The archive.

Ask a general-purpose AI assistant to research the planning rules for a small community hydro or solar scheme in Wales, and it will produce a fluent, well-organised, confident answer. It will also, more often than not, be wrong. Not because the model is weak. Because the internet is not evenly written. England publishes more planning guidance, indexes it better, and gets linked to more often than Wales does, so an assistant reaching for "the rules" reaches for England's rules first, whether you asked for them or not.

That sounds like a narrow problem for a narrow industry. It isn't. It is the shape of the single most common failure in AI-assisted research, and it shows up wherever a business asks AI to look something up on its behalf: the nearest, best-documented answer quietly stands in for the one actually needed.

A trap with a name, once you have seen it

We recently helped a member put together a research brief for exactly this kind of work, mapping the consenting rules for community-owned hydro, solar and battery schemes across Wales. The brief opens with a single instruction, ahead of everything else: Wales is not England. Confirm every source applies here before you use it.

That instruction earns its place at the top because the failure it prevents is so easy to miss. The regulator most guidance assumes is the Environment Agency, which has no authority in Wales at all. The planning framework most guidance cites is the National Planning Policy Framework, which Wales replaced with its own years ago. A generation threshold that keeps appearing in older material, ten megawatts, was superseded by a fifty megawatt figure some time ago, and nothing marks the old number as retired. Each of these produces an answer that reads perfectly. Each is about the wrong place, or the wrong year.

Why the model does not just know better

An AI model reflects the shape of what has been written, not what is true everywhere. Wherever one version of a rule has been published more, linked more, and repeated more than its neighbour, the model will lean toward it, because that is what leaning toward the evidence looks like from the inside. This is not a Welsh problem, and it is not a planning problem. It is the same pattern behind an assistant that quietly assumes English employment law, sizes a market using a bigger competitor's numbers, or explains a tool using last year's version of it. Whatever is best documented wins the model's confidence, unless a person has told it, in advance, which contest it is not allowed to enter.

What actually fixes it

None of this is an argument against using AI for real research. It is an argument for briefing it properly, the way you would brief a very fast, very literal researcher who has read everything and understood none of the context. A few moves carry most of the weight.

Name the trap before you start. Do not assume good judgement will turn up unprompted. State plainly which neighbouring, plausible, wrong answer to refuse.

Decide what counts as evidence before you have seen an answer. A source hierarchy chosen in advance is honest. One chosen to justify an answer you already like is not.

Force structure, not prose. A fixed template with required fields, including explicit permission to write "unknown, here is what would confirm it", stops a fluent answer from papering over a gap.

Ask one narrow question at a time. A narrow brief produces something you can check line by line. A sweeping one produces something that reads beautifully and cannot be checked at all.

Verify separately. A second pass, run in a fresh session with no sight of how the first answer was produced, told plainly that its job is to find problems rather than to be helpful, catches what the first pass was too invested to notice.

The real opportunity

None of this is really about hydropower, or Wales. Every business that starts using AI for research that actually matters inherits the same problem, in whatever domain it works in. The businesses that end up with dependable answers will not be the ones with the cleverest prompts. They will be the ones who decided, before they asked the question, exactly what they were and were not willing to accept as an answer.

That is the quieter half of the question this note keeps returning to. It is easy to ask what work a machine could now do for you. It is more useful to ask what you would need to insist on, in writing, before you trusted its answer.