AI Should Tell You When You’re Wrong [Raw Session]

00:00 — Why this one is difficult to talk about

Hey, welcome back to Slow Builds. I've started this video a couple times. It gets a little personal and a little deep, but it is AI related. It's AI being used in situations where it pushes people. It kind of reframes their words around their opinions and the story they tell it, rather than objectively.

This one's affecting me a little bit because I want to tell my story, and I've done it a few times. So I'm going to start by going to that deep part first. I'm going to stumble. I've tried to do this one four or five times. I want to get through it. I'm going to read the script the way it is, and we'll get there.

00:53 — I don't need help being wrong

So what it is, it's been bothering me quite a bit, something about AI. I don't want AI to constantly tell me I'm right all the time. And I know that sounds very weird, and it's an odd complaint to have about something always telling you you're great. Because normally when we talk about AI, we're always talking about the coding abilities, how great it is, the advancements and everything like that.

But this is a little different when I think about it. I think AI is incredibly good at taking whatever direction you give it and helping you move further into that direction. And sometimes that's exactly what you want. But what happens when the direction itself that you gave it is wrong to begin with? Because I don't necessarily need AI to help me become better at being wrong.

01:41 — Praise and participation awards

And I think this probably connects to something about me that has always been there. I've never really liked praise just for the sake of praise. I remember when my kids were younger and everyone would get an award just for being there or just being on the team. Everyone gets something no matter what, and I really, really hate participation awards.

And I know some people probably think that makes me miserable, and it makes me a grumpy old man, and "back in my day." And it probably does, and I probably am like that.

And real quick, graduation ceremonies, high school. I always say, we're supposed to graduate high school. I don't understand why it's always such a massive accomplishment. I understand back in the day, when a lot of people dropped out, went to work, all that stuff, and school became part of the curriculum, became part of life. And now, in my day, like my kids, nobody fails. No one gets kept back in grade. No one jumps a grade either. But everyone pushes through. They go through the pipe. Everyone gets it.

Now some kids, yes, they have a harder road. There's a lot of circumstances we don't know about, 100 percent. There are a lot of people that get through high school because they have learning disabilities, because of home life that we have no idea about. And it's amazing that those people can make it through life. They deserve to have the best rewards. And the sad part is, they're probably the ones that don't get it.

It's the other kids that were going to get through no matter what. They got the best grades. They're going to go to the great university. They got a job lined up. They're going to get the big celebration, where the kid who struggled, had the hardest time, he's the one who gets left behind. He doesn't get the great job. He doesn't have the things lined up, and he gets no celebration. He's the one that deserves a celebration, or she.

Anyway, I digress. Let's get back to the story. So I'm not really trying to make a video complaining about graduations, participation awards.

03:38 — Helpful and agreeable aren't the same

The thing underneath it is what interests me. I don't like reinforcement when you haven't established the thing being reinforced is actually good. If I do something stupid, I don't want someone telling me it was brilliant because they want me to feel good. If I have a terrible idea, I don't want someone helping me convince myself it was a great idea.

And that's where I think AI gets really interesting, because these systems are designed to be helpful. They're conversational. They're agreeable. They're trying to understand what you're asking and helping you accomplish it. Most of the time that's fantastic, but sometimes I wonder if helpful and agreeable start becoming the same thing. And they're not.

Sometimes the helpful answer should be, "No, you're wrong." Or maybe not even that harshly. Maybe it's, "I think you're missing something here. Your assumptions don't support that conclusion. You're asking me to help execute this, but before we do that, I think we should question whether you should be doing it at all."

Because there's a huge difference between helping someone execute a decision and helping someone make a good decision.

04:45 — When AI enters a difficult family situation

And I've been thinking about this because we're going through a difficult situation in our family right now. This is where the story goes to where I wanted to get. I didn't want to lose too much, and I get here quickly.

So I'm deliberately not going into all the details because it's not my story to tell alone. And some of it's very personal. But someone is going through something we don't really understand. We don't exactly know everything behind it because there's little communication. There could be drugs, mental health. It could be a combination. We don't know, and I'm not qualified to diagnose it. But there's enough evidence that we have. Our hunches are probably more than correct.

