Spotting synthetic prose before the LLM flood

What it does

GLTR is a web-based forensics tool that scores text against a language model’s own expectations. You paste in prose, and it highlights how predictable each token is to GPT-2 or BERT — the theory being that machine-generated text tends to stick to the model’s safest bets. The result is an interactive visualization that maps suspiciously likely words across a passage.

The interesting bit

Built in 2019 when GPT-2 was still considered alarming, GLTR turns the generative model against itself: instead of sampling likely tokens, it uses the model’s probability distribution to flag them. The frontend renders this analysis visually, making statistical detection feel more like reading a weather map than parsing log probabilities.

Key highlights

• Visualizes per-token probabilities to surface synthetic text patterns.
• Ships with backends for gpt-2-small and BERT, swappable via a server flag.
• Exposes an extensible AbstractLanguageChecker API for adding custom models.
• Includes a live demo at gltr.io and can run fully offline as a local server.
• Born from a collaboration between MIT-IBM Watson AI Lab and HarvardNLP.

Caveats

• Demo texts are only bundled for the gpt-2-small backend; the BERT mode leaves you to supply your own examples.
• The frontend build workflow is manual, and the README offers no guidance on modern model support beyond GPT-2 and BERT.
• The project’s visual and architectural language predates the current LLM era, so its effectiveness against newer models is unclear.

Verdict

Worth a look if you study AI-generated text detection or want a concrete 2019 baseline for model forensics. Skip it if you need a turnkey detector for modern LLMs.