Products SEO & GEO Agent Paid Ads Agent Website Agent Conversion Agent
How it works Pricing FAQ Blog Book a call

GLTR AI checker: how it works and its limits today

Updated · 5 min read

Short answer GLTR is a 2019 research tool by MIT-IBM Watson AI Lab and HarvardNLP researchers. It colors each word by how likely the small GPT-2 model was to predict it. It gives no score, its public demo is offline, and edited or prompted text from today's chatbots can beat it. Treat it as a teaching aid, not a verdict.

Want help applying this to your business? Talk it through on a free 30-minute call.

Book a call

What is the GLTR AI checker?

GLTR (Giant Language Model Test Room, pronounced "glitter") is a free, open-source research demo from 2019. Its authors, Sebastian Gehrmann, Hendrik Strobelt and Alexander Rush, built it as a joint project of MIT-IBM Watson AI Lab and HarvardNLP, Harvard's language-technology group (GLTR project page). It doesn't say "AI" or "human." It paints each word a color and leaves the call to you.

Four facts decide whether it's worth your time today:

QuestionAnswer as of November 2026
What model does it check against?GPT-2 117M, OpenAI's small 2019 model (GLTR project page)
Can I use it online?No. The demo page at gltr.io/dist returns "404 Not Found"; Internet Archive captures show it down since early 2024
Is it still updated?The code on GitHub was last changed in January 2024 (GitHub)
Does it give a score?No. Colors and three charts only

How does GLTR work?

A language model (the engine behind tools like ChatGPT) writes by guessing the next word, one word at a time. GLTR runs your text through GPT-2 and asks, at every word: where did the actual word rank among GPT-2's guesses? The authors explain the bet behind it:

"We make the assumption that computer generated text fools humans by sticking to the most likely words at each position." (GLTR project page)

People, by contrast, more often pick words a model didn't see coming.

What do GLTR's colors mean?

ColorThe word ranked in GPT-2's...Reads as
Greentop 10 guessesvery predictable
Yellowtop 11–100predictable
Redtop 101–1,000surprising
Purplebeyond 1,000very surprising

Ranks from the GLTR project page. A page that's nearly all green and yellow looks machine-made. Lots of red and purple looks human. The authors also noted that the words carrying the real message, like names and specific facts, "almost always appear in red or purple" (GLTR project page).

Three charts sum up the whole text: how many words fall in each color, how the actual word's probability compared with GPT-2's top guess, and how sure the model was at each step (called entropy: low entropy means the model had one obvious next word). Human writing tends to show more uncertainty (GLTR project page).

How accurate is GLTR?

It made people better guessers in one small 2019 test, nothing more. Its own paper reports that 35 students in a college language-technology class, shown news-style texts, picked out the fake ones 54.2% of the time unaided and 72.3% with GLTR's colors (Gehrmann et al., ACL 2019). Most fakes came from GPT-2, the same model family GLTR checks against; the rest from Heliograf, the Washington Post's news-writing bot. The paper didn't test any of today's chatbots.

Why does GLTR miss modern AI text?

GLTR was a fair test for 2019 text. Five things weaken it now:

  1. It rests on one assumption about how AI picks words. The authors wrote that their assumption "is also limited in that it assumes a simple sampling scheme" and that adversaries "might change the sampling parameters per word or sentence" (GLTR project page). Today's tools let anyone change those settings, and their makers tune them.
  2. The judge is a small 2019 model. GPT-2 117M scores the text. It never saw how current chatbots phrase things, so its sense of "predictable" is a 2019 one.
  3. Prompts and edits blur the signal. A 2023 benchmark on ChatGPT answers found GLTR's color counts still carry a signal when a computer model is trained on them (F1 of 0.865 on one dataset; F1 is a 0-to-1 accuracy measure). That's a lab setup, not the color view you read. Adding a prompt asking for human-sounding text cut that score by 0.029, and most detectors struggled with texts of 25 words or fewer (He et al., MGTBench, 2023). Hand edits change the colors too, because each swapped word gets a new rank.
  4. Plain, clear writing is predictable too. A short service page ("We repair furnaces in Denver. Call for a free quote.") is mostly green whoever wrote it. The tool can't know that simple wording was a choice.
  5. Reading it takes skill. The authors say it "requires at least an advanced knowledge of the language to know whether an uncommon word does make sense at a position," and it handles "only individual cases," not large volumes (GLTR project page).

Can I still use GLTR?

Only by running it yourself. The GitHub code installs with Python and starts a local server on your computer (GitHub). That's a developer's job, and it still checks against GPT-2. For a business owner, it's useful as a lesson in how detectors think, not as proof.

Does a GLTR result matter for SEO?

No. Google judges the page, not the tool that wrote it:

"Our focus on the quality of content, rather than how content is produced, is a useful guide that has helped us deliver reliable, high quality results to users for years." (Google Search Central Blog, February 8, 2023)

A page full of green can rank; a page full of purple can fail. What counts is covered in Does Google penalize AI content? and AI content that Google likes.

Example

A roofer checking a freelancer's blog post

A developer friend runs GLTR on his laptop for a 600-word post. It's mostly green. That's no proof: plain roofing advice is predictable. Better check: does the post name local permit rules, show his crews' photos, and get the facts right?

A bakery owner checking her own writing

She writes her "About us" page herself in short, simple sentences. A detector would color much of it green. She keeps it. Customers read it on phones, and Google doesn't score how text was made.

A marketing agency promising "100% human" content

The agency shows GLTR screenshots full of red and purple as proof. The owner asks instead for sources for every claim and a named writer. Colors are easy to change by swapping words; checked facts aren't.

FAQ

Is GLTR free?

Yes. The code is open source under the Apache 2.0 license (GitHub). There's no paid version and no hosted demo today.

Does GLTR work on ChatGPT text?

Not reliably. It was built and tested on GPT-2 text, and its authors flagged that changed sampling could fool it. Its authors never tested it on today's chatbots, and edits change its colors.

What is the difference between GLTR and newer AI detectors?

GLTR shows word-by-word colors and leaves the judgment to you. Most newer detectors are trained on samples from current models and output a single percentage. Both make mistakes; see AI detectors and SEO.

Who made GLTR?

Researchers at MIT-IBM Watson AI Lab and HarvardNLP: Sebastian Gehrmann, Hendrik Strobelt and Alexander Rush. Their paper was presented at ACL 2019 in Florence, Italy (Gehrmann et al., 2019).

Different for you?

Keep learning

NextMake ChatGPT undetectable? Reddit tricks vs. what worksThe popular tricks to make ChatGPT text undetectable, checked against Google's rules and detector research: which help readers and which backfire.

Ask us one question

One line is enough. We reply by email.

Sources

  1. GLTR project page (gltr.io, via Internet Archive): Catching a Unicorn with GLTR
  2. Gehrmann, Strobelt, Rush: GLTR: Statistical Detection and Visualization of Generated Text (ACL 2019 demo paper)
  3. GLTR source code on GitHub (detecting-fake-text)
  4. Internet Archive: capture history of the GLTR demo page
  5. He et al.: MGTBench: Benchmarking Machine-Generated Text Detection (arXiv, March 2023)
  6. Google Search Central Blog: Google Search's guidance about AI-generated content (February 8, 2023)