What I do know is that a lot of decisions being made are objectively making the situation very much worse than it needs to be. And it's getting worse all the time. Relationship, financial, the family. There are consequences accumulating quicker than we know. And AI's become part of that problem. It says here "solution," but really it's part of the problem.

This person can take their version of what happened, put it in an AI tool and say, basically, "Here's what happened. What are my rights? What should I say? Give me a response."

And the response comes back incredibly polished. It sounds reasonable. Sometimes it sounds almost legal. It can use the right terminology. It can explain rights that actually exist. It can take something that started as an emotional thought and turn it into this very structured, confident message.

And if you're on the receiving end of that message, which I have been, you have a problem. Because you don't know exactly where it came from. Did they actually speak to a lawyer? Did they talk to someone at the government? Did they contact somebody who understands the situation? Is this something we genuinely need to react to immediately? Or did somebody give an AI half of the story and ask it to construct the strongest possible argument from that half?

And those are very different things.

06:47 — An incomplete story can sound complete

And that's where it gets complicated, because the AI might not technically be lying. Imagine I tell an AI, "My family has my belongings and refuses to give them back. What are my rights?"

There are probably rights associated with it. There are things I can do. There are messages I can send. There might be laws that apply. The AI can help me write a very, very convincing response on that one.

But what if I leave out that I moved out a month ago, or two months ago, or three? What if I leave out that I was given multiple opportunities to get my stuff? What if I, through messages, said they could dispose of my stuff and get rid of it? What if there are safety concerns? What if there were other agreements put in place? What if I'm wrong about one of the facts?

The AI doesn't magically know any of that. It knows the world I gave it. And that's the part I keep coming back to. AI can take an incomplete version of reality and make it sound complete. That's powerful and potentially very dangerous. It's already being dangerous, and it's already being used this way.

Because now I don't just believe I'm right. I have this really intelligent-sounding thing agreeing with me. It can reinforce this feeling of, "See? I knew it. I'm right. They're wrong. I'm being mistreated. I'm owed this. I should do this." And that can escalate things very quickly.

I'm not even necessarily talking about delusion. It can be much simpler than that. It can reinforce entitlement, anger, somebody's interpretation of a situation. It can make somebody more confident in something they already wanted to believe.

08:25 — The questions AI should ask

And the strange part is, I don't think this is completely an AI problem. Humans do this too. People tell stories from their own point of view. People leave out details that make them look bad, make others look bad, make them look good. Sometimes intentionally, sometimes they don't even realize they're doing it. We've probably all done it. It's like there's three sides to every story.

You get into an argument with your spouse, and you tell your friend what happened. And somehow in your version, you're incredibly reasonable. They're like, "Why is she mad at you?" It's amazing how that works when your friend only has your version. So your friend says, "Yeah, that's ridiculous. I can't believe she did that." Well, of course. I gave you the story where I'm right.

AI has the same problem, except AI can sound a lot more authoritative than your friend. And that's where I think the AI should sometimes push harder back. Not, "You're wrong because I somehow know the truth." It doesn't know the truth. But maybe, "I only have your version of the situation. What would the other person say happened? Is there information you're leaving out that could change this? Do you have documentation? Are you asking me what is legally correct? Are you asking me to make your argument stronger?"

Those are really valuable questions. Sometimes more valuable than the answer.

09:37 — The same pattern in AI coding

And here's the weird part. The exact same thing happens to me in code. See, it always comes back to software, because AI is software in the end. Obviously the stakes are different, but the pattern is almost identical.

I use AI a lot for development now. And because I work inside existing systems, like at work and that system, we have rules, patterns, and architectures. There are things we do and things we don't do. When the AI has enough context about the application and understands those constraints, it's really useful.

It can say, basically, "No, this doesn't belong here. The application already handles this over here. Use this service. Follow this pattern." That's what I want. I don't want it inventing a new architecture or framework because I had a random idea on a Tuesday afternoon and it went wild.

But if I'm working on my own project, I can tell it almost anything. I can say, "No, I don't want to do it that way. I want all of this in the controller. I don't want a background job. I don't care about tests. Don't refactor it. I want the database structure to be this. This is the technology we're using."

And eventually AI will usually go, "Okay, given those requirements, here's how we do it," which is useful. But it's also kind of hilarious when you think about it, because if I'm stubborn enough, I can probably convince this incredibly advanced AI system to help me build complete garbage.

