Can AI be wrong?

Yes. Regularly. And it’s often wrong in the same calm, confident voice it uses when it’s right.

That’s the single most useful thing to understand about AI, and it isn’t a reason to avoid it. It’s a reason to use it in a particular way.

---

Why it happens

Generative AI writes by predicting what a good answer looks like, based on patterns it learned. Most of the time, the pattern of a good answer and the true answer line up. Sometimes they don’t, and it produces something that fits the pattern and isn’t true.

It can:

• State a wrong fact, date, or number
• Invent a quote, a statistic, or a source that doesn’t exist
• Describe a rule or policy that’s out of date, or that never existed
• Mix up two similar things, like two medications or two companies
• Leave out the one detail that changes the answer

You may hear this called “hallucinating.” The word matters less than the fact.

There’s no reliable signal in the writing when it happens. Sometimes it hedges when it’s right and sounds certain when it’s wrong.

---

When checking matters most

Match your checking to what happens if it’s wrong.

Check carefully, against a source you trust or a person who’s accountable:

• Medical questions, and anything about medications
• Legal matters, contracts, and your rights
• Money: taxes, loans, investments, benefits
• Current prices and fees
• Dates and deadlines
• Policies at your workplace, school, bank, or insurer
• News and recent events
• Quotes, from anyone
• Statistics and figures

Checking matters less when:

• You’re brainstorming ideas you’ll judge yourself
• It’s rewording something you wrote, and you can read the result
• You’re drafting a message you’ll review before sending
• You know the subject well enough to notice a mistake

For health, legal, and money questions especially, AI can help you understand what you’ve been told, organize your questions, and prepare for a conversation with a professional. It isn’t a replacement for that conversation.

---

Use it to think, not as proof

The practical rule: use AI to help you think and work, not as automatic proof that something is true.

What that looks like:

• Ask it to explain a letter, then check the deadline against the letter itself
• Ask it for questions to bring to your doctor, then ask your doctor
• Ask it to compare two quotes, then confirm each number against the quotes
• Ask it for a draft, then read every fact in it before you send

In each case AI does the part it’s good at (organizing, explaining, drafting), and you or a source handle the part where being wrong costs something.

---

A three-step check

When an answer matters:

1. Find every specific in it. Names, numbers, dates, quotes, rules.
2. Sort them. Anything that came from material you gave it, you can check against that material. Anything else is unverified, even if it came with a confident explanation.
3. Check the ones the decision depends on against something outside the chat: the original document, an official website, or a person.

If you can’t check a specific and your decision depends on it, don’t rely on it.

---

Something you can try

Ask about something you know well: your town, your job, a hobby, a book you love.

Tell me five specific facts about [SOMETHING YOU KNOW WELL],
including at least one number or date.

Check each one against what you know. If one is off, notice that it read exactly like the others. If they’re all right, try a less well-known detail of the same subject, since mistakes tend to show up where less has been written.

---

Something important to know

Asking “are you sure?” isn’t a check. It may apologize and change a correct answer, or stand by a wrong one. A check means comparing against something outside the conversation.

---

If it doesn’t match what you see

It won’t answer a medical, legal, or financial question. Some apps limit these answers. You can usually still ask it to explain terms, summarize a document you provide, or help you prepare questions for a professional.

It gave you a source, and the source doesn’t exist or doesn’t say that. That’s one of the ways it goes wrong, and the reason to open links rather than trusting that a link exists.

Two AI apps gave you different answers. That tells you at least one is wrong, not which one. Go to a source outside both.