Announcing an update to Pangram Text, state-of-the-art AI text detection. Pangram Text now officially supports the top 20 languages on the internet, and performs well on many other languages outside of that!. See the thread for technical detail on how we did it! 1/5

RT @brian_jabarian: AI-generated text is everywhere: hard for orgs to assess human performance. Can we detect it while min false accusation….

Generate videos in just a few seconds. Try Grok Imagine, free for a limited time.

RT @qwoted: Can you believe it’s already time for our summer Media Mixer?. On August 21st, we’re partnering with @pangramlabs, the AI detec….

Read about our false positive rate and how we measure it here:

Thanks for reading! Find our full blog post on day one GPT-5 tests here:

However, we do not believe this is an inherent "tradeoff"- we believe we can train models with both low false positive and false negative rates, and that is why we conduct benchmarks on both sides!.

We calibrate and tune the threshold such that our false positive rate is approximately 1/10th of our false negative rate, treating false positives much more seriously than false negatives.

Of course, accuracy on AI text isn't enough. We must consider the false positive rate: how often does Pangram flag human-generated text as AI?. Our measured false positive rate is 1 in 10,000 across all domains, and is often lower in practice.

We had time today to run one last test, which is to see how well Pangram is able to detect all the different flavors of GPT-5, as well as the new GPT-OSS models, using the API. We tested a few thousand examples each with our standard prompts and a minimum of 50 words.

Pangram correctly identified all 10 responses as AI-generated with 99-100% confidence, while GPTZero incorrectly classified them all as human-written, with most responses as having 0% chance of being AI-generated. Full prompts and responses here:.

One of the claims made in today's livestream was that GPT-5 demonstrates "significantly improved writing." We came up with prompts that would showcase GPT-5's ability by feeding it prompts that force the model to stretch both its reasoning abilities and its creativity.

Other AI detectors, GPTZero, ZeroGPT, Grammarly, UndetectableAI, and Originality all predict that this text is human-written. We believe this speaks to the strength of our research approach and generalization abilities backed by a powerful model with a very high quality dataset!

Pangram is able to predict that the entire document is AI-generated with extremely high confidence.

Our first test of GPT-5 started with an office inside joke: trying to convince someone that eating shrimp is better than eating lobster. We asked GPT-5 to write an impassioned speech for why eating shrimp is superior to eating lobster.

This afternoon, we ran a more in-depth test, and we were quite pleased with the results: we have yet to find a failure case of Pangram's AI detector against out most creative prompts on GPT-5, and we find the competition's results are spotty at best.

During OpenAI's announcement livestream, we tested Pangram against the three written examples generated by GPT-5. Pangram was still able to catch these outputs - despite never being trained on GPT-5 outputs!.

GPT-5 was released twelve hours ago. How well does Pangram do at detecting GPT-5 outputs? We found that while Pangram is able to detect 99.8%+ of GPT-5 outputs, other AI detectors consistently marked the texts as human. Let's take a look at the results: 🧵

RT @max_spero_: Happy GPT-5 day everybody! I will be live-tweeting Pangram's results on OpenAI's demo outputs during the livestream today.….

RT @max_spero_: New @pangramlabs feature. coming soon!

We’re excited to announce our partnership with Qwoted, a network built to support working media and connect them with credible, human voices. As journalists and content creators face growing pressure to vet the authenticity of their sources, we’re proud to help Qwoted provide a.