Concept · part of AI in the chain

Before trusting someone’s advice, you might ask who is whispering in their ear right now. An AI chatbot usually runs under instructions you never see: a hidden prompt from the company operating it, rules about what to refuse, limits on which tools it may use.

Deployment capture is the book’s name for tilting a model at this last stage. Nothing about the model changes. The same trained system is steered differently, request by request. Research on prompt formatting shows how far the wrapper around a question can move the output, and here the operator writes the wrapper. Its settings are the defaults every user gets without choosing, and the book expects a commercial operator to tune them for its revenue.

The book rates this the cheapest capture and the easiest to reverse: a different prompt or policy can fix it, if the operator wants to. Unlike slanted training data or a slanted reward, it shapes a surface and leaves the model alone. Its weakness is secrecy. Configurations are rarely published and can change silently between requests, so it is auditable in principle and opaque in practice. That is why the book’s institution-run model keeps deployment in the institution’s hands.

Read the full argument

How it connects


← Objective capture · AI in the chain (6 of 7) · LLM as capability extender (inside substrate custody) →

Explore all the ideas · Read the book in order · Open the interactive map