Ask an AI chat assistant for what you need, spot the ways a fluent answer goes wrong, check the claims that matter a second way, and know what to keep out of the chat. No technical background needed.
Learn the few ways a confident AI answer goes wrong, and a short routine for checking each one before you rely on it.
Written and reviewed with AI assistance, and checked by automated rules. No subject-matter expert has signed it off, and it is not a certification.
Learning outcomes
Preview
Read the complete course here. Reading does not create a private copy, save progress, or record learning evidence.
Preview material
An AI chat assistant writes the reply that most plausibly follows your message. It learned what plausible looks like from a very large amount of text, and it builds each reply piece by piece, every piece chosen to fit the ones before it. That is all the machinery this course needs, and it has one consequence that matters more than any other: an assistant is built to sound right, and sounding right is a different property from being right.
Three things feed an answer. The first is patterns from its training text, which stops at a cut-off date, so anything that changed later is invisible to it. The second is your conversation: your message, anything you pasted, and the earlier turns. The third is anything a tool hands it, such as a web search or a calculator, when the assistant has such tools and they are switched on. With no tool, it looks nothing up. It cannot see your files, your employer’s rules or the page you have open unless you put them in the chat or have connected them: an uploaded file, a feature that reads your browser page or files, or one that remembers earlier chats. If the answer shows no sign of a search and no links, or you are unsure, treat it as having no tool. Even with a search, open the page: the assistant can summarise it wrongly.
When those three sources hold what your question needs, the likeliest text is often the true text, because true statements are common in what it learned from. When they do not, the assistant still writes something, and the something has the same smooth shape. A precise number, a confident tone, a section reference and a tidy list are all features of fluent writing, so none of them shows that the content was checked. Even its wording of confidence is only more text, so it does not reliably follow whether the content is true.
Three smaller facts follow. Ask the same question twice and the wording, and sometimes the content, can differ, because the choice of each piece involves some chance. An assistant that is wrong will often explain its wrong answer well, since explaining well is the thing it does best. And assistants are tuned to give replies that people rate well, which helps explain a tendency you will meet in lesson 3: agreeing with you.
Preview material
Step 1. Ask what each answer rests on. Answer A rests on advice that appears in many places, and Ana can judge it by reading it and asking whether it would work for her team. Answer B can only rest on the hall’s own opening hours, and the assistant has no way to see them.
Step 2. Look for what would make B checkable. Answer B contains two claims, statements that could be true or false and that Ana could check: the 8:00 opening and the kitchen from 9:00. The hall’s website, a phone call or Ana’s booking confirmation each hold the real hours. The answer cites none of them. The precise times and the detail about the kitchen make it read as if one had been used.
Step 3. Act on the difference. Ana treats A as a starting point she can judge for herself. She treats B as a claim to check before she tells anyone to arrive at 8:00. If the assistant had replied “I cannot see the hall’s hours; check with the hall”, it would have done the useful thing. If Ana pushes with “just give me your best guess”, she gets a guess in the same voice, and the push changes nothing about whether it is true.
Result: the difference was in the question, not in the writing. Both answers read the same, and only one of them rested on anything the assistant could see.
Check a claim a second way, keep sensitive material out of the chat, know when a professional comes first, and turn it all into a routine.