# AI in Plain English

By [DYLIT Chronicles](https://dylit.info/user/dylitmediabuzz)

[Everything AI - beyond the hype](https://dylit.info/pr/everything-ai-beyond-the-hype/6a9efac02e92664f4d50cf9d) > [AI in Plain English](https://dylit.info/ch/ai-in-plain-english/6ab3536999638e119f592c02)

What AI Is Actually Doing When It Answers You Almost everything confusing about AI stops being confusing once you understand one thing about how it works. The systems people mean when they say AI today, the ones you type a question into and get paragraphs back from, were built to do a single job. Read a stretch of text. Guess what word comes next. Then guess the one after that. Then the one after that, and keep going until the answer looks finished. That is the mechanism. There is no second system underneath doing something cleverer while the word-guessing handles the surface. People resist this, and reasonably so, because the output does not feel like guesswork. It feels like understanding. So it is worth sitting with the idea for a minute, because once it lands, six or seven other things fall into place at once. How guessing words produces a memo During training, the system was shown an enormous quantity of written text. Books, code, articles, manuals, arguments, transcripts. Far more than any person could read in a thousand lifetimes. Across all of it, the system learned which words tend to follow which other words, in which contexts, at a level of detail that is genuinely hard to picture. Not "the word the is common." More like: in a paragraph that has been discussing a contract dispute in formal English, after the phrase "the parties agree to," these are the continuations that are statistically plausible, weighted by everything else in the passage. Stack enough of that together and you get grammar. Then tone. Then structure. Then something that looks a great deal like reasoning, because human writing contains reasoning, and reproducing the shape of reasoning convincingly requires having absorbed a lot of it. What you do not get is a lookup. When you ask a question, the system is not searching a database of facts and retrieving the answer. It is producing text that fits the pattern of a correct answer to that question. Very often those two things land in the same place. Sometimes they do not, and nothing in the mechanism flags the difference. What this one fact explains Why it is so good at drafting. A first draft is a text-shaped problem. You need something in the right register, with the right structure, covering the expected points. Producing plausible, well-formed text is precisely what the system does. Drafting is not a weak application of the technology. It is the center of it. Why it invents things. A made-up citation and a real one look identical from the inside of the mechanism. Both are plausible continuations. The system has no separate check that asks whether the journal exists. When it produces a case name, a statistic, or a date that nobody ever wrote down, it is not malfunctioning. It is doing exactly what it does, in a situation where the plausible answer and the true answer came apart. Why it sounds so certain. Hesitation is a pattern in text like any other, and the system reproduces confident prose because confident prose is what most of its training material looks like. Tone is not connected to accuracy anywhere in the machinery. A wrong answer and a right one are delivered in the same voice. Why it knows nothing about your company. Your pricing decisions, your customer list, last quarter's actual numbers, the reason the Leeds project got cancelled: none of that was in the training text. If you did not put it in the conversation, it is not there. The system will still answer, because answering is what it does. It will fill the gap with whatever is plausible for a company like yours. Why the same question twice gives two different answers. Each response is generated fresh. There is deliberate variation in how the next word gets chosen, which is what keeps output from being flat and repetitive. Ask again and you get a different walk through the same territory. What it does not have It holds no beliefs. It has no intentions. It is not aware of you, and it does not remember you between conversations unless a specific feature is carrying that forward. When it opens with "Great question," that is not enthusiasm. It is a phrase that frequently follows a question in helpful-sounding text. Same with the apologies when you correct it, and the confidence when you do not. None of this makes the tool less useful. It does change what you are doing when you use it. You are not consulting an expert. You are operating a very capable text engine, and the quality of what comes out depends heavily on what you put in and how carefully you read the result. The part that catches people Here is the uncomfortable bit. The system is genuinely excellent at this. Not passable. Excellent. The prose is clean, the structure is sensible, the tone adapts to what you asked for. That quality is exactly what makes it risky. Humans have spent our entire lives using fluency as a proxy for competence, because for humans it usually works. Someone who explains a topic in clear, well-organised language probably understands it. That heuristic has served us well and it fails completely here. Fluency is the one thing this technology is guaranteed to deliver. Accuracy is a separate question that the mechanism does not address. So the instinct to relax when the answer reads well is precisely backwards. The better it sounds, the more deliberately you have to check it. What to do with this Three habits follow directly from the mechanism, and they will carry you further than any list of tips. Supply the context. Whatever the system needs to know and could not possibly know, put it in the message. Background, constraints, the actual figures, an example of what good looks like. Missing context does not produce an error. It produces confident invention. Treat every specific as unconfirmed. Names, numbers, dates, citations, quotes, URLs. The reasoning and structure of an answer usually hold up well. The details are where it falls apart, and the details are what get quoted back at you in a meeting. Use it where you already know the answer's shape . If you can recognise a good result when you see one, you are in the right territory. If you are asking about something you cannot evaluate, you have no way to catch a confident mistake, and the mechanism guarantees there will eventually be one.