GenAI Field GuideReference
Myths, busted
Most misuse comes from a handful of wrong mental models. Skim this, or use it as a lookup; every row links to its page.
| Myth | What is actually true | Read more |
|---|---|---|
| "It remembers our last conversation." | It doesn't, unless a memory or project feature re-sent that text into this prompt. Open a fresh chat and ask. | Projects |
| "The prompt is what I typed." | The prompt is everything the model receives this turn: preamble, tools, memory, history, then you. | What's in the prompt |
| "1M tokens means it can read my whole archive." | It can hold it. Material in the middle of a long context is used less reliably. | Context window |
| "I gave it the whole document, so it has all the facts." | It has all the tokens. Whether it uses the one on page 214 correctly is a separate question. Check what you will sign. | Documents |
| "It's all still in the conversation." | After compaction the model sees a summary of the older turns, not the turns. | Compaction |
| "It's on GenAI.mil, so the output is approved." | IL5 clears the platform to hold CUI. It says nothing about whether the answer is right. | Where your text goes |
| "The best model wins." | Fit-for-task plus verification wins. A cheaper model with the right document beats a frontier model without it. | Which tool |
| "There's a magic phrase." | Source, permission to abstain, checkable citations, small steps, second look. That is the toolkit. | Prompting |
| "Telling it 'don't hallucinate' helps." | No evidence it changes the error rate meaningfully. It changes the tone. | Hallucination |
| "RAG makes it accurate." | RAG makes it checkable and reduces the error rate. It does not remove it. | RAG |
| "It searched the web." | Only if the transcript shows the tool call. Otherwise it answered from training. | Tools, MCP, skills |
| "Agents are autonomous." | Autonomous inside a loop a human designed, with tools a human granted, until a stop condition a human set. | Agents |
| "The subagent said it succeeded." | The parent only knows what the child reported. A quiet failure can report success. Ask for evidence. | Subagents |
| "It worked last time." | After compaction it may remember that it worked and not why the earlier attempt failed. | Coding sessions |
| "It's the tool's fault." | Accountability doesn't move. The signature is yours. | Verification |
| "Newest is best for everything." | Newest is usually best at hard reasoning. For ordinary work the fast tier with the right document attached is as good and quicker. | Choosing a model |
| "Thinking mode fixes hallucination." | It reduces reasoning errors. It does not add missing facts. | Thinking modes |
| "AI today is basically all LLMs." | The chat assistants are LLMs. Image, video, and voice tools are cousins built on different model families. | What an LLM is |