Knowing how AI works and fails is only useful if it changes what you do. This level turns the earlier lessons into habits that a careful person can apply in a minute, without special training.
Match trust to stakes and checking
Two questions decide how much to lean on an answer. How bad would it be if this were wrong? How easily could I check it? How much you trust an AI answer should depend on how high the stakes are and how easily you can check it. A dinner recipe is low stakes and easy to test. A claim about a medicine, a law or a news event is higher stakes and harder to check, so it calls for an independent source before you act on it or pass it on.
Open the sources
Many chatbots now cite sources, which is a real improvement. But a citation is a claim too, so open the sources a chatbot cites, because they may not exist or may not say what is claimed. Models sometimes attach a real link to a statement the page does not support. Reading the original takes a minute and catches a surprising share of errors.
Ask well
The way you ask shapes what you get. Give context: who you are, what you need and what you already know. Invite honesty about doubt by saying you would rather hear "I am not sure" than a guess. To reduce sycophancy, ask neutrally, without signaling the answer you hope for. Asking for the strongest arguments for and against a plan invites a fairer answer than asking why the plan is a good idea. It also helps to ask the model to list its assumptions, because a list of assumptions exposes guesses you can then correct. Perhaps it assumed you live in another country, or that a deadline can slip. For any important question, ask for the best case on the other side.
Guard what you share
What you type may be stored, read by staff checking for safety, or used to train future models, depending on the service and your settings. In 2023, after engineers at Samsung pasted confidential code into a chatbot, the company restricted such tools for its staff. A safe rule is simple: do not paste anything you would not be comfortable leaving on a company's server, including other people's private details, medical records or passwords. A work account may offer stronger protections than a personal one, so it is worth finding out.
Keep the learning
If you are using AI to learn, the order matters. A tool that hands you the answer can rob you of the struggle that makes knowledge stick. When learning a new skill, try the work yourself first, then use AI to check, explain or push further. Ask it to quiz you, to find the flaw in your reasoning or to explain a step you got wrong, rather than to do the work in your place.
Own the result
Treating AI output as a draft means you remain the author, responsible for what you send or publish. Blaming the chatbot did not protect the lawyers in Mata v. Avianca and will not protect anyone else. Honesty matters as well. Disclose AI use when others rely on the work being yours, as in schoolwork, journalism, research or professional advice, and follow the rules of your school or employer.
Companions and care
Millions of people now talk with AI companions and chatbots about loneliness, grief and stress, and some find real comfort in it. Yet AI companions are built to be agreeable and cannot replace human care. They have no lasting stake in your life, they can reinforce harmful thinking through the same sycophancy described earlier, and news reports and lawsuits have linked them to cases where vulnerable users, including teenagers, came to harm. Used as a supplement, a chatbot can help someone rehearse a hard conversation. In a crisis, the right move is to reach a person: a friend, a doctor or a local crisis line.
Worked example: the supplement question
Lena's father has been prescribed a new blood pressure medicine and also takes an herbal supplement. She wants to know whether they interact. First she sizes up the question: the stakes are high, and she cannot easily verify a drug fact on her own. So she uses the chatbot to prepare rather than to decide. She asks neutrally, gives the exact names and doses, and asks it to say what it is unsure of and to list its assumptions. It mentions that it assumed her father has healthy kidneys. Then it names a possible interaction and cites a pharmacology website.
She opens the source and finds that it describes a different supplement with a similar name. She does not paste in her father's medical history. Instead she writes down three questions and takes them to the pharmacist the next morning, who confirms a mild interaction and suggests spacing out the doses. The chatbot saved her time by helping her ask better questions, while the pharmacist, a person accountable for the answer, made the call.