To clarify something that tripped me up for other readers: F1-score is the harmonic mean of precision and recall, so that Table 2 (given in a screenshot in this post) from the Oct 2025 EditLens paper is meaningful on its own (it's not just "accuracy" with no mention of false positives).
(The recent Pangram 3.3 has false positive rates of less than 0.5-1.5%, sometimes less than 0.05%, depending on the human-written dataset it's tested on.)
1One thing I've never seem discussed is, given that Pangram allows anyone to use it, can't you just adversarilly iterate any text you want against Pangram until you get a low AI score? You can even get an AI agent to do this.
One issue I see with AI text prediction services is that over time normal humans' output could trend stylistically towards being LLM generated and increase the false positive rate significantly. The human version of model collapse if you will.
That is to say, right now, an "average human" likely has 20, 40, 60 etc years of natural language mimicry to go on. However, as time goes on the ratio of "LLM-input" to "non-LLM-input" increases.
I would be interested to know whether this is being monitored, for example I would expect it to be most visible first in language learning communities e.g. new learners of a language beginning to sound more "LLM-y" over time.
[Edit: an earlier version of this post claimed that Pangram has announced plans to expand to Canada. A reliable source of information (Pangram's CEO) informed me that this is false].
Pangram Labs builds the most accurate AI text detector in the world. Team is ~25 FTE; they are active on Twitter, you can engage directly, look for "affiliates" tab of @pangram.
Here is a table of their performance on adversarially modified AI text (source paper):
Language
AI Text Detection %
Humanized AI Text Detection %
GPTZero
95.60%
34.53%
Binoculars
94.40%
29.73%
Pangram Baseline
100.00%
73.07%
Pangram Humanizers (current model!)
100.00%
93.66%
Note that "current model!" is not current as of July 2026. Their classifier now provides a percentage instead of a binary verdict. They released an open source model (Llama-3.2-3B QLoRA) which was SOTA at the time. The paper does not test adversarially modified AI text, but you are welcome to try running this test (repo); it may trigger agent safeguards.
Note again that "Pangram" in this table is not current as of July 2026. Their production model detects Fable 5 outputs with 99.64% accuracy (blog). See prompts. Reasoning effort level (High, Max, etc) is not disclosed. Just to be clear, Pangram knows that the output came from some AI model, but their technology does not predict the specific model used.
I expect that Pangram's business will grow faster than the following AI companies that I am somewhat familiar with but haven't studied in detail: Ideogram, Elevenlabs, Cerebras, Cognichip, Decagon, and Cohere. I lack sufficient information to forecast their ultimate size. Pangram's Chrome extension has ~10k users, I would upper bound this at ~50m if they can maintain their leader position.
If I had to guess, the neolab most likely to acquire them is Thinking Machines, based on the premise that we primarily care about the provenance of text when humans will read it. If the target is a machine and not a human, the content of the payload matters more than its origin (Edit: I have changed my mind about this after gwern pointed out that for TMI, if customers find some text acceptable, TMI will train on it, so customers are doing data filtering).
Detection is the first step in mitigating security threats, so initially, better detectors sound great. But evasion co-evolves with detection. Currently, humans can easily detect AI text. But if Pangram succeeds (operating as a filter on most text shown to humans in environments where LLM text is excluded), this may no longer be the case, and we may become dependent on AIs for accurate AI text detection sooner. It is not guaranteed that natural text will be trusted text, but LLM generated text is more likely to be unsafe in cases where many egregiously misaligned AIs are hostile to humanity without being strong enough to take over. AI verification is an important prerequisite for an agreement in Plan 2040, but I haven't thought about whether AI text detection counts.