Can You Predict the AI?
An AI like ChatGPT writes by doing one thing over and over: it looks at the words so far and predicts the next word. So here's a puzzle worthy of this whole site — can you predict the thing whose entire job is predicting? Sometimes easily. Sometimes not at all. Let's find out where the line is.
🎮 The next-word game
Read the sentence, type the one word you think comes next, then see the AI's top guesses and how sure it is. Some sentences are easy. Some are basically impossible — and that's the point.
A sharp peak vs. a flat field
"The capital of France is ___" has one towering bar — one likely future, easy to predict. "The mysterious object began to ___" is a flat field of little bars — many possible futures, impossible to call.
That spread of possibilities is exactly what predictability is. One sharp peak = predictable. A flat field = not.
The predictability horizon
Predicting the next word is often easy. Predicting the next page is hopeless — the possible stories explode with every word, just like two weather forecasts pulling apart.
A sentence, a swinging pendulum, and a turbulent plasma all ask the same question: how far ahead can we see before the futures fan out?
🌫️ Watch the fog roll in
One word changes everything. Pick a first word for the story and watch how quickly the AI loses track of where it's headed — confidence high for the next word, then fading fast.
The next word is a fairly safe bet. A few words later, the AI is mostly guessing — and the whole rest of the story could go anywhere. That drop-off is the horizon.
🔬 Do the real experiment
Now try it on a real AI — ChatGPT, Claude, Gemini, or any chatbot. No advanced math required. (Teachers: each part is a ready-to-run prompt.)
Human vs AI
Paste this, then feed the sentence:
"I'll give you a sentence. Don't finish it — instead list the 5 most likely next words with probabilities."
→ The capital of France is
Did its #1 word actually show up when you then asked it to finish?
Your predictability score
For 10 prompts, write your guess for the next word before asking the AI. Score yourself out of 10.
Easy: The capital of France is
Medium: The dog ran into the
Hard: The mysterious object began to
Which were easy? Which were nearly impossible — and why?
The chaos test
Run both, save the outputs, compare:
Write a story about a dog.
Write a funny story about a dog.
How different are they by the first sentence? The first paragraph? The whole story? One word in → a whole new trajectory out. That's chaos.
Long-range forecast
Write a 500-word story about a robot.
Stop it after ~20 words. Predict the ending, the characters, the plot. Then let it finish. The next word was easy; the next page was the predictability horizon.
Quick, Draw!
Go to quickdraw.withgoogle.com and draw: cat · bicycle · house · airplane. How many strokes before Google guesses right?
After 1 line: many possibilities. After 10: almost only one. More information → easier prediction.
Rank your systems
weather · sports · the stock market · ChatGPT · Quick Draw · social media trends
Which is most predictable? Which is least? How far ahead can you see for each? What information would improve your prediction?
You just ran the same kind of experiment fusion and AI researchers run on turbulent plasmas and language models.
For teachers & grown-ups
The next-word game uses small, hand-built probability distributions rather than a live language model, so it runs entirely in the browser with no account, key, or internet call — but it faithfully shows the real mechanism: an LLM converts a context into a probability distribution over the next token and samples from it. "Predictability" here is just how peaked that distribution is (formally, its entropy). The deep link to the rest of Chaos Lab: text generation is an iterated map (each step's output feeds the next input), so a single different word is an initial-condition change whose consequences amplify — sensitive dependence, the same idea as the double pendulum and the logistic map. The "predictability horizon" of a chatbot, a weather model, and a fusion plasma are all set by how fast nearby possibilities diverge. The hands-on parts mirror a printable worksheet; the Quick, Draw! activity links information to predictability (more strokes → lower entropy → easier classification).