How confidence, evidence, approval states, and undo buttons turn a model's output into something a person can safely act on.
Imagine a friend tells you, "Don't go to that restaurant, it's bad." Do you believe them? It depends. Did they actually eat there, or did they just hear it from someone? Can you ask "why, what happened?" If they turn out to be wrong and you go anyway, can you just... go, or do you have to fight them on it first?
That's basically how trust works with AI too. Not because AI is magic, but because trust always works the same way, whether it's a friend, a doctor, or a computer program.
Think about a kid in class who always raises their hand fast and answers confidently — even when they're wrong. After a while, you stop trusting how confident they sound, because confidence stopped meaning anything.
AI has the exact same problem, except it's worse, because AI never sounds nervous. It says "I'm 94% sure" in the same calm voice whether it's actually right or about to make an expensive mistake. So the first job of good AI design is simple: don't let the AI just sound sure. Make it show you something real underneath that number.
You know how a math teacher says "show your work," not just the final answer? It's not because they don't trust you can add. It's because if your answer is wrong, "show your work" is the only way to find out where it went wrong — and if your answer is right, seeing the work is what actually convinces someone else you got there for a real reason, not by luck.
AI should have to show its work too. Not the confidence score by itself, but the actual evidence: here's the document I looked at, here's the line that made me think this, here's what I compared it against. A person reading that can catch a mistake in five seconds. A person reading just "94% confident" can't catch anything at all — they either blindly trust it or blindly ignore it.
If a doctor says "I think you should get this surgery," that's a recommendation. You still get to ask questions, get a second opinion, or say no. Nobody would want a doctor who just performs the surgery the moment they decide it's a good idea.
AI systems need the same kind of pause button, especially for anything that's hard to take back. Let the AI suggest. Let a person say yes or no before anything actually happens, especially when the AI is about to do something big — send money, delete something, tell a customer bad news. The pause isn't slowing things down for no reason. It's the difference between a recommendation and a decision, and those are not the same thing.
Most apps you trust have some kind of undo. You can un-send an email for a few seconds. You can restore a deleted photo from the trash. That safety net is a big part of why you're willing to click things quickly without triple-checking every time.
AI products that let a person undo or reverse an action easily earn trust faster, because people relax a little — they know that if the AI gets it wrong, it's not the end of the world. When there's no undo button, every single AI suggestion suddenly feels much scarier, even if the AI is actually right most of the time.
None of this is about making AI sound smarter. It's the opposite. People don't trust AI because it sounds intelligent. They trust it when they can understand why it's saying what it's saying, push back on it if something feels off, and undo it if it turns out to be wrong. That's not a technical detail buried in the code — it's something you can see and feel directly in the screen in front of you. That's why the interface, not the model underneath it, is where trust actually gets decided.