Make your questions clear

AI chatbots can’t read your mind

The questions or messages you send to an AI chatbot are called ‘prompts’. The way you write your prompts will change the answers you get.

The advice in this section is not specific to healthcare. You can use these tips any time you interact with an AI chatbot.

There have been some scientific studies looking at what makes a good healthcare prompt. At the moment though, there isn't enough research to say for sure what the best approach is. The suggestions below are based on how these tools are built, and what experienced users have found most helpful.

When you write your prompt, try including any specific instructions that you want the chatbot to follow.

Here’s how this might look:

“I want you to help me understand my type 2 diabetes. Give me a longer answer with plenty of detail, but use simple language and bullet points to make it easier to understand. Include links to information from the NHS and diabetes charities. Tell me about the pros and cons of different treatments so I know what to ask when I next see my doctor.”

  1. The prompt gives a clear task for the chatbot – to help you understand your type 2 diabetes.
  2. The prompt says how long the reply should be. Some people prefer shorter answers, and other people prefer longer ones. For some questions, you might prefer the chatbot to go into more detail. For other questions a quicker and simpler answer will be more helpful.
  3. The prompt tells the chatbot to use simple language and bullet points. This means it will avoid using complex medical terms. Using bullet points makes it easier to read text quickly. You can also try telling it to explain any technical medical terms in plain language first.
  4. The prompt gives the chatbot specific sources you want information from – the NHS and diabetes charities. This means the chatbot is more likely to show you content from these trustworthy organisations. This doesn't always work though, so remember to check the facts.
  5. The prompt asks for information about pros and cons of different treatments. Sometimes there will be good evidence that a specific treatment is the best way forwards. Other times there may be several good options to choose from. Occasionally the best choice may be to have no treatment. Chatbots can help you explore the different options. But remember, only a qualified health professional has the training and experience to give medical advice. Your doctor may know more about your overall health than the chatbot, so talk to a person when you need to.
  6. The prompt does not hint at which treatment you would prefer. This is important, because chatbots are designed to be helpful and agreeable. If you ask a question which hints at a particular answer, it may be more likely to agree with you rather than giving a balanced answer.
  7. The prompt makes it clear that you want advice on what to ask when you next see a trained health professional. This can help you prepare questions so that you understand your health problems and the different treatment options.

One thing to consider is that by asking the question in this way, you have told the company that makes the chatbot you have type 2 diabetes. This is personal information about your health. Sharing this information means the chatbot can give you a response which is more tailored to you. You can learn more about sharing personal information in Ask yourself: "What am I sharing?".

If you find that the answer you get back isn’t very helpful, there are some things you could try:

  1. You could carry the conversation on, but tell the chatbot what you want it to do differently in your next message.
  2. You could start a whole new conversation and ask the question differently this time. This might be better in some situations. It means the chatbot starts again without being influenced by your original question. Starting a new chat can help clear up misunderstandings, but any information saved in its memory may still influence the reply. Read more about memory.

Don't be afraid to try different things. Chatbots are tools that are meant to be helpful, but different people have different needs. Find what works best for you. But if you are finding that the conversation is becoming long or unhelpful, it might be best to stop and talk to someone you trust.

Chatbots can only work with the information you give them. “What should I have for dinner” would get a generic answer, but “Give me a recipe for vegetarian lasagne” would help you make a delicious meal.

Sometimes AI chatbots will ask you questions so they can give you more specific responses. You can also try telling the chatbot to ask you questions to improve its replies.

But remember – they are just computer systems. They are designed to keep the conversation going. You don't have to tell them anything you don't want to.

Some people get the chatbot to interview them – to keep asking questions until they have worked out what matters most to them. Some people find this helpful for thinking through health goals.

Some chatbots have different modes and advanced features that you can use. These might help you get more helpful results for particular tasks. Each chatbot will be better at some things and worse at others, so don't be afraid to experiment and see which ones work for you. You can usually find out more about these different modes and advanced features by visiting the chatbot’s website.

Below are some things that you could experiment with:

Memory. Memory lets some chatbots store details from previous conversations. It can then use these details when giving you answers in the future. For example, you might ask the chatbot to remember that you prefer answers in plain language. Read more about memory.

Custom instructions. You can tell a chatbot to always stick to certain rules. For example, you could say that all responses should be written in simple language. All the responses that the chatbot then makes will be written in simple language. You can often set these instructions in the chatbot's settings.

