GoKawiilOur classifier uses a traditional language model architecture. It receives input text and tokenizes it. Then, the model turns each token into an embedding, which is a vector of numbers representing the meaning of each token. The input is passed through the neural network, producing an output vector. A classifier head transforms the output vector into a prediction of human, AI, or AI-assisted. We train an initial model on a small but diverse dataset of approximately 1 million documents composed of publicly licensed human-written text. The dataset also includes AI-generated text produced by GPT-5 and other frontier language models. The result of training is a neural network capable of reliably predicting whether text was authored by human or AI.
Pangram – AI detector for text and images
Why This Matters
Pangram introduces an AI text and image detector built on a standard language-model classifier architecture, trained on a curated dataset of human-written and AI-generated content. As AI-generated content proliferates online, tools like this matter for verifying authenticity in journalism, education, and content moderation, though their reliability and potential for false positives remain key concerns for the industry.
Key Takeaways
- Pangram uses a transformer-based classifier that tokenizes text, generates embeddings, and predicts human, AI, or AI-assisted authorship.
- The model was trained on roughly 1 million documents mixing publicly licensed human text with AI-generated content from GPT-5 and other frontier models.
- AI detection tools like this are increasingly important for combating misinformation and verifying content authenticity, though accuracy and bias in such classifiers remain ongoing challenges.
Get alerts for these topics