Catching Unicorns with GLTR
Spot AI-generated text at a glance with GLTR’s color-coded statistical forensics.
Quick facts
- Best for
- Spot AI-generated text at a glance with GLTR’s color-coded statistical forensics.
- Pricing
- Freemium
- Editor rating
- 4.5 / 5
- Community saves
- 0
About Catching Unicorns with GLTR
GLTR (Giant Language model Test Room) is a visually forensic, open-source tool for detecting AI-generated text from large language models like GPT-2. It highlights the predictability of each word using a color-coded overlay (green/yellow for likely, red/purple for unlikely), provides histograms of rank distributions and entropy, and shows top predicted tokens on hover. With a live demo, pre-loaded examples, and research-backed improvements in human detection accuracy (from 54% to 72%), GLTR empowers journalists, educators, researchers, and the public to distinguish machine-written from human-authored content. Developed by MIT-IBM Watson AI Lab and HarvardNLP, GLTR debuted as an ACL 2019 demo and is available on GitHub.
Pros
- Color-coded token visualization (green/yellow/red/purple) by probability rank
- Live demo to paste any text for instant analysis
- Histograms showing distribution of top-10/top-100/top-1000 ranks
- Entropy and predictability summaries for entire texts and sentences
- Hover tooltips with top predicted words and ranks
- Sentence-level breakdowns with per-sentence charts
- Model selection for different GPT-2 variants
- Pre-loaded real and fake text examples for comparison
- Analysis examples of algorithmic journalism (e.g., sports reports)
- Open-source code available on GitHub for local use and customization
- Developed by MIT-IBM Watson AI Lab and HarvardNLP researchers
- ACL 2019 demo, nominated for best demo
Cons
Pricing
- • Paste text for analysis
- • Color-coded word probability visualization (green/yellow/red/purple bins ~ top-10, top-100, top-1000, >1000 rank buckets)
- • Uses pretrained GPT-2 (e.g., 117M) for token likelihoods
- • Downloadable results
- • No registration/login required
- • No stated usage limits