It finishes sentences.
It doesn't know things.

Here's a two-minute test you can run tonight, free. Open one of the AI chat tools and ask it something you know cold — a question from your own trade, specific to your town. The kind of thing you'd bet a paycheck on. Then read the answer slowly.

Sometimes it'll be right. Run enough of these and you'll hit the other thing: an answer that's smooth, complete, confidently worded — and wrong. Not "I'm not sure" wrong. Wrong in the same even, professional voice it uses when it's right.

Your gut reaction — don't trust this thing — is the correct read. This page is what the machine actually is, in plain language, so that instinct turns into something better: knowing in advance where it'll be useful and where it'll bluff you.

The plain answer: it's a prediction machine

The plain answer first. An AI chat tool is a prediction machine. It was trained on mountains of text, and its whole job is to estimate what word most likely comes next — over and over, until it has a full answer. Google says so itself, in its own machine-learning documentation (last updated January 9, 2026, quote verified live July 20, 2026): "By modeling the statistical patterns of tokens, modern language models develop extremely powerful internal representations of language and can generate plausible language." A token is just a chunk of text, roughly a word or a piece of one.

Read Google's word choice there. Plausible. Not true. Plausible.

You already own a small version of this machine: the autocomplete on your phone. You type "running a little" and it offers "late." It doesn't know you're late. It has seen that pattern a billion times. The chat tools are that same idea with a vastly bigger engine — big enough that the finished sentences sound like a person who knows things.

But it holds patterns, not knowledge. There's no fact-shelf inside that it goes and checks — it generates from statistical patterns learned from training text, not by consulting a database of facts. Whatever it has picked up about your town or your shop is secondhand — swept in from the open web while it was training, frozen on that date, and never checked against what's true now. It isn't looking anything up. Ask it about your corner of the world and it can hand you a five-year-old price it once crawled off some page about your own business, in the same confident voice it uses for everything else. It isn't remembering. It's predicting what an answer about a place like yours would probably sound like.

Why it bluffs instead of saying "I don't know"

Here's the part the sales pitches skip — and it explains that confident wrong answer.

When the machine doesn't have the pattern, it doesn't stop. It produces the most plausible-sounding answer anyway — because plausible-sounding is the only thing it's built to optimize. The industry word for this is "hallucination," which is a fancy word for making things up. OpenAI researchers — OpenAI is the company behind ChatGPT — published a paper on September 4, 2025 titled "Why Language Models Hallucinate": "Like students facing hard exam questions, large language models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty." Their diagnosis of why: the training "reward[s] guessing over acknowledging uncertainty."

So the confident wrong answer isn't the machine failing. It's the machine doing what it was graded on. A smooth guess scored better than an honest "I don't know," so a smooth guesser is what got built.

The vendors say this themselves, in writing

You don't have to take our word for any of this. The companies selling these tools put it in their own documentation. Just not in the ads.

And xAI's own documentation, rendered live July 20, 2026, says of its Grok model: "Grok has no knowledge of current events or data beyond what was present in its training data."

That's the makers telling you the machine guesses and the machine is frozen in time. The fine print agrees with your gut — and it's written by the people who cash the checks.

Frozen in time — and the asterisk on the freeze

One more piece and you have the whole machine.

The mountain of text it learned from was gathered up to a certain date, then training stopped. Past that date, its knowledge doesn't stop dead — it thins out. The vendor's own fine print says so: the published date is where the model's knowledge is "most extensive and reliable," not a clean wall where everything after is simply gone. It's not even a single clean date. Anthropic's own model table, as of July 2026, lists two cutoff dates per model — a "reliable knowledge cutoff" and a broader "training data cutoff" — and both change every time a new model ships. So we don't print dates on this page. The cutoff board — the reference page in this manual that tracks each tool's freeze date — keeps the current ones, read live off each vendor's page, including the vendors that won't publish a date at all.

The asterisk: some tools can now search the live web before answering — Anthropic announced that for Claude back on March 20, 2025: "You can now use Claude to search the internet to provide more up-to-date and relevant responses." That bolts fresh information on top of the frozen machine. It does not change what the machine itself learned, and it does not remove the guessing habit.

Now you can predict it

Here's the payoff. Once you know it's a prediction machine holding patterns, you don't need a tech blog to tell you where it fails. You can call it yourself, the way you can call where a cheap fitting will leak.

So you can sort any AI job into two piles yourself. Pattern work is where it does real work: rewording something you already wrote, turning scattered notes into clean structure, adjusting the tone, drafting a shape you'll fill with your own facts. All of that is shape-of-language, no facts required — and shape of language is the one thing the machine genuinely holds, because that's what the patterns are.

Fact work is where it bluffs: prices and rates, names and phone numbers and addresses, codes and specs and anything you'd stake a job on, anything about your town or your shop, anything recent. That's specifics, recency, anything checkable about the real world — exactly the things a customer would catch, and exactly what the machine doesn't have.

Under both piles sits the standing rule from the hub board: the facts in anything it writes for you come from you, and you read every line before it goes anywhere near a customer. The machine drafts. You inspect. It never gets to be the source of a single fact, because it holds patterns, not facts it can vouch for.

That's the whole machine. It finishes sentences; it doesn't know things. Say that to yourself once before you use it — or before anyone tries to sell you something powered by it — and you'll predict its failures better than most of the people doing the selling.

Next: the cutoff board — what each vendor publishes about its freeze date, and which vendors publish nothing at all, checked live.