DennisKennedy.Blog

Building the Stochastic Sandpit for AI

The article proposes a "stochastic sandpit" as a thinking workspace where generative AI is used for exploration rather than as a vending‑machine answer engine. It contrasts two usage modes: insurance mode, which enforces tight guardrails for compliance and predictability, and sandpit mode, which encourages productive wrongness to surface hidden assumptions, edge cases, and alternative frames. By deliberately allowing the model to be imperfect, users can generate "impressions"—traces of reasoning that reveal uncertainty and leaps. The sandpit approach is presented as a protocol for high‑stakes fields such as law and policy, ending with a hard rule that only human‑rewritten, source‑anchored output reaches production.