Which of your words were yours?

You write. A small language model reads it one word at a time, and at every word it guesses what comes next. Where it guessed right, your word goes dark. Where it could not have guessed, your word stays lit. What stays lit, you walk out with.

Why this exists

For twenty years, much of the internet's business has been predicting people. Shoshana Zuboff described the product as behavioral data turned into "prediction products" and sold to "business customers with a commercial interest in knowing what we will do now, soon, and later."

Language models are prediction made fluent. They write the expected next word, very well, for everyone at once. Words drift toward what is expected. Martin Fowler named one version of this semantic diffusion in 2006: a word coined with a good definition spreads through a wider community in a way that weakens it. Kyle Chayka called the cultural version flattening. When models are trained on their own output, Shumailov and colleagues found, "tails of the original content distribution disappear." Whether the same thing happens to people who read mostly predicted text is an open question. This room is a place to notice it in yourself.

It uses the same kind of machine that predicts you, turned around, run on your own device, and then it shows you the door. In Plato's cave the prisoners face a wall of shadows cast by a fire behind them. Education, Socrates says, is not putting sight into eyes, but "turning the whole soul" toward the light (Republic VII, 518d).

Nothing leaves the room

• The model runs inside your browser. It is SmolLM2-135M (Hugging Face, Apache 2.0), converted to ONNX, run by transformers.js.
• The first visit downloads it once, about 136 MB. Your browser keeps it after that.
• There is no server behind this page. Its Content Security Policy lets it fetch from two places only: huggingface.co, where the model lives, and cdn.jsdelivr.net, where its runtime lives. Your text is never sent. To check, open your browser's developer tools and watch the Network tab while the fire reads.
• No analytics, no cookies, no account. Two things are kept, in your own browser only: the date you last walked out, so the room can tell you it will be here tomorrow, and, for each day you walked out, how many of your words stayed lit (numbers, never your words), for the calendar on the daylight screen. Your browser may clear them; Safari does after about a week away.
• The page also asks Chrome not to compute advertising topics from your visit (Permissions-Policy: browsing-topics=()). Google is phasing that system out. The header stays until it is gone.
• The model is pinned to one revision, and the software that runs it to one version. Both matter: run by different software, the same model ranked about half the words in a test differently. Pinned, the fire never changes under you, so today's reading can be compared with next year's.

What "saw it coming" means

• At each word, the model ranks every piece of its vocabulary, 49,152 of them, by how likely it is to come next. If your word was among its first three guesses, the fire saw it coming. It goes dark.
• Guesses four through twenty are grey. Beyond twenty, your word stays lit.
• Grammar (the, of, it, was, and about 160 more) is shown but never counted. Everyone's grammar is predictable.
• The first word is never counted. Nothing comes before it.
• Touch any word to see what the fire expected there, where yours ranked, and the odds it gave.

The daily card

Each day the wall opens on the same prompt for everyone, written by hand. The day turns over at midnight in New York. Your page works out which card is today's by itself, from the date; nothing is asked of a server and nothing about you is sent. Ask for another question and you get the rest of the library.

For scale

The number under your verdict is a share, not a grade and not a percentile: out of every 100 words that carry meaning, how many the fire saw coming.

The two lines beneath it come from HC3 (Guo et al. 2023): questions from Reddit's Explain Like I'm Five and from an open-domain question set, each answered once by a person and once by ChatGPT in December 2022. Several hundred of those pairs were run through this exact fire. HC3 stored the human answers with a scraper's spacing (it 's, word .); that was undone first, so people are measured as they wrote.

Length matters, so the room compares you with the first words of each answer, as many as you wrote. A short piece moves a lot: at 15 meaning words, one word shifts the share by almost seven points.

One register (people explaining things to strangers), one model, one vintage of ChatGPT. What you write here is probably none of those. A reference line, not a verdict.

What it cannot do

• It is not an AI detector. Do not use it to accuse anyone. Plain human writing can be predictable, and edited machine writing can surprise.
• It cannot tell you whether you meant it. Some of the truest sentences are the most predictable. "I love you" is in shadow. The fire shows where you were expected, not whether you were honest. That judgment is yours, and it is the point.
• Unpredictable is not the same as good. Nonsense stays lit. When a text is mostly noise, the room says so and gives no number.
• It reads English best. Writers whose first language is not English may see more of their words go dark, for reasons that have nothing to do with thinking for themselves. Detectors built on the same kind of measure have rated non-native English writing as more machine-like (Liang et al. 2023). We are working on it, and more languages are planned.
• This fire is small. 135 million parameters, trained on about 2 trillion tokens of educational web pages and code. A larger model would see more coming. Its habits are not everyone's.

Verbicide

"Verbicide, the murder of a word, happens in many ways. Inflation is one of the commonest." C.S. Lewis, Studies in Words (1960)

Struck-through words come from three places:

• Style words that Kobak and colleagues found in measurable excess across more than 15 million PubMed abstracts after LLMs arrived (delve, underscore, pivotal, meticulous and others).
• Lewis's own examples.
• A short list chosen by hand (journey, curate, authentic, passionate and others). These make no frequency claim. They are words that used to weigh more.

Touch one and it gives back what it used to mean. Etymologies are from the Online Etymology Dictionary.

Underlined phrases are discourse frames, counted in HC3: phrases like it's important to note, which ChatGPT used about sixty times more often than people did.

Sources

1. Plato, Republic VII, 514a to 520a; 518b to d.
2. Lewis, C.S. (1960). Studies in Words. Cambridge University Press.
3. Kobak, D., González-Márquez, R., Horvát, E.-Á., & Lause, J. (2025). Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Science Advances 11(27). arXiv:2406.07016. Word list: berenslab/llm-excess-vocab.
4. Guo, B. et al. (2023). HC3: Human ChatGPT Comparison Corpus. Hugging Face.
5. Shumailov, I. et al. (2024). AI models collapse when trained on recursively generated data. Nature 631, 755 to 759. doi:10.1038/s41586-024-07566-y.
6. Zuboff, S. (2019). Interview in the Harvard Gazette, on The Age of Surveillance Capitalism.
7. Fowler, M. (2006, December 14). Semantic Diffusion.
8. Chayka, K. (2024). Filterworld: How Algorithms Flattened Culture. Doubleday.
9. Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns 4(7). arXiv:2304.02819.
10. Allal, L. B. et al. (2025). SmolLM2. arXiv:2502.02737.
11. Online Etymology Dictionary, entries checked September 2026.

Back to the room