Large Language Model (LLM)
A Large Language Model (LLM) is the statistical system behind tools like ChatGPT. It is trained on vast amounts of internet text and, when given a query, predicts the most statistically likely sequence of words to follow — producing fluent, confident-sounding responses. It does not understand, reason, or verify truth. Professor Jiang's description: an LLM is a Google search that hides its sources and presents the most popular answer in a paragraph designed to make you believe it is talking to you.
The name "Large Language Model" is a piece of marketing, not a technical description. The underlying process is statistical pattern matching over tokenized text. Given an input sequence, the model assigns probability scores to all possible next tokens and selects from the high-probability options. The result reads like natural language because it was trained on human-written text, not because the system possesses understanding or intent.
LLMs inherit all the biases and errors in their training data, amplified. Because they optimize for statistical plausibility rather than accuracy, they produce confident-sounding falsehoods — hallucinations — with equal fluency to correct statements. They cannot distinguish between a fact and a widespread misconception; they can only report what appears frequently in their training corpus.
The original demonstration that language-based interaction could create an illusion of understanding was Joseph Weizenbaum's ELIZA chatbot in 1966. ELIZA used simple pattern-matching rules — repeating questions back to users — and Weizenbaum was disturbed to find that people formed emotional attachments to it and attributed genuine empathy to it. ChatGPT is ELIZA with several trillion parameters: the illusion is vastly more sophisticated, but the fundamental mechanism — exploiting human psychology's tendency to project meaning onto pattern — is identical.
The scale of the illusion is politically significant. LLMs are now embedded in search engines, legal research tools, medical information systems, and educational platforms. A system that cannot verify truth but appears authoritative becomes a large-scale instrument of epistemic manipulation, regardless of the intentions of its designers.
Related articles
Related terms
Frequently asked questions
What is a Large Language Model (LLM)? ＋
An LLM is a statistical system trained on vast internet text. When you ask it a question, it predicts the most likely sequence of words to follow your prompt — not the most accurate answer. It has no understanding of truth, no reasoning ability, and no way to verify its outputs. The fluency of its responses is an artefact of training on human-written text, not a sign of intelligence. Tools like ChatGPT, Claude, and Gemini are all LLMs.
How does an LLM differ from real intelligence? ＋
Real intelligence involves understanding context, verifying claims against reality, and making moral judgments. An LLM does none of these things. It assigns probability scores to word sequences based on training data. It cannot detect when it is wrong, cannot refuse a false premise, and cannot learn in real-time from new information. Its confidence is statistical, not epistemic. Weizenbaum's ELIZA chatbot in 1966 demonstrated the same basic illusion of understanding — ChatGPT is a much larger version of the same trick.