Connectors or plugins. Some chatbots can use extra tools beyond having a conversation with you. Common examples include searching the web for up-to-date information, reading and analysing a file you upload, and writing and running computer code. Some chatbots can also connect to your other accounts, such as email, calendar or cloud storage. This is often called a connector or plugin. Before you turn one of these on, think about what you're giving the chatbot access to.

Projects (or similar). Many chatbots let you create a dedicated space where you can store background information, files or instructions that it can reuse. This can be useful if you keep returning to the same topic, such as planning a wedding or managing a long-term health condition. It means you don't need to explain the context again each time. You may need to remove outdated files or instructions if you use a project for a long period of time.

Skills. Some chatbots let you build reusable sets of instructions for specific tasks. These are called Skills. Once you have added a skill to a chatbot, it can use it automatically whenever needed. This means you don't need to write out detailed instructions every time. Before you use a skill, read its description and check who made it. If you would prefer to make your own, explain the steps, the expected format and anything the chatbot should avoid doing. Try using the skill for simple things first and check the results are what you expect. If you do use a skill, it's important to check facts, dates and wording every time. Skills can make chatbots repeat mistakes as well as useful steps. You may also need to update a skill over time when requirements change or problems appear.

Live conversation modes (ChatGPT Voice, Gemini Live, Copilot Voice, Claude voice mode). As well as typing out messages to an AI chatbot using your computer or your phone, you can also have a live conversation and speak to it out loud. The chatbot will then reply to you using a computer-generated voice. This might be helpful for particular tasks, or when you'd rather not be looking at a screen.

Image, audio and video generation. Some chatbots can generate images, audio, and video. Some people find this helpful when they're trying to understand a complex topic. For example, you could ask a chatbot to produce a summary image if you're trying to understand a scientific research paper.

Attaching files to a chat. As well as replying to text, many chatbots can look at files you send and describe or explain them. You can also ask questions about these files. Before uploading, open the original yourself and check for important details. Medical documents may contain your name, date of birth and NHS number. Sharing these details with the chatbot could put your privacy at risk. Read more about sharing information with chatbots.

The chatbot might miss images or tables, or search only parts of a long document. This means it's always important to check yourself.

In-depth research modes. Several chatbots have research modes. This means the chatbot will spend a long time preparing an answer. It will often search multiple websites and pull together information from lots of different places. This can save you time if you're trying to understand a new area. The summary the chatbot produces may not contain all of the most important information though, so it's important that you check for things that might be missing.

Model size and version. Some AI chatbot services let you choose the specific model size or version, and adjust settings called ‘effort’ or ‘thinking’. In general, larger models can do more advanced things. However, they use more computer power to run. This means that they cost more, so if you have a free account, you might not be able to use the largest models. Even if you have a paid subscription to a chatbot, you may only be able to use the largest models a certain number of times per day. Smaller models require less computer power and are cheaper to run. In many cases, medium or small models will be perfectly fine for simple questions. Some AI chatbots also let you choose the model version. If a model that you use frequently gets updated to a new version, you might find that it does things differently. Newer models are typically more powerful than the older ones. Some people prefer to use older models though, because they are cheaper, or because they are familiar with how they work. You can also set the model to use a higher ‘effort’ or ‘thinking’ level. The higher you set this level, the longer the model will spend processing before it comes up with an answer. More complex problems might need a higher thinking level to get a better response. Using a higher thinking level uses more computer power, and so costs more.

Thinking or reasoning trace. You might also be able to click and see a thinking or reasoning trace, sometimes called chain-of-thought. This is a summary that the chatbot has written to explain how it came to its response. However, research has shown that these explanations may not reflect what actually drove the chatbot's answer.1 It can be interesting to look at the thinking or reasoning trace, but you shouldn't see this as an accurate reflection of how the chatbot formed its response.

AI agents taking action on your behalf. Some chatbots can do more than just reply. Modern chatbots can complete complex multi-step tasks, such as researching a complex topic and writing you a summary report as a Microsoft Word document that you can download and read on your computer. They might even be able to carry out tasks on your computer for you, such as visiting websites, clicking buttons and submitting a form. When AI systems can take multiple steps towards a goal, this is often called agentic AI. Because the chatbot is acting with more independence, it is critical to check what it is doing carefully. It's also important to clearly state the result you want and set clear boundaries. You may be asked to give permission for the AI agent to take actions on your behalf. Think carefully, and always review permissions before it performs important tasks such as sending information or deleting files.

One risk worth considering is that when a chatbot reads a web page or document, it cannot always tell the difference between your instructions and any text on the page that tells it to do something. This means an AI agent could be tricked into sharing your private information or taking actions you do not want. This is why it's so important to review any permissions you have granted regularly.