GLTR was never a product. It is a 2019 visualisation demo from the MIT-IBM Watson AI Lab and HarvardNLP, built around GPT-2 small, with no pricing, no accounts and no support desk. If you are looking for a GLTR alternative, you almost certainly want a maintained tool that does the job GLTR only illustrated.
Paste some text or load the example to estimate AI likelihood.
GLTR stands for Giant Language Model Test Room. Its project page credits Hendrik Strobelt and Sebastian Gehrmann, reviewed by Alexander Rush, as a collaboration between the MIT-IBM Watson AI lab and HarvardNLP, published as an ACL 2019 demo track paper (gltr.io, checked 16 August 2026).
The method is simple and elegant. The page explains that GLTR "has access to the GPT-2 117M language model from OpenAI" and analyses "what GPT-2 would have predicted at each position". Every word in your text is coloured by where it ranked in that model's prediction: green for the top 10, yellow for the top 100, red for the top 1,000, purple for everything else. Hovering a word shows the model's top five predictions with probabilities. Three histograms aggregate the picture across the whole passage.
The underlying assumption is stated openly on the page: "computer generated text fools humans by sticking to the most likely words at each position", whereas "natural writing actually more frequently selects unpredictable words". So a wash of green and yellow suggests generation, and a scatter of red and purple suggests a person.
It is one of the clearest explanations of how statistical AI detection works that anyone has built. That is the reason it still gets traffic seven years later.
Checked 16 August 2026, three things are worth knowing before you send anyone to the link.
Over plain HTTP, gltr.io serves the original project page, titled "Catching Unicorns with GLTR", last modified 29 August 2023. Over HTTPS, the host presents a TLS certificate issued for *.vizhub.ai that expired on 27 October 2025, and the content served on that port is a different project page titled "CSI framework". Browsers that upgrade to HTTPS by default will show a certificate warning or the wrong page.
The live demo at demo.gltr.io does still respond, and its backend still returns per-token rankings. Its project list endpoint returns exactly one model, gpt-2-small. It is reachable over HTTP only, with the same certificate problem on HTTPS.
The source repository, github.com/HendrikStrobelt/detecting-fake-text, is Apache 2.0 licensed, was created in April 2019, and its most recent commit is dated 27 June 2023. The repository is not archived and carries open issues (GitHub API, checked 16 August 2026).
None of that is a criticism. Research demos are not maintained products, and this one has outlived almost every commercial detector launched alongside it. But it does mean nobody is on the other end if it stops working.
For teaching, it is unmatched and nothing here replaces it. If you are explaining to a seminar, a newsroom or a faculty committee why detectors work at all, a coloured passage with hoverable token probabilities does the job in ninety seconds. A commercial tool that returns "83% AI" teaches nobody anything.
For interpretability, it shows its working in a way no commercial product on this page does, ours included. You can see the exact evidence, word by word, and disagree with it. That is a real property and it is rare.
For researchers, the Apache 2.0 licence means you can run it locally against your own model rather than GPT-2 small. The repository documents a backend API for plugging in custom models. If your question is methodological rather than operational, self-hosting GLTR against a current model is a more honest instrument than any vendor's black box.
And it costs nothing, requires no account, and stores nothing behind a login you have to trust. For a one-off look at a paragraph where privacy matters more than accuracy, that is a legitimate reason to keep using it.
GPT-2 small is the whole limitation. GLTR measures how predictable your text is under a 117 million parameter model from 2019. Modern systems generate text that GPT-2 small could not have predicted well, and modern human writing is not calibrated to it either. The tool was designed and evaluated before ChatGPT existed, which is not a flaw in the research and is a decisive problem for anyone using it to judge a 2026 submission.
There is also no score, no threshold, no report, no history, no file handling, no API for production use, and no one to email. GLTR asks you to interpret a picture. That is fine for a researcher and unworkable for anyone who has to justify a decision to a third party.
We are not going to publish a pricing table comparing GLTR to commercial tools, because GLTR has no pricing to compare. It is free, it is unsupported, and those two facts travel together.
| What you got from GLTR | Maintained option | Cost, as published |
|---|---|---|
| A picture of how detection works, for teaching | Keep using GLTR, over HTTP | Free |
| Token-level evidence you can inspect yourself | Self-host the repository against your own model | Free, Apache 2.0, your own compute |
| A defensible score on a modern draft | Any maintained detector, ours included | Free tiers exist across the field |
| Checking work in another language | Detecting AI, which publishes multi-language coverage | Free tier, $84/yr annual plan |
| A quick free box with no account | BrandWell's detector at 2,500 characters, or Ahrefs' at 2,048 | Free |
| Detection inside a documented workflow with export | Paid detector tiers | From $9.99/mo, ours |
Our detector is an English-only ensemble of DeBERTa, RoBERTa and ELECTRA, maintained and supported, with a documented false-positive bias against writers who learned English as a second language. That bias is the reason we say a score is an input to a judgement rather than the judgement.
Here is the honest concession to GLTR: we do not expose a per-token probability visualisation. There is no colour-coded prediction ranking here and no hoverable model internals. What we return is a score in three bands, under 30 "likely human-written", 30 to 69 "mixed signals", 70 and above "likely AI-generated", with sentence-level analysis from Starter upward showing which sentences drove the result. That is more actionable than GLTR's picture and less transparent than it. Both statements are true.
The box above takes 3 checks a day anonymously, pooled across the AI Detector and the Document Detector, capped at 1,000 words, with a 25-word minimum. A free account gives 60 checks a day at up to 5,000 words. Paid tiers add the operational parts: Starter at $9.99 a month or $7.49 annually brings sentence-level analysis, PDF export and 200 daily checks; Pro at $19.99 or $14.99 annual removes the daily cap and adds a Chrome extension, scan history and a REST API; Business at $39.99 or $29.99 annual adds teams and branded reports; Enterprise is $79.99 or $59.99 annual. File upload, URL scanning and bulk scanning run across the paid range.
If your interest in GLTR was intellectual, none of that is a substitute and you should keep the demo bookmarked. If your interest was operational, a maintained tool with a support address is the actual upgrade.
The deeper head-to-head: feature table, where GPTZero wins, where TextSight wins, and migration steps.
Read the comparePerplexity, burstiness, and transformer scoring explained, so you can read any detector's verdict critically.
Read the guideWhy detectors over-flag ESL and formal writing, and how sentence-level evidence helps you defend real work.
Read the guideThe full ranking with detection approach, pricing, and use-case fit side by side.
See the rankingStart with TextSight's free tier. No card, no signup, no commitment. Your first scan in about six seconds, with the evidence shown line by line.
Honest comparisons vs other tools.