Annamieka Aerts

How Pangram measures accuracy, and how Pangram 4 performs on false-positive and false-negative benchmarks across writing domains, AI models, mixed authorship, diverse authors, and text length.

Yes. Pangram detects Gemini 3. 8 Flash in 998 of 1,000 tested examples (99.

Yes. Pangram detects Claude Sonnet 5. 5 in 1,019/1,022 (99. 71%) of our tested examples.

Yes. Pangram detects Claude Opus 5. 5 in 997/1,000 (99. 70%) of our tested examples.

How AI text watermarks like SynthID and Claude’s work, which models use them, how they can be beaten, and what they mean for AI detection.

A brief report on how Pangram 4 performs when given a collection of items written by AI agents using an obscure web forum to help each other cheat on their assigned task.

Yes. Pangram detects GPT-6 Astra in 982/986 (99. 59%) of our tested examples.

Can Pangram 4 detect Anthropic's newest model, Fable 5. 1?

Pangram scanned 178,672 PR Newswire press releases and found AI-generated text in 34% of them over the past year, climbing to 48% by June 2026.

We evaluated Pangram 4 on 25,996 human-written student essays from the PERSUADE 2.

A technical overview of Pangram 4, our latest AI text detection model: how it was built, how it was evaluated, and how it compares to Pangram 3.

A plain-English explanation of how Pangram's AI detector learns to tell human writing from AI-generated text.

We scanned over 1 million social media posts for AI content.