GPT-2 Output Detector and GLTR: The Hugging Face AI Detector

When people search for the “Hugging Face AI detector,” they usually mean OpenAI’s GPT-2 Output Detector, a RoBERTa-based classifier released in 2019 and hosted as a demo on Hugging Face. Alongside GLTR, a visual tool from MIT-IBM Watson AI Lab and HarvardNLP, it was one of the first free ways to detect machine-written text. Both were built for GPT-2, and as of November 2026 neither official demo is working, so they’re best treated as history rather than a tool for checking modern AI text.

Here’s what each tool did, how it worked, what happened to it, and what to use now.

What was the OpenAI GPT-2 Output Detector?

In 2019, OpenAI released GPT-2 in stages, partly over concerns that it could be used to produce convincing fake text. Along with the largest 1.5-billion-parameter version, it released a detector to help people identify GPT-2 output.

The detector is a RoBERTa model (a language model from Facebook AI) fine-tuned to tell apart two kinds of text: output from GPT-2 1.5B and human-written web text from WebText, the dataset GPT-2 was trained on. You pasted in text and it showed a “real” versus “fake” probability.

According to its model card on Hugging Face, the classifier detected GPT-2 1.5B text with about 95% accuracy. OpenAI itself said that wasn’t high enough for standalone detection and should be paired with metadata, human judgment and public education. The card also notes the model works only on English and is released under an MIT license. OpenAI published two sizes, a base model and a larger one, and the card adds that detection gets harder as the source model gets bigger.

The demo got a second life after ChatGPT launched in late 2022. It was free, fast and easy to find, so plenty of people tried it as a quick “ChatGPT detector,” even though it was never trained on ChatGPT. That’s largely where search terms like “huggingface openai detector” and “gpt 2 detector” come from.

It’s worth separating three things that often get mixed up under the “OpenAI detector” name:

• The GPT-2 Output Detector (2019): the RoBERTa model and Hugging Face demo covered here.
• The AI Text Classifier (2023): a separate tool OpenAI launched for newer models and later withdrew.
• Community copies: unofficial Hugging Face Spaces and websites that wrap the 2019 model under new names.

How did the GPT-2 Output Detector work?

It was a classic classifier model:

1. Researchers generated a large set of GPT-2 outputs using different sampling settings.
2. They paired those with real human web text.
3. They fine-tuned RoBERTa to label each sample as human or machine-generated.
4. At check time, the model read your text and output a probability for each label.

The key detail is step 1. The detector only ever learned what GPT-2 text looks like. The code and training notes live in OpenAI’s gpt-2-output-dataset repository on GitHub, which explains how to run the detector on your own machine.

What was GLTR and how did it work?

GLTR, short for Giant Language Model Test Room, took a different approach. Instead of giving you a verdict, it showed you the evidence.

Built by Hendrik Strobelt and Sebastian Gehrmann with Alexander Rush, as a collaboration between the MIT-IBM Watson AI Lab and HarvardNLP, it was presented at ACL 2019 (paper). According to the GLTR introduction page, it ran your text through the small GPT-2 model (117M parameters) and checked, for every word, how highly GPT-2 had ranked it as the next word. Each word was then colored:

Color	Word’s rank in GPT-2’s predictions	What it suggests
Green	Top 10	Very predictable
Yellow	Top 100	Fairly predictable
Red	Top 1,000	Less expected
Purple	Beyond the top 1,000	Surprising

A page that was almost all green and yellow looked machine-generated. Human writing usually had more red and purple, because people pick less predictable words, and those surprising words tend to carry the real meaning. GLTR also showed histograms summarizing the whole text.

The research result was striking for its time. In a human study, the authors found GLTR’s color coding raised people’s ability to spot fake text from 54% to 72% without prior training. It’s essentially a visual version of perplexity, which we explain in how AI detectors work.

Is the Hugging Face AI detector (and GLTR) still online?

We checked both in November 2026.

Tool	Status in November 2026
OpenAI detector demo (huggingface.co/openai-detector)	Redirects to the official Space, which Hugging Face shows as paused; the demo address returned a “service unavailable” error
GPT-2 detector model weights	Still published on Hugging Face, where community Spaces and code can load them
OpenAI detector code	The GitHub repository is marked as archived and read-only
GLTR intro page	gltr.io still loads over plain HTTP; the secure version showed an expired certificate
GLTR live demo	The demo link (gltr.io/dist/index.html) returned a “404 Not Found” error
GLTR code	Still on GitHub, where it can be run locally

So you can still run either tool yourself if you’re comfortable with Python. You’ll also find many unofficial Hugging Face Spaces that wrap the GPT-2 detector. Just know that they’re running the same 2019 model.

Why aren’t they reliable for ChatGPT and modern AI?

Both tools were designed for GPT-2, a model that is now several generations old. That causes three problems:

• They learned the wrong target. The GPT-2 detector was trained to recognize GPT-2’s habits. Modern models write differently, and classifiers often struggle with models they weren’t trained on, as the RAID benchmark found across 12 detectors.
• Their own documentation says so. The GPT-2 detector’s model card warns against using it to make academic misconduct allegations about ChatGPT text, since results may be inaccurate.
• Independent testing agrees. In a 2025 study by Jabarian and Imas, summarized by Chicago Booth Review, an open-source RoBERTa detector performed close to random guessing in many cases and was judged unsuitable for high-stakes use, while commercial detectors did much better.

GLTR has a related problem. It measures predictability under small GPT-2. Plain human writing can look predictable, and modern models can be prompted to write in less predictable ways. Even OpenAI’s later detector for newer models, its 2023 AI Text Classifier, was withdrawn for low accuracy.

These tools still have value for learning. GLTR in particular is one of the clearest ways to see why AI text tends to be predictable. They just shouldn’t be used to judge someone’s work today.

What should you use instead?

For checking modern text, look for a detector that:

• Is updated for current models like recent versions of ChatGPT, Claude and Gemini.
• Shows sentence-level results so you can see which parts look AI-written, in the spirit of GLTR’s highlighting.
• Explains its signals rather than only giving a number.
• Is honest about uncertainty, especially for short or formal text.

Our free AI detector gives you an overall AI percentage, highlights each sentence as likely AI, possibly AI or human, lists the signals behind the score, and shows stats like burstiness (variation in sentence length) and vocabulary diversity. It’s free to try without an account, with more daily checks on a free account. See pricing for higher limits.

For more options, see our roundup of the best free AI detectors and the best ChatGPT detectors. Whichever tool you use, compare at least two, and read our guide on how to check if text is AI generated before acting on a result. As the evidence in Are AI detectors accurate? shows, any detector score is a signal, not proof.