If you want to know how to ask AI a question and get a usable answer back, give it three things: what the question is about, what you want to do with the answer and what form you want it in. Leave those out and you get a generic piece of text that applies to everyone and fits no one.
You don’t need an account or a subscription to get started, and you don’t need to be a prompt engineer. Below, you’ll learn how to phrase a question, how to steer when the answer is off and how to check whether you can actually use it.
Can you just ask AI a question?
Yes, you can ask AI a question for free and without installing anything. ChatGPT, Claude, Gemini and Copilot all have a free version where you can start typing straight away. Most of them do ask for an email address and an account, but you don’t have to pay.
Where to start without paying
Without paying, the big three take you furthest. ChatGPT is the best known and the most all-round. Claude is more pleasant to work with when you paste in long pieces of text. Gemini is built into Google’s services and comes in handy when your question involves search.
The free versions are not stripped-down demos, but they do have limits on usage, on file uploads and on the more capable models.
What you shouldn’t type into an AI tool
Not everything belongs in an AI tool, not even in a free version. Leave out personal data about customers or colleagues, medical information, passwords, client files and commercially sensitive figures.
The reason is simple. With consumer versions, what you enter may be used under the terms of service to improve the model. If you still want to work with a real document, first remove names, addresses and anything else that could identify someone.
How do you ask AI a question that gets a useful answer?
You get a useful answer from AI when you don’t make the model guess what you mean. The difference between a mediocre and a good answer rarely lies in clever wording and almost always in how much context you provide.
The four building blocks of a good question
A good question has four building blocks. First, the role the model should answer in, for example as a policy adviser or as a teacher. Second, the goal: what you’re going to do with the answer. Third, the context: everything the model can’t know but does need. Fourth, the format, for example the number of words or the layout you want.
The role is often skipped, even though it has the most influence on the answer: an explanation for a colleague looks different from an explanation for a customer.
A weak and a strong question side by side
The difference between a weak and a strong question becomes clear once you put them side by side. A weak one is “write something about sustainability for our newsletter”. The question doesn’t say who it’s for, how long it should be, what tone to use or what is already known, so the model fills all of that in itself.
A strong one is: “You are writing for our monthly customer newsletter, read by purchasing managers in the construction industry. Write 150 words about our switch to recycled packaging material. Businesslike tone, no superlatives, end with an invitation to reply. The facts are below.” Then you paste the facts underneath.
The strong question gets a better answer in every tool, whichever one you use. What you provide matters more than which model you choose.
Build your question from the four building blocks
Fill in the role, goal, context and format. Your prompt grows as you type. The text stays in your browser; this page sends nothing.
Your prompt
How do you ask ChatGPT a question?
You ask ChatGPT a question with the same four building blocks: role, goal, context and format. What you type into the input field is called a prompt, so everything above applies in ChatGPT too. If you’d rather not start from scratch, ChatGPT prompts with a fixed structure are a good starting point that you add your own context to for each task.
ChatGPT also has three places where you can store context, so you don’t have to type it again in every prompt:
- Settings: under personalisation you set down who you are and how you want answers, for example businesslike and without bullet points. ChatGPT takes this into account in every new conversation.
- Projects: in a ChatGPT project you keep fixed instructions and files together, which is handy for a client or a recurring task.
- Attachments: upload a document instead of pasting the text, and refer to it in your prompt, for example “use only the figures from the attachment”.
If you’re not sure exactly what you need, let ChatGPT ask you questions first. End your prompt with “ask me three questions before you start”. That way, the context you forgot still comes to the surface.
How do you steer when the answer is wrong?
If the answer is wrong, you steer it in the same conversation instead of starting over. Within a conversation, the model keeps track of what was said earlier, so you build on what is already there.
Follow up instead of starting again
Following up works better than rephrasing the question and sending it again. Say what is going wrong and what you want instead: too long, too vague, the wrong tone, or an assumption that isn’t correct.
Feel free to ask for an explanation too. “Where did you get this figure from?” or “which part of this did you infer rather than take from a source?” often shows within one sentence where the weak spot is.
