. I sat with my knees drawn up and watched the thin walls shudder and bow under the wind. Water gathered along a seam, swelled into bright beads, then fell on the corner of my sleeping bag. The temperature rises in a storm like that, so the tent stayed warm enough. Sometimes a storm is a welcome rest. I had pulled hard for weeks. Every morning I dragged myself awake feeling as though I had only just gone to sleep, and the work among the crevasses wore on my mind. Two or three days of being sent to bed was a pleasure. The air smelled of wet nylon and of the wool of my socks. When the lightning flashed, the boots and pots lying about showed sharp against the fabric, then went back into gray. Thunder rolled through the ground under me. I held a cooling mug in both hands and counted the gaps between the gusts. Most of the time I slept without dreaming. I roused out for meals. Once or twice I woke in the warmth of the bag and heard the deep boom of the tent flapping in the wind, and once I drifted half awake through other parts of the world while the rain purred against the fabric by my head. The canvas stopped snapping. The rain settled into a steady patter, and past the entrance water ran over stones. I loosened my grip on the mug and stretched my cramped legs.

265 words

Run Pangram

AI Generated

one of Cherry-Garrard's sentences, before and after the rewrite

You may sleep dreamlessly nearly all the time, rousing out for meals...

Most of the time I slept without dreaming. I roused out for meals.

This basic humanizer did not get past Pangram. Both rewrites scored 100% AI.

Passage A scored the same after its rewrite. An AI model rewrote a paragraph that an AI model had written, so the result was still AI writing. In the mixed paragraph the AI model replaced Cherry-Garrard’s sentences with its own paraphrase of them. Pangram also flagged the model’s paraphrases of Cherry-Garrard as AI-written.

The blader/humanizer rule list is a mild humanizer. Its rules make prose read more plainly, and an AI detector trained on examples flags the result anyway.

The paid services are built to defeat AI detectors. In an independent test posted in August 2026, one of them worked. Researchers at Notre Dame put AI-written abstracts through it, and fewer than 4 in 100 stayed flagged by Pangram. The researchers also found that a human abstract given a light polish by an AI model was flagged as AI-written between 38 and 80 times in 100. The rate depended on the field. In the researchers' own words, honest AI editing carries a higher risk of being punished than paid evasion does.

Both of those figures came from an older version of Pangram. The independent tests cited here do not tell us how paid humanizers perform against the latest detectors.

Pangram then retrains against the humanizers it finds. Pangram reports that it catches nearly all the text from thirteen humanizer services. The company measured that figure on its own set. Its chief executive puts Pangram's catch rate at 50 to 70 percent against a humanizer it has not trained on.

Cat and mouse game between detectors and humanizers

With scripts off every move shows at once.

Detectors like Pangram

Humanizers

Detectors like Pangram · journalist2 January 2023Edward Tian, a Princeton student, releases GPTZero, an AI detector. It scores how predictable the wording is and how much the sentences vary.

Humanizers · independent2 August 2023Researchers have GPT-4 rewrite 20 abstracts, then run them through Undetectable AI, a humanizer. Originality.ai’s average AI score falls from 91% to 28%.

Detectors like Pangram · company23 January 2025Pangram Labs ships a detector trained on humanized text. On Pangram’s own test set it catches 94% of humanized AI text; GPTZero catches 35%.

Humanizers · journalist30 May 2026A reporter at The Atlantic gets AI text past Pangram with Walter Writes AI, a humanizer.

Detectors like Pangram · company29 July 2026Pangram Labs tests Pangram 4 on text from 13 commercial humanizers. On the company’s own test set it flags 98.8% as involving AI.

Humanizers · independent6 August 2026Undetectable AI version 11, a paid humanizer, gets over 96 in 100 AI rewrites of abstracts past GPTZero and Pangram 3.2. Pangram 4 was untested.

Detectors like Pangram · journalist8 September 2026A Fast Company reporter writes that Pangram has added training aimed at each humanizer. In his own test it seems to catch Walter Writes AI.

Humanizers · company8 September 2026Pangram’s chief executive, Max Spero, puts Pangram’s catch rate at 50 to 70% for a humanizer it has not trained on.

2 Jan
20232 Aug
202323 Jan
202530 May
202629 Jul
20266 Aug
20268 Sep
20268 Sep
2026continues

continues

Next move Show all moves

Humanizer makers keep releasing new tools and Pangram retrains on the ones it finds, so the exchange continues after the last box.

In summary

Pangram caught the AI-written passage in the opening quiz. When human and AI sentences were mixed together, it identified eleven of fourteen correctly. Those results are useful. They also show why a detector’s answer needs interpretation: it can recognise a mixed paragraph and still accuse the wrong sentence.

Each technique gives us different evidence. A detector judges the finished words. A watermark can reveal a model’s involvement, including when it edited someone’s writing. An editing history records how the words reached the page. None gives a complete account of who did the thinking.

Humanizers make that uncertainty harder to resolve. Detectors learn to catch their rewrites, and humanizers change again. The independent tests cited here cannot tell us which tools are ahead today. When authorship matters, a score should start a closer look. Read the work, examine the drafts, and ask the writer to explain their choices.

Who wrote a text is a question about the past. Soon most of what you read will have passed through a machine somewhere. The question that will still matter is whether it is true.

Back to Artificial Minds.