This Word Does Not Exist: the website that invents fake words

This Word Does Not Exist is a website that invents a dictionary word that has never existed, writes a straight-faced definition for it, and drops it into an example sentence. It is worth a look because the results are uncanny: the words look like English, read like English, and mean nothing at all.

The whole page is one word at a time. A headword sits at the top, tagged with a part of speech, followed by a definition and a usage example, exactly as a print dictionary would lay it out. A button fetches the next invention. There is nothing to win and nothing to install.

In brief

• Every entry is generated by a neural network, not written by a person, and the words are filtered so they are not real.
• It was built by former Instagram engineering director Thomas Dimson and launched in May 2020.
• It runs on GPT-2, the OpenAI language model, trained to imitate the shape of dictionary entries.
• The source code is public under an MIT license, so anyone can inspect how the fakes are made.

What the website actually does

Open the site and a single word appears. During one visit it offered prebendian, tagged as an adjective, defined as "having a branch or branch crest that is partly pointed upward," with the example "a prebendian temple." None of that is real. The word is not in any dictionary, the definition describes nothing, and the example sentence is a decoy.

A control labeled "New word" asks for another. Each press returns a fresh invention within a second or two. The layout never changes, which is the point: the format is borrowed wholesale from lexicography, so the eye trusts it before the brain catches up.

There is no scoreboard, no account, and no feed to scroll. The site does one thing and stops there, which is the quiet discipline that links it to the rest of the single-purpose web. Every element on the page exists to present one invented word as convincingly as possible, then get out of the way so the next one can arrive.

You can also feed it your own

There is a "Write your own" option that lets a visitor type a made-up string and watch the model define it on demand. Enter a nonsense cluster of letters and the network treats it as a serious lexical problem, assigning a part of speech and composing a definition that sounds like it belongs in a reference book. The machine never refuses. It has no concept of a word being too strange to define.

That mode is where the illusion gets personal. A reader can invent a string that means something privately, a nickname or an inside joke, and receive an official-looking definition for it in seconds. The result feels like a verdict, even though the model is only guessing at what such a word might mean if it existed.

The person who built it

The site is the work of Thomas Dimson, a software engineer who had been an engineering director at Instagram before turning to side projects. According to Nerdist, which covered the launch, the project went public on 13 May 2020. Dimson credits collaborators in the acknowledgements, and the code lists the people who helped shape it.

What stands out is the restraint. The site sells nothing and explains almost nothing on the surface. A visitor who wants to know how it works has to go looking, and the trail leads straight to the source code rather than a marketing page. The project reads less like a product and more like a demonstration left running in public.

How a machine learns to make up words

The engine is GPT-2, a language model released by OpenAI. Per Wikipedia, the full 1.5-billion-parameter version was released on 5 November 2019, and it was trained on WebText, a corpus of roughly 8 million documents scraped from pages linked on Reddit. GPT-2 is autoregressive: it predicts the next token in a sequence, one piece at a time.

A model that only predicts text does not know what a word is. It knows what dictionary entries tend to look like. Dimson fine-tuned GPT-2 on actual dictionary data so that, when prompted, it continues the pattern of headword, part of speech, definition, and example. The output is not retrieved from a list. It is improvised, token by token, in the style of a reference book.

GPT-2 carried a small reputation for danger when it arrived. OpenAI released it in stages through 2019, holding back the largest version at first out of concern that fluent machine text could be misused at scale. Seen from today, a generator of imaginary dictionary words is a wonderfully low-stakes use of a model that was once treated as too convincing to publish all at once. The threat model aged into a toy.

The dictionary as a training format

A dictionary entry is unusually regular text. It has fixed slots in a fixed order, and the same shape repeats across tens of thousands of examples. That regularity is a gift to a language model, because the pattern is easy to learn and hard to break. Once the model has seen enough entries, producing a new one is just a matter of filling the slots in sequence, the same way it would continue any other sentence it had learned to imitate.

The example sentence is the sly part. A definition alone can hide behind vagueness, but a word used in a sentence has to behave grammatically, agreeing with its claimed part of speech. When the model tags a word as a verb and then conjugates it correctly in the example, the entry gains a second layer of credibility that most readers never consciously check.

Built on open tooling

The project uses the Hugging Face Transformers library to load and run the model, the same toolkit that powers a large share of applied language-model work. That choice is why the repository is approachable: the heavy machinery is a dependency, and the interesting part is the training setup and the filtering around it.

Two models, pointing in opposite directions

The clever structure is visible in the public repository. It ships two models. A forward model takes a word and produces a definition. An inverse model runs the other way, taking a definition and producing a word to fit it. Together they let the system start from either end and meet in the middle, which is how the "write your own" feature defines a string the model has never seen.

The repository also describes a blacklist that filters generated words against real ones. That step matters more than it sounds. Without it, the model would keep reinventing words that already exist, and the site would quietly become an ordinary, error-prone dictionary. The filter is what guarantees the promise in the name: the word really does not exist.

Open to copy

The code is released under an MIT license and has gathered roughly a thousand stars on GitHub, along with dozens of forks. The repository includes scripts for pulling training data from dictionary sources, including the system dictionaries that ship with macOS and a scraper for Urban Dictionary. A companion account, @robo_define, posts fresh inventions on its own.

