Skip to main content

How AI Reads Your Words

Before AI can process your words, a tokenizer breaks them into pieces — like cracking a chocolate bar. Different models use different tokenizers, so the exact pieces will change.

Welcome to the Tokenizer

AI models like ChatGPT don't read English directly. They read tokens. Tokens are small chunks of text (letters, syllables, or whole words). Let's see how a model physically breaks sentences apart.

Sandbox: Break Your Own Words

Type a short sentence below. We'll run a simulated tokenizer to show how AI might chop it up into smaller chunks.

Notice: This is a simplified simulation. Real AI tokenizers use massive dictionaries to decide these splits.

Surprising Splits

Long words aren't always broken down logically. Click each word to see how a real AI tokenizer might snap it apart based on statistical frequency.

Notice how the capital 'S' in 'Summarize' changes how the word is split!

Context Matters

Select an example to see how different data types (like code or emojis) are handled.

Language Inequity

Because tokenizers are trained mostly on English text, non-English languages are often chopped into far more pieces. Hover over the bars to see inside.

Chart comparing 'What is artificial intelligence?' token counts: English 5, Spanish 10, Japanese 8.

More tokens mean slower processing and higher costs for non-English users.

How Tokenizers Learn: BPE