AI Mistakes to Avoid UK: 8 Fixes for Beginners

After running AI sessions with UK adults for the past year, the same eight mistakes come up again and again. A few take thirty seconds to fix. Left uncorrected, they either produce poor answers or, in a couple of cases, create real safety risks.

This guide covers each mistake once, explains why it matters in plain English, and tells you exactly what to do differently. Start with whichever one sounds most familiar.

1. Sharing personal details without stopping to think

This is the mistake with the most serious consequences, so it comes first.

People often type things into AI chatbotsA program you can talk to in text or voice. that they would not share with a stranger: full names combined with dates of birth, NHS numbers, bank account details, National Insurance numbers, or a family member's address alongside a health question. It feels natural because the conversation is private and the interface looks like a text message. It is not private in the way a phone call to your GP is private.

The risk is not that the chatbot attacks you directly. The risk is that the company running the AI may store your conversations and use them to train future models. If their systems were ever compromised, your data could be exposed. The Information Commissioner's Office publishes specific guidance on AI chatbots. The principle it applies: do not share details you would not want a third party to read.

The fix: Use placeholders for sensitive details. Instead of "My bank account number is 12345678, what do I do about an unexplained charge?", write "My bank account shows an unexplained charge. What steps should I take?" You still get useful advice, without handing over your real details. Our guide to what not to share with ChatGPT has a full list of what is and is not safe to include.

2. Trusting AI health answers without checking your GP

AI tools are genuinely useful for understanding medical terms, preparing questions for a GP appointment, or finding out what a drug interaction might mean in plain English. They are not a substitute for a diagnosis, and they should not be used as one.

The problem is that AI answers sound authoritative. The phrasing is calm and clear, and it mimics the style of medical writing. This can make it easy to accept a health conclusion without checking it against an actual clinician.

AI cannot account for your personal medical history, the medications you are already taking, or the other symptoms you have not mentioned. A good GP does all of that. A chatbot only knows what you type in that particular session.

The fix: Use AI to prepare for the appointment, not to replace it. Ask it to explain what a medical term means, or help you write down your symptoms clearly. Then take that to your surgery. The NHS remains the right place for anything that needs clinical assessment. Our guide to AI and health questions explains where the line sits and which questions are genuinely safe to ask a chatbot.

3. Believing everything the AI tells you

AI chatbots sometimes make things up. They do it confidently, in complete sentences, with no obvious sign that anything is wrong. This is called hallucinationWhen an AI confidently makes something up that is not true., and it happens more often than most people expect.

In documented cases, lawyers submitted AI-generated legal references to courts, only to discover the cases cited did not exist. Dates, statistics, quotes from public figures, links to articles: all of these can be fabricated, and they look convincing because AI is very good at producing text that sounds plausible.

Which? has tested multiple AI tools and found factual errors in every one, even the most widely used. The AI is not lying. It is pattern-matching, and sometimes the pattern leads somewhere that sounds right but is not.

The fix: Check anything that matters before you act on it. If the AI gives you a statistic, search for it on a trusted news site or government page. If it names a law or a service, look it up directly. Use AI as a starting point, then verify.

4. Writing vague prompts and wondering why the answer is poor

AI tools can only work with what you give them. A vague question produces a generic answer. Ask "tell me about pensions" and you will get a broad overview that probably does not address your actual question at all.

The difference between a useful answer and a frustrating one is usually in the promptThe message or instruction you give an AI., and the change is smaller than most people expect.

Before: "Help me write about pensions."

After: "Write a short, friendly explanation of how the State Pension works in the UK, for someone who retired last year and has never had a workplace pension. Keep it under 150 words."

Two details alone change the quality of almost any AI answer: who it is for and how long it should be. We have tested both versions with readers, and the specific prompt consistently produces something you could actually use. The vague one produces something you would delete.

The fix: Include three things: who the answer is for, what you want to do with it, and how long or detailed it should be. Our guide to ChatGPT prompts for beginners has ten copy-and-paste starting points.

