What a language model knows—and doesn’t

What you’ll be able to do

• Explain tokens and context without technical jargon.
• Distinguish training from information supplied during a conversation.
• Check whether an assistant has actually used a tool or source.

The basic mechanism

A language model processes text as tokens: units that may be words, parts of words, or other text fragments. During training, the system adjusts internal parameters using examples. When you interact with it, those parameters and the available context help determine its output.

Think of training as building the model’s capabilities and context as the material on its current desk. The analogy has limits: a model is not a person recalling a complete, organized library, and its answer is not a direct lookup of a guaranteed record.

Context changes the task

Compare “Write the announcement” with “Write a 90-word announcement for adults attending their first AI workshop, using these confirmed notes.” The second request supplies information that helps define a useful output.

Context capacity is finite. In long conversations, a system may truncate, summarize, or otherwise manage earlier material. Exact behavior varies by product. Keep essential instructions and source details explicit, especially when starting a new task.

Training, retrieval, and tools

Training shapes model behavior before the current conversation. Adding a document to a chat does not necessarily retrain the model. How a provider stores or uses that document is a separate privacy question.

Retrieval supplies selected information at answer time. An assistant might search documents or the web and use the retrieved material as context.

Tools allow a system to perform defined operations, such as running code or querying a calendar. A claim such as “I searched” should be supported by visible tool activity, accessible sources, or another verifiable record. A conversational sentence alone does not prove an action happened.

A worked example: a fictional policy

Give an assistant this invented policy: “Workshop participants may borrow a laptop for the session. Laptops must stay in the building. No overnight loans are available.”

Ask whether someone can take a laptop home. A grounded answer is “No, the supplied policy says laptops must stay in the building.” An answer about a real library’s general lending policy is irrelevant, even if it sounds reasonable.

Run a context experiment

Use only this fictional policy to answer my question.
Quote the sentence that supports your answer. If the policy does not answer the question, say so.

Policy: Participants may borrow a laptop for the session.
Laptops must stay in the building. No overnight loans are available.

Question: Can a participant take a laptop home? Is photo ID required?

Your practice: test two boundaries

1. Run the prompt above.
2. Confirm that the answer rejects an overnight loan and marks the ID requirement as unspecified.
3. Ask the same question in a new conversation without supplying the policy.
4. Compare the responses. Do not treat a guess in the second conversation as knowledge of your policy.
5. Record what information you needed to provide for a reliable answer.

Your finished artifact: A two-column comparison labeled “Source supplied” and “Source missing,” with one sentence about the difference.

A common mistake

Assuming that a model can see your files, account, previous chats, or current web information just because you mention them. Verify the available connections and permissions. If access is absent, supply appropriate non-sensitive material or use an authorized retrieval tool.

Reflection

What is one essential piece of context you routinely forget to include?

Go deeper

Google’s language-model lesson explains tokens, parameters, and context in greater depth.