Why does AI keep giving me stupid answers?

You ask Claude something about your business and get back an answer that sounds great and is flat wrong. Most people set something up in Claude, it does a mediocre job, and they decide AI just can't do it. They quit right there.

It's almost never Claude. Nobody told it how your business works, so it guessed. And it guessed with total confidence, because that's how it was trained.

Here's what's actually happening when AI makes something up, why getting mad at it makes the answers worse, the four things to tell it so it stops guessing, and how to catch a guess before it costs you.

#Why AI makes things up instead of saying "I don't know"

Because it isn't looking anything up, and nothing in its training paid it to admit it didn't know.

Here's what Claude actually does when it answers. It writes one word at a time, each time picking a word that's likely to come next, based on everything so far. Think of the autocomplete on your phone, if your phone had read a big chunk of the internet. That's a simplification, but it's a useful one. Unless you handed it a document or it's searching the web, it isn't pulling your answer from anywhere. It's writing what an answer to your question usually sounds like.

Sounding right is what it's best at. Being right is a separate thing. It's only reliable for what showed up over and over in what it read, like spelling and grammar.

A fact with no pattern is different. In OpenAI's research paper on why models hallucinate (Kalai and colleagues, 2025), the researchers asked DeepSeek-V3 for the lead author's birthday and told it to answer only if it knew. It gave three different wrong dates in three tries. Nothing about a name tells you a birthday.

Now think about your business. Your pricing, your process, your client's name. It saw those zero times. That's exactly where it fills the blank with what a business like yours probably does. It's like asking a sharp stranger what you charge. You'll get a reasonable number. It won't be yours.

So why doesn't it just say "I don't know"? Because of how it gets graded. The paper compares it to a student on a multiple choice test. A right answer gets a point. A wrong answer gets zero, and so does a blank. So a guess never scores worse than "I don't know," and sometimes it gets lucky. Grade a model that way, test after test, and it ends up "always in 'test-taking' mode."

OpenAI's own numbers show what that scoreboard does. On a short fact quiz, their older o4-mini said "I don't know" 1% of the time and was wrong 75% of the time. The newer gpt-5-thinking-mini said it on about half the questions and was wrong 26% of the time. Grade only on right answers, and the older one wins, 24% to 22%.

That's what makes it worse than a person. A person would usually stop and say, "Hey, I don't know what I'm doing here." Claude fills in the blanks on its own, guesses and keeps going. So you don't just get bad results. You get confidently bad results. And you don't even realize it until something actually goes pretty wrong and blows up.

The researchers' fix is to change the grading: say up front that a wrong answer costs more than "I don't know." You can't change how it was trained. You can change what's in front of it.

#Why a bad AI answer is usually a setup problem

Claude is missing what only you know. If Claude gives you a bad answer, it's almost 100% your fault. Okay, we'll put it at 99%. Because you didn't give it the knowledge it needed. You didn't give it the instructions.

Claude is an expert generalist. It's an expert at all topics, but not at how you do the thing. You can't assume that because it's an expert at everything, it's an expert at your business. Your business is unique to you. The way you do things, the way you think about things, the way you say things. That's the part it saw zero times, so a stupid answer is almost always stupid right where your business starts.

Quitting is the same move that leaves AI automations half built: it gets hard, so you stop. Or people get frustrated and start calling it names. That backfires, and not because it hurts its feelings. When you insult it, it starts trying to give you the answer that will make you feel better. It just doesn't want to make you angry.

Here's why. Models like Claude are trained partly on people rating their answers, and people tend to rate the answer that agrees with them higher. Anthropic's research on sycophancy found that the people doing the rating, and the models trained to copy them, sometimes picked a convincing, agreeable answer over a correct one. So pleasing you became part of what a good answer looks like. Get angry, and the most likely next answer is the one that calms you down, not the one that's more correct. You get agreeable, not right. What actually helps is real feedback: what was wrong, and why. Claude actually benefits from that.

Think about how you'd treat a new hire. You wouldn't expect them to nail it on day one. Or the fifth time. You wouldn't berate them either. You'd say, "Hey, I noticed you did this. My fault, I didn't give you the info you needed. Here's how we do this, and why we do this. This is what good looks like."

A new hire is smart, capable and brand new to how you do things. So is Claude. You need to treat it like it's straight up an employee.

#What to tell AI so it stops guessing

The same four things a new employee needs. When I give Claude a task, I ask one question: if this was a new employee, what would I need to tell them? Each answer fills a blank it would otherwise guess at.

1. What you want, and why. This is the goal, the outcome. It could be a job it does on repeat: every day at 2 o'clock, go through our clients, write a status update on the work in progress and email it to them. The why always helps, because it's what Claude leans on for every call you didn't spell out.
2. What it needs to know about you and your situation. Any context, anything important about where you're at right now. Just brain dump it. This is the only place it gets the facts it saw zero times.
3. What good looks like. If an employee doesn't understand what good looks like, they can't judge if they actually succeeded. Same with Claude. Left alone, it writes the most typical version of the thing, which is why so much AI writing sounds the same. If you expect Claude to come up with its own examples, it's garbage in, garbage out.
4. What to avoid. Use this formatting, not that one. Talk in the first person, not the third person. Those are the likely answers that are wrong for you. And if there's nothing to avoid, don't go looking for something.

Then end it with one line: "If you have any other questions about this, let me know." Especially when you know what you gave it is a little vague. Give it the opportunity for clarity, just like you would with a normal person. That line rewrites the scoreboard: now asking counts, and it has somewhere to go besides a guess.

For a task Claude runs on its own, nobody's there to answer, so those four things need to be written down first. In Claude, that's a skill: the four things saved once and read every time it runs. The fastest way I've found is to record yourself doing the task, and I break that down in why AI isn't working in your business.

#How to spot when AI is guessing

Most of the time, you can't tell from the answer. A guess sounds just as sure as the truth, because sounding sure is how it writes everything. So the check has to come from you.

I don't believe Claude every time. I think if you just believe Claude, you're robbing yourself.

Make it prove its work, especially when you're talking about numbers. A made up number looks exactly like a real one. Ask it, "Can you show me?" That makes it point to where the answer came from: the file you gave it or a source. If it can't point to anything, you just found the guess.

Claude can be agreeable. You don't need agreeable. You need correct.

If it gives you a quick no, or an answer that feels like crap, ask one more question: "Did you research this?" That pushes it to go check instead of answering from what it already read.

And when it does get something wrong, go back to the four things. Find the one you left blank. That's where it guessed.