For anyone who wants to run it rather than just read it, the project documents downloadable pretrained models, hosted on Google Cloud Storage, and supports inference on either a processor or a graphics card, with quantization options to shrink the model for slower hardware. None of that is required to enjoy the website, but it is the difference between a magic trick and a method. The trick is on the page; the method is in the repository, in full view.

Why the fakes feel so real

English is full of productive patterns. Latin and Greek roots, prefixes, and suffixes combine in predictable ways, which is how real coinages like podcast or blog ever caught on. A model trained on enough definitions absorbs those patterns and reassembles them. The invented word sclerotoxin looks medical because sclero- and -toxin are familiar fragments doing familiar jobs.

Nerdist put the effect plainly, noting that every word the tool produces "sounds like it both is, and isn't, a real word." The definitions cooperate. They use the cautious, hedging register of lexicography, which is exactly the tone that makes a claim feel authoritative even when it is empty.

There is a human reflex behind the trick. Readers judge whether a string is a possible word long before they check whether it is an actual one, and that sense of wordlikeness is driven by sound and shape, not meaning. The model is tuned to satisfy exactly that reflex. It assembles letters into shapes the eye has been trained to accept, so the first impression lands as plausible and the second thought, the one that asks what the word means, arrives too late to spoil it.

The site is honest about the limits. A note on the page warns that the words are not reviewed and may reflect bias in the training data. That caveat is doing real work: a model that learned English from the open internet learned its blind spots too, and an unreviewed generator will surface them.

How real dictionaries do the opposite

The site is funniest when read against how words actually enter a dictionary. Lexicographers do not invent words, they document them. A term earns an entry only after editors gather evidence that real people use it, in print, over time, with a stable meaning. The definition follows the usage. It is written last, as a summary of behavior already observed in the wild.

This Word Does Not Exist runs that process backward. It produces the artifact of settled usage, a polished entry, with no usage behind it at all. There is no community of speakers, no first citation, no drift in meaning over decades. The entry arrives fully formed and completely hollow, which is why it can be generated a thousand times a minute while a single real dictionary entry can take years of watching to justify.

Part of a larger family

This Word Does Not Exist belongs to the "does not exist" genre that spread across the web around 2019 and 2020, a wave of single-serving sites each demonstrating one generative model. The best known conjures human faces, using a generative adversarial network to produce portraits of people who never lived. Others produce invented cats, rental listings, or resumes. A catalog site at thisxdoesnotexist.com keeps a running index of the format.

Words are a quieter entry in that lineup, and in some ways a sharper one. A fake face is impressive but hard to interrogate. A fake word comes with its own argument, a definition, which the reader can test against meaning and watch collapse. The failure is legible, and that legibility is the fun.

It sits comfortably beside the other deadpan, single-purpose projects in the Boring Websites network, the kind of site that commits fully to one small idea and refuses to explain the joke. The commitment is the joke.

What it is actually good for

Beyond the novelty, the site is a clean teaching example. It shows, in a form anyone can poke at, what a language model does and does not understand. The grammar is near perfect. The morphology is plausible. The meaning is absent. That gap, between fluent form and missing substance, is the single most useful thing to grasp about modern text generation, and here it is rendered as a toy.

Writers and gamers have found uses too. The inventions make serviceable names for characters, bands, spells, and startups, since they are designed to sound like real words without colliding with trademarks. A tabletop group short on fantasy vocabulary can mine the "new word" button for an evening. The output is free of the obligation to mean anything, which is precisely what makes it adaptable.

There is a gentler use as well. Spend a few minutes pressing the button and the mystique around machine-generated language quietly deflates. The words stop feeling like oracles and start feeling like what they are, confident guesses with nothing underneath. For a reader trying to build intuition about where these tools help and where they mislead, that small deflation is worth more than most explainer articles.

The takeaway

This Word Does Not Exist takes one of the stranger capabilities of machine learning, fluent nonsense, and frames it with the straightest face available: the dictionary entry. It is funny, briefly unsettling, and genuinely instructive, all from a single button. Visit thisworddoesnotexist.com, press for a new word, and try to decide, before you read the definition, whether the thing in front of you is real.

FAQ

Who created This Word Does Not Exist?

It was built by Thomas Dimson, a software engineer and former engineering director at Instagram. He released it in May 2020 and published the source code on GitHub, where it is maintained under the handle turtlesoupy.

How does the site generate the words?

It fine-tunes GPT-2, an OpenAI language model, on dictionary data so the model learns to continue the pattern of a headword, part of speech, definition, and example. A separate step filters the results against real words, so the output stays genuinely invented.

Are the words and definitions ever real?

By design, no. The system uses a blacklist to screen generated words against existing ones, which is what lets the site promise that each word does not exist. The definitions are improvised to fit and describe nothing in the real world.

How is it different from This Person Does Not Exist?

Both belong to the same "does not exist" family, but they demonstrate different models. The faces site uses a generative adversarial network to synthesize images, while This Word Does Not Exist uses a GPT-2 text model to synthesize language. Words are easier to fact-check than faces, which makes the fakery more obvious and more fun.

Can I use the invented words for names or projects?

Yes. Many visitors mine the generator for character names, band names, and product ideas, since the words are built to sound plausible without matching anything that already exists. The code is open under an MIT license if you want to run your own version.