An assistant summarizes a forty-page report in six seconds and the summary reads well. That is exactly the moment to slow down. Before you act on it, you need to know what the tool actually read, what it was allowed to see, and whether anyone has checked the one claim that matters.
The application assembles the request, manages your account, selects files or search results and presents the answer. The model processes the supplied input and generates output. Infrastructure runs that computation. A chat product may use several models and outside services, so its capabilities and data boundaries extend beyond the model itself.
Names in this space are often overloaded. ChatGPT is a product built around GPT models, while names such as Claude and Gemini may refer to product experiences, model families or both depending on context. When comparing systems, specify the application, model and enabled features. A response with web citations may owe its freshness to search. A coding assistant may see repository files and tool output you never typed into the chat.
Five habits make everyday use more dependable:
- Check consequential claims against an authoritative source. Treat fluent language as a style, not as evidence.
- Establish what information was actually available. A model cannot reliably report a private fact it was never given.
- Inspect the relevant context. Old messages, retrieved text and conflicting instructions can change an answer.
- Distinguish a suggestion from an executed action. A displayed plan is not evidence that a business record changed.
- Identify the data boundary before uploading work material. Account type, enabled integrations and organizational rules matter.
- Assume the product has changed since you last used it. Models, features and defaults move without notice.
A useful first prompt. “Using only the attached public report, summarize the three operational risks most relevant to a service manager. For each, identify the supporting passage. Separate stated facts from your interpretation. If the report does not establish a point, say so.”
This request narrows the evidence and the decision. It still needs review. A citation can exist without supporting the claim beside it, and a confident interpretation can omit an important exception. Ask whether the answer would change if the missing exception were included.
Chapter 2 explains why context affects output so strongly. Later chapters add retrieval and tools, but neither removes the need to decide what counts as adequate evidence.