Why Pangram Became the AI Detector Everyone Tests Against

When a humanizer company wants to prove its product works in 2026, it rarely brags about beating ZeroGPT anymore. It names Pangram.

That shift says a lot. StealthGPT’s founder announced a new model in September 2026 after saying Pangram 4 had broken the company’s earlier ones (RuntimeWire, September 2026). WriteHuman advertises a Pangram score on its homepage. Third-party testers pick it as their yardstick.

How did a detector founded around 2024 by two Stanford graduates become the bar everyone measures against? And should it be?

From Newcomer to Default Yardstick

What Pangram does differently

Older detectors mostly asked one question about each word: would a language model have predicted this? Text full of predictable choices got flagged. That approach, known as perplexity analysis, has a famous weakness. It tends to label very familiar human writing, like the Declaration of Independence, as machine-made, because models have read it so many times.

Pangram’s team describes a different approach. Instead of reverse-engineering how a model picks the next word, it trains a classifier to answer “who does this sound like?” Its 2024 technical report says the human training text was verified as written in 2021 or earlier, before AI writing flooded the web, and that a technique called hard negative mining cut false positive rates “by 100x-1000x” on holdout sets (Emi and Spero, 2024). Those are the company’s own numbers. What changed its reputation came later, from people who don’t work there.

The independent results that mattered

Two studies did most of the heavy lifting.

The Chicago Booth comparison

Brian Jabarian and Alex Imas compared commercial and open-source detectors on human and AI passages (NBER working paper 34223, September 2025). Pangram was the only tool that held a false positive rate at or below 0.5% without giving up accuracy. When the researchers ran AI text through StealthGPT, GPTZero lost much of its ability to detect it. Pangram’s performance stayed largely intact.

The expert-reader study

Jenna Russell, Marzena Karpinska and Mohit Iyyer asked a sharper question: can people who use ChatGPT a lot spot AI writing, and how do machines compare? Their ACL 2025 paper found that frequent users were very accurate, and that among the automated detectors, only Pangram matched them. On 300 articles, it caught 98.0% of AI text with a 2.0% false positive rate. A version tuned for humanized text caught 99.3%, with false positives at 2.7%.

Neither study is the final word. But both came from researchers with no stake in Pangram, and both put it at the top.

The money and the new model

In July 2026, Pangram raised $9 million in a round led by Menlo Ventures (TechCrunch, July 2026) and released Pangram 4. The company’s launch post claims the new model catches humanized AI text 98.83% of the time across 13 commercial tools, with a false positive rate of 0.0041%, roughly one in every 24,000 documents. It also says its false negative rate fell to 0.34%, from 1.99% in Pangram 3.

That’s an extraordinary claim, and it’s the company’s own. In the same coverage, CEO Max Spero gave a looser figure, around one false flag per 10,000 human documents. So, by Pangram’s own accounts, somewhere between one in 10,000 and one in 24,000. No independent test yet has been large enough to confirm either number.

Why Everyone Measures Against It Now

A harder bar makes a better test

If you’re testing a rewriting tool, the weakest detector tells you nothing. Plenty of tools “pass” detectors that miss half of plain ChatGPT output anyway. The strongest available detector is the only one where a pass means something.

That logic shows up in Phrasly’s September 2026 Benchmarks, which ran four humanizers through the Pangram API using the pangram-4 model. The gaps were wide. Phrasly Ultra averaged a 99.8% human score, StealthGPT 50.0%, Undetectable AI 24.5% and WriteHuman 13.3%. Phrasly sells Ultra, so treat that as one company’s test. But the choice of detector is telling. Nobody bothers benchmarking against a pushover.

The evidence so far, side by side

Study or claim Who ran it Key finding Independent?
Jabarian and Imas, 2025 Chicago Booth researchers Only tool with false positives at or below 0.5% Yes
Russell et al., ACL 2025 University researchers 98.0% caught, 2.0% false positives on 300 articles Yes
DAMAGE paper, 2025 Pangram staff 98.26% of humanized text caught at 5% false positives No
Pangram 4 launch, 2026 Pangram 98.83% on 13 humanizers, about 1 in 24,000 false flags No
Humanizer benchmark, 2026 Phrasly Three of four tools averaged 50% human or lower No (vendor of one tool)

 

The Caveats Worth Knowing

Being the toughest detector isn’t the same as being infallible.

It can be beaten

In March 2026, a LessWrong writer showed that repeated, carefully prompted rewrites pushed one essay from 63% AI to “100% human” on Pangram 3.2. The author added a note that Pangram 4 is much stronger, and warned that results on short texts under about 250 words were unreliable. That’s the real pattern: Pangram raises the cost of evasion. It doesn’t make evasion impossible.

Its best numbers are its own

The headline figures for Pangram 4 come from Pangram. That doesn’t make them wrong, but it puts them in a different category from the Booth and ACL results. Competing vendors publish rival benchmarks where they come out ahead, too.

Scope matters

Most of the strong results involve long English prose. Short texts, heavy human editing and mixed authorship are harder for every detector, Pangram included.

A score is not a verdict

Even at one error in 24,000, a large school or publisher will eventually flag someone innocent. Any detector score, from any tool, should start a conversation, not end one.

What This Means if You Use or Face a Detector

For editors, teachers and hiring teams

If your organization picks a detector, the research so far suggests the choice matters more than most people assume. The gap between the strongest and weakest tools isn’t a few points. In the Booth study, some tools missed large shares of humanized text outright, while the open-source detector they tested was judged unsuitable for high-stakes use. Picking a cheap tool to save money can mean flagging the wrong people, or missing what you were looking for.

Whatever you choose, write down a process for what happens after a flag. Who reviews it? What evidence counts? Can the writer show drafts or version history? A strong detector paired with a weak process still produces unfair outcomes.

For writers

If you write your own work, the encouraging part of this research is the low false positive rate in the independent studies. The cautionary part is that no rate is zero. Keep your drafts and notes, especially on anything important. If you use AI tools in your process, follow whatever rules apply where you’re submitting, and be ready to explain how the piece came together.

A Higher Bar, Not a Final Answer

Pangram earned its status the slow way, through independent studies that kept putting it at or near the top, especially on the false positives that hurt real people. That’s why rewriting tools now quote it, testers pick it, and rivals race to beat it.

But a yardstick only stays useful if people keep checking it. Pangram 4’s boldest numbers still need outside confirmation, and every new humanizer release is, in effect, a fresh challenge. For anyone reading test results, the takeaway is simple. Ask which detector was used, which version, and who ran the test. If the answer is “Pangram 4, published data, someone with nothing to sell,” you’ve found a result worth paying attention to. Anything less deserves a raised eyebrow.

READ MORE: Tech winks

Leave a Reply

Your email address will not be published. Required fields are marked *

Captcha loading...