5. Stopping after one answer instead of having a conversation

Most people treat an AI chatbot the way they treat a search engine: ask once, read the result, move on. AI works much better as a back-and-forth.

If the first answer is not right, you can say "That is too long, can you shorten it?" or "Can you give me a more formal version?" or "I am 72 and have arthritis, can you adjust those exercise suggestions?" Each follow-up refines the answer. AI chatbots remember the whole conversation within a single session, so you do not have to repeat yourself. You build on what came before.

This is what makes chatbots different from a search engine, and it is the most underused feature for people who are new to them.

The fix: After the first answer, ask one follow-up question. "Can you make that shorter?" is enough to start. Get used to treating it as a conversation rather than a single query.

6. Sending AI-written letters without reading them first

AI is genuinely good at drafting letters: complaint letters, letters to a council, letters requesting a GP referral, thank-you notes, letters disputing an energy bill. This is one of the most useful things it can do for everyday life.

The mistake is sending the letter without reading it carefully first.

AI does not know the specific history between you and the organisation you are writing to. It does not know that you already called twice, that a previous adviser gave you a reference number, or that an earlier letter was ignored. It fills in those gaps with reasonable-sounding guesses, and those guesses are sometimes wrong.

One reader asked an AI to write a letter about a delayed hospital appointment. The draft mentioned a follow-up date that had not actually been agreed. It would have confused the department receiving the letter and potentially delayed the appointment further.

The fix: Always read the draft in full before sending anything. Check that the facts are correct, add the details the AI left out, and adjust the tone if it does not sound like you. Treat it as a first draft from a capable helper, not a finished letter ready to go.

7. Falling for AI-generated scam messages

Scammers now use AI to write phishingA scam that tricks people into giving personal information. emails and text messages that are much harder to spot than the old attempts. The warning signs most people learned, spelling mistakes, clumsy phrasing, generic greetings, are disappearing. AI-generated scam messages can be fluent and convincing in a way the old ones were not.

The National Cyber Security Centre has warned about AI-assisted phishing, highlighting messages that impersonate HMRC and NHS appointment services. Voice cloningAI that can mimic a person's voice using samples., where a scammer creates an AI replica of a family member's voice and calls asking for money or login details, is also increasing across the UK.

If you receive an unexpected message asking you to click a link, call a number, or send money, do not act on it without checking through a separate channel. Call the organisation directly using a number from their official website, not the one in the message.

If you think you have been targeted, Report Fraud (the replacement for Action Fraud) is the right place to report it in England, Wales, and Northern Ireland. Our guide to reporting an AI scam walks through the full process step by step.

The fix: Pause before acting on any unexpected message. A thirty-second check through a separate channel stops most attacks. Read our staying safe with AI guide for a fuller picture of the main threats and how to protect yourself.

8. Trying five tools at once instead of learning one properly

There are hundreds of AI tools available and new ones arrive every week. Some people start by signing up for four or five at once, hoping to find the best one quickly. The result is surface-level familiarity with each and real ability with none.

Each AI tool has its own strengths, quirks, and ways of responding to prompts. You get better results from one tool you understand than from five you have barely used.

The most widely used free option in the UK is ChatGPT. It requires no subscription for the basic version and works in plain English. Starting there makes sense for most people. Once you are comfortable with it, adding another tool is much easier because you already understand the patterns of how chatbots work.

The fix: Pick one tool and use it for a month before trying another. ChatGPT is a sensible starting point. The WellWired Academy teaches you how to get the most from it in a structured course written for UK adults who are new to AI tools.

A quick reference checklist

• Before you type: would you share this information with a stranger in a public place? If not, leave it out or use a placeholder.
• For health: use AI to prepare questions, not to replace your GP.
• For important facts: check anything that matters against a trusted source before acting on it.
• For prompts: say who the answer is for, what you want, and roughly how long it should be.
• After the first answer: ask one follow-up question to improve it.
• For letters: read the draft in full before sending.
• For unexpected messages: pause and check through a separate channel.
• For tools: pick one and learn it properly before adding more.

Frequently asked questions