Last updated 08-11-2026
Category:
Reviews:
Join thousands of AI enthusiasts in the World of AI!
Catching Unicorns with GLTR
GLTR is a forensic text analysis demo that highlights how predictable each word in a passage looks to a language model and colors the text so you can spot machine-written prose at a glance. Paste a sample into the live demo, and GLTR ranks every next-word prediction from GPT-2 117M, painting likely picks green or yellow and surprising human choices purple.
Unlike black-box AI detectors that return a single score, GLTR shows the forensic footprint word by word, plus histograms for top-k counts, probability ratios, and prediction entropy across the passage. That makes it useful for teaching how generative models stick to safe, high-probability wording while human writers reach for rarer terms that still fit the topic.
Researchers, journalists, and educators use GLTR to compare sample texts, explore model behavior, and sanity-check suspicious copy. The project is free, open source, and backed by an ACL 2019 demo paper from MIT-IBM Watson AI Lab and Harvard NLP, though the maintainers note it was built for GPT-2 era text and may struggle with newer models like ChatGPT.
Colors each word green, yellow, red, or purple based on GPT-2 117M top 10, 100, 1,000, or rarer predictions
Live demo at demo.gltr.io accepts custom text plus bundled real and fake samples
Hover overlays show the top 5 predicted next words with probabilities for any token
Three histograms chart top-k counts, probability ratios, and prediction entropy across a passage
Open-source code lives on GitHub under detecting-fake-text for local deployment
ACL 2019 demo paper reports human fake-text detection rising from 54% to 72% with GLTR overlays
Maintainers link to the newer RADAR demo for testing text from recent large models
Word-level color overlays make model behavior visible instead of hiding it behind one probability score.
Free demo and open-source release let classrooms and newsrooms experiment without procurement hurdles.
Bundled real and fake samples, including the famous unicorn GPT-2 passage, teach what generated prose looks like.
Histograms summarize entropy and top-k patterns across an entire document for quicker comparisons.
Published ACL study shows GLTR helped non-experts catch fake text more often during user testing.
The underlying model is GPT-2 117M, so results on ChatGPT or other modern generators are unreliable.
GLTR inspects one passage at a time and is not built for large-scale automated moderation pipelines.
Interpreting purple or red words still requires language fluency to judge whether an uncommon term is legitimate.
What does GLTR stand for?
Catching Unicorns with GLTR stands for Giant Language model Test Room. The demo from MIT-IBM Watson AI Lab and Harvard NLP shows how likely each word in a passage would be predicted by a language model.
How does GLTR detect generated text?
Catching Unicorns with GLTR compares each word in your text to GPT-2 117M next-word rankings. Words in the top 10 predictions turn green, top 100 yellow, top 1,000 red, and rarer choices purple because generated prose tends to stay on highly likely tokens.
Is GLTR free to use?
Yes. Catching Unicorns with GLTR is free through the public demo at demo.gltr.io, and the source code is open on GitHub. There is no paid plan or account requirement to analyze sample text.
Can GLTR analyze my own text?
Yes. The Catching Unicorns with GLTR demo lets you paste custom passages or load bundled examples, then click analyze to render the color overlay and histograms. You can hover individual words to inspect the top five model predictions.
Does GLTR work on ChatGPT output?
Catching Unicorns with GLTR was built in 2019 to test GPT-2 era text and may not be reliable on newer models such as ChatGPT. The demo disclaimer points users to the RADAR demo at radar-app.vizhub.ai for a newer detection approach.
Who created GLTR?
Catching Unicorns with GLTR was created by Hendrik Strobelt and Sebastian Gehrmann with review by Alexander Rush as a collaboration between the MIT-IBM Watson AI Lab and Harvard NLP. The ACL 2019 demo track paper is publicly available on arXiv.