chapter six

6 All that glitters is not smart

 

This chapter covers

  • Analyzing why language models produce hallucinated knowledge.
  • Interpreting the probabilistic nature of model responses.
  • Evaluating how safeguards and jailbreaks reshape model outputs.
  • Assessing how context limits shape model responses.

Large language models often produce responses that feel precise, informed, and closely aligned with the questions they receive, creating the impression that the system understands the request and assembles a reasoned reply. In many interactions, this impression holds, with outputs that appear confident, coherent, and consistent with the task.

Yet this surface reliability can crack in ways that are difficult to anticipate. The model does not reason over a stable store of facts or apply rules in the way traditional software might. Instead, each response emerges from a process that predicts the next fragment of text based on patterns learned from vast collections of human language and on the information available at that moment. Within that process, facts, instructions, examples, retrieved evidence, and prior context all become signals that can reinforce, weaken, or displace one another.

6.1 Sorry, are you hallucinating?

6.1.1 When fluency outruns grounding

6.1.2 Optimization without truth

6.1.3 Extrapolating beyond experience

6.1.4 When fabrication becomes operational risk

6.2 Only facts, please

6.2.1 Constraining generation

6.2.2 Monitoring confidence

6.2.3 Reliability without guarantees

6.3 Breaking the rules

6.3.1 Rules without rule-keeping

6.3.2 The impact of small details

6.3.3 The difficulty of improving prompts

6.3.4 Tilting the balance

6.3.5 Containing instability

6.4 A budget for memories

6.4.1 Where the answer comes from

6.4.2 A window, not a memory

6.4.3 Present but not decisive

6.4.4 Borrowed memory

6.4.5 Retrieval is not grounding

6.5 Summary