Short answer: generative AI is the chatbot kind — software that writes, drafts and explains in plain language by predicting the most plausible next words. In a distillery it’s genuinely excellent at language-shaped work: drafting the maturing-stock report, explaining a process, structuring a tasting note. It’s genuinely unreliable at remembering specific numbers, because to it a fabricated cask strength reads exactly as plausible as the real one. The whole game is grounding it in your documents so it quotes instead of invents.
This is part three of the foundations series. We’ve covered what “AI” means and machine learning. This post takes the form everyone has actually touched — the chatbot — and explains why it’s so useful and so untrustworthy at the same time.
A different machine entirely
Machine learning predicts a number from your data. Generative AI predicts words — the most plausible next word, then the next, building a fluent answer one token at a time. That single design fact explains everything about how it behaves. It is, in effect, an extraordinarily well-read colleague who has skimmed most of the internet, writes beautifully, explains patiently — and occasionally states something completely wrong with total composure, because composure is all it has. There is no internal truth-check. There is only plausibility.
For a distiller this is not a flaw to be outraged by; it’s a property to be worked with, the same way you work with any fast, fluent, occasionally-wrong colleague: listen carefully, then measure.
What it’s genuinely good at
Language-shaped work, grounded in your own documents:
- Drafting — the maturing-stock report, a board summary, a supplier email — from the data you give it. You edit; it does the blank-page labour.
- Explaining — “why does cask position change the angel’s share?” in plain terms, at whatever depth you ask for. Reasoning is the safest thing it produces because you can audit each step.
- Structuring — turn a warehouse manager’s rambling voice note into a clean tasting table; pull the fields out of a scanned COA into a row.
- Asking — point it at your actual ledger and ask “which casks enter their bottling window this quarter?” — and it reads, rather than recalls.
The thread: it shines when the work is about words and shape, and when the facts come from a document in front of it rather than its memory. This is the same pattern behind pointing a copilot at the COA instead of trusting its recall.
Why it makes things up
Because “this cask is 61.2% ABV” is exactly as plausible-sounding a sentence as “this cask is 58.4% ABV” — and the model picks the plausible continuation, not the true one. It has no way to know which is right unless the figure is in front of it. So a fabricated number arrives in the same fluent, confident prose as a real one. This isn’t lying; it’s the machine doing precisely what it was built to do. The danger is only that we read its confidence as evidence. It isn’t — telling the good answers from the confidently wrong ones is a discipline of its own, and the final post in this series returns to it.
Grounding: the one habit that fixes most of it
The single most effective move with generative AI is to stop asking it to remember and start making it read. Give it your actual cask ledger, your actual regauge log, your actual COA, and ask your question of that document. A model reading your real numbers has far less room to invent than one answering from a hazy global average. Grounding beats vigilance — it removes the opportunity to fabricate rather than relying on you to catch it.
Where this breaks
Honesty, as always. Grounding is not a force field — a model can still misread or misattribute a figure in a document, so a number that matters still gets a human check. Fluency masks shallowness — a beautifully written explanation can be subtly wrong, and the better the prose, the harder it is to doubt. It will agree with you too eagerly — push back on a correct answer and it may cave just to be agreeable, so a retraction is a flag to verify, not proof you were right. And the slow one: verification fatigue. After weeks of good drafts you stop checking the numbers — which is precisely when a wrong cask figure slips into a report that goes to the board. The discipline has to outlive the novelty.
The bottom line
Generative AI is the word-machine: superb at drafting, explaining and structuring, structurally unreliable at recalling specific facts. Use it for language-shaped work, ground every factual question in your real documents, and keep a human on anything touching compliance, safety or a measurement of record. Treat it as a brilliant, fast, occasionally-wrong colleague and it earns its keep daily. Next, the post that puts all three side by side: AI vs machine learning vs generative AI — which is which, and which one your problem actually needs.
Frequently asked questions
What is generative AI? Generative AI is the kind of AI that produces new content — text, tables, summaries, images — in plain language in response to a request. The chatbots everyone has tried are generative AI. It predicts the most plausible next words, which makes it excellent at drafting and explaining and unreliable at recalling specific facts, because a plausible-sounding number and a correct one look identical to it.
How can a distillery use generative AI? For language-shaped work grounded in your own documents: drafting the maturing-stock report from the cask ledger, explaining a process in plain terms, structuring a messy tasting note, or answering “which casks are in their bottling window?” when pointed at the actual data. It drafts and explains; a human owns anything touching compliance, safety or a measurement of record.
Why does generative AI make things up? Because it generates the most plausible continuation of text, not the true one — and from its point of view a fabricated cask strength reads exactly as plausible as the real one. It has no internal check for truth. The fix is grounding: give it your actual COA, ledger or report to read, so it quotes the document instead of inventing from memory.