Pangram has been billed as the “gold standard” of AI detection and the “world's best,” claiming accuracy rates above 99% for text from mainstream AI models and offering an AI image detection feature. Pangram says its tools use proprietary AI models, that third parties have independently verified its accuracy and reliability, and that its tech outperforms humans, including on text from tools that aim to humanize AI content. Its development includes mirrored prompts, in which a human writes from a prompt and the same prompt is given to an AI model. Testing showed Pangram consistently picks out AI content but also regularly gets things wrong. Pangram requires submissions of at least 50 words. After adding a Gemini-generated sentence to a draft, Pangram failed to detect any AI content; after adding another Gemini sentence, it still failed. Only after adding three sentences, going from an original 250 words to 311, did Pangram find anything amiss, labeling only two sentences as AI-generated, one of which the author wrote. When the flagged AI-generated sentence was rephrased, Pangram could not detect anything.