What to do if the model keeps drifting off
When the model keeps drifting off, the conversation has usually become cluttered with earlier instructions that contradict each other. In that case, do start again, but give everything you learned in the previous conversation in one go.
If it keeps going wrong, the task is probably not suitable for AI. A question where the model has to supply the facts itself won’t get better by phrasing it differently.
How do you check the answer before you use it?
You check an AI answer before you use it, because a wrong answer looks just as convincing as a right one. A language model predicts likely text and does not test whether that text is true.
The three types of output that always need checking
Three types of output always need checking: figures and statistics, especially without a source; names of people, organisations and products; and references to legislation, article numbers, years and dates.
Those are exactly the three things a model slots effortlessly into the pattern of a believable sentence, even when they don’t exist.
The intern question as a final check
The intern question is the quickest final check before you send something: would you believe this if an intern had handed it in, or would you check it? That question forces you to judge the content rather than the form.
For work that goes out under your own name, that is the minimum. Smooth wording is precisely what a language model always gets right, even when the content rests on nothing.
Which questions should you not ask AI?
Some questions are better not put to AI, and that has nothing to do with how you phrase them. These are tasks the model was never built for.
Questions where the model has to supply the facts
Questions where the model has to supply the facts itself carry the greatest risk. “What are the recent changes to this law?” or “which studies show this?” lead to answers that are right in form but not in substance.
Turn such questions around where you can. Supply the legal text, the report or the figures yourself and have the model summarise them, compare them or rewrite them. That moves the risk to source material you can check.
Questions with personal, medical or legal consequences
Questions with personal, medical or legal consequences should not end with an AI tool. A model can help you organise your thoughts on a situation or draft a letter, but it has no knowledge of your case and takes no responsibility for the outcome.
In that case, use the answer as preparation for the conversation with a doctor, a lawyer or an adviser, not as a replacement for it.
From one question to a fixed way of working
Moving from one question to a fixed way of working is the step that makes the difference for anyone who uses AI every day. One good question gives you one good answer; a way of working gives you the same level every time.
Why the same question can give two different answers
The same question can give two different answers because a language model chooses from several likely continuations at each step. That variation is built in and doesn’t go away.
Providing more context does make the outcome more predictable: the narrower you make the playing field, the less room there is to deviate.
When you should ask a person instead
You should ask a person instead as soon as the answer affects a decision with consequences for someone else, or as soon as you can’t check the outcome yourself. The same applies within organisations that use AI in customer contact: a system that can’t answer a question with certainty should pass it on to an employee rather than take a guess.
If you want to practise this rather than read about it, the free ChatGPT course takes ten minutes and walks you through exactly this way of working, with a checking technique you can apply to your own work straight away. If you want to build a fixed way of working with ChatGPT for your job, the full ChatGPT Course takes it further.
Frequently asked questions about how to ask AI a question
Besides the wording of your prompt, asking AI a question also depends on the channel you use and the form your question takes.
Can you ask ChatGPT a question without an account?
Yes, you can ask ChatGPT a question without an account: go to chatgpt.com and type your question into the input field. According to OpenAI, if you don’t log in, you can have one conversation at a time and your chat history is not saved. If you leave, you can’t continue that conversation. For features such as sharing conversations and setting custom instructions, you need an account.
Can you attach a photo or document to your question?
Yes, in ChatGPT you can attach a photo or document when you ask AI a question, including in the free version. According to OpenAI’s pricing page, the number of messages with uploads is limited in that free version. Remove names and confidential figures from a document before you upload it.
Can you speak your question to AI instead of typing it?
Yes, in ChatGPT you can also speak your question through a voice chat, and hold the conversation out loud. According to OpenAI’s pricing page, the number of voice chats is limited in the free version. For a question with a lot of context, typing often works better, because you can reread and add to the text before you send it.
Can you ask ChatGPT a question via WhatsApp?
Yes, since 13 July 2026 you can once again ask ChatGPT a question via WhatsApp in the European Economic Area, using the number 1-800-CHATGPT. Between 15 January and 13 July 2026 the service on WhatsApp was suspended because of a change in WhatsApp’s policy and terms. Apart from WhatsApp, you can ask your question via the website, the apps for iOS and Android or the desktop app for macOS and Windows.
