How to minimise bias and hallucination — and why it matters for the discoverability of your prompts
Extract from Melina Efstathiou’s article, “The Promptosphere: How to minimise bias and hallucination — and why it matters for the discoverability of your prompts.”
Online culture has given us the manosphere and the femosphere — internet ecosystems with their own dialects, grievances and bad habits. Legal AI now has its own: the Promptosphere. And in May 2026, a federal court gave practitioners a very clear reason to start taking what happens inside it seriously.
The recipe, not just the dish
On 18 May 2026, Magistrate Judge Thomas Farrish of the US District Court for the District of Connecticut ordered the plaintiff in Conservation Law Foundation v. Shell Oil Co. to produce the actual AI prompts its expert witness had used in preparing her expert report. This appears to be the first federal decision treating AI prompts as discoverable methodology in their own right — not as a peripheral curiosity, but as squarely within the scope of Rule 26. The court rejected the argument that a stipulation protecting “expert notes, drafts or communications” was broad enough to shield the prompts, holding that the language simply was not “quite clear” enough to cover them specifically.
The line from the commentary that followed has stayed with me: If an expert’s analysis relies on AI, expect to turn over the recipe, not just the dish.
