Does AI Really Think? The Guessing Game Behind ChatGPT, Claude and Gemini

Key takeaways

• AI chat models write one small piece of text (a token) at a time: for every next piece they give each option a score, pick one, and repeat, so a one-page answer is hundreds of guesses in a row.
• They learn by playing fill-in-the-blank on a huge library of text: GPT-3 (2020) learned from about 300 billion tokens, roughly 5,000 years of reading for one person at 8 hours a day, and Meta said in 2024 its new model learned from over 15 trillion.
• Inside, words become points on a map of meaning, and the Transformer's attention (2017) lets each word look at the others; Anthropic's 2025 research found a Claude model can plan a rhyme word before writing the line, so the guessing has rich structure but is still not human thinking.
• Confidence is not proof: AI can invent names, numbers, books and court cases in the same smooth voice as true answers. In 2023 a New York lawyer was fined after filing at least six cases ChatGPT made up, and questions about Cambodia in Khmer need extra checking.
• Five rules: check important facts in a second source, open the sources yourself, ask clearly and let it say 'I don't know', never paste private information, and think for yourself first.