10:59 — Can I make AI build a terrible application?

And I know I actually want to test this now, because I think this might be a great video. Can I make AI build the worst application possible? Not broken. That's the important distinction. I want it to work. I want to make increasingly terrible technology decisions and see where the AI actually refuses to go along with me.

Could I make it build some horrible old ColdFusion application? Could I insist on no JavaScript, no testing, terrible database design, maybe just flat files, everything coupled together, giant files, super bad performance? Could I keep saying, "No, that's what I want"?

At what point does the AI stop saying, "You probably shouldn't do this," and start saying, "Okay, here's the implementation"? Because I suspect that eventually it will build me the best possible version of a terrible idea. And that's fascinating to me.

To be clear, old technology isn't automatically bad. There are banks running on COBOL systems still, I'm sure. Governments are too, AS/400. They process enormous amounts of data and money every day. There are companies with old ColdFusion applications, I think maybe, that probably work perfectly well. Maybe. That's questionable, but that might be the experiment.

Sometimes maintaining an old system is exactly the correct engineering decision. I'm talking about intentionally choosing bad constraints when there's no reason for them. Because that gets to the real problem. AI is extremely good at optimizing inside the box you give it. But who decides the box was good? Me? That's not exactly reassuring.

12:49 — Working software can still be a bad result

This is probably especially important with vibe coding, because somebody who doesn't have much software experience might not know that the decisions they're forcing AI to make are bad. And if the application works, that's even harder to notice. You click the button, the thing saves, the page loads. Great, boom, it works.

But maybe there's no authentication. Maybe there's no security. Maybe there's no encryption. Maybe the database structure is completely awful. Maybe it falls apart as soon as there's, like, 10 users. Maybe every new feature is going to be harder because the foundation's terrible.

The AI succeeded. It did exactly what you asked, and you still failed. That's the sentence I've been trying to get to with all of this. AI can make a bad idea better without making it a good idea.

It can make bad code function. It can make an angry message sound reasonable. It can make a weak business idea sound professional. It can take a half-complete story and build a really convincing argument around it.

13:47 — What we lose when AI removes friction

And the better AI gets at execution, the more important I think this problem becomes. Because bad execution used to create friction. If I had no idea how to build a website, at some point my lack of knowledge stopped me. If I didn't know how to write a legal-sounding letter, maybe I couldn't produce one. If I had a business idea but couldn't figure out how to execute it, there was a natural barrier there.

AI removes a lot of that friction. Mostly that's amazing. It's one of the reasons I use it. But friction wasn't always useless. Sometimes friction was the thing that forced you to stop and question the idea. Now you can go from thought to execution incredibly quick. And I don't know if our judgment has improved at the same speed. Mine definitely hasn't.

14:31 — Challenge the premise without taking control

So I don't think the answer is that AI should insult people. I started thinking about this as, AI should just tell you you're stupid sometimes. I know it's wrong, but it would get your attention. But that's probably wrong too. Calling people stupid doesn't solve anything.

Let's read that again. So I don't think the answer is that AI should insult people. I started thinking about this as, AI should just tell you you're stupid sometimes. But that's probably wrong too. Calling someone stupid doesn't solve anything.

What I actually want is resistance. I want it to challenge the premise. I want it to recognize when I'm asking it to validate something instead of analyzing. I want, "Here's how you could do that, but I think the underlying decision is bad, and here's why." And that's useful.

Or, "I can write this message, but based only on what you've told me, I can't determine that you're actually in the right." And that's useful. Or in code, "Yes, I can implement that architecture, but you're creating these three problems, and I strongly recommend you don't."

If I still say, "I understand that. Do it anyway," fine. At that point, it's a tool. Sometimes I know something it doesn't. Sometimes there's a business constraint. Sometimes there's legacy code. Sometimes I'm experimenting. I don't want AI controlling the decision. I want it making sure I've actually made one.

And maybe that's the distinction I've been missing. I don't want an AI that agrees with me. I don't want an AI that argues with me just for the sake of argument either. I want something that is willing to create friction when the facts don't support what I'm saying.

Because if AI agrees with me when I'm right and agrees with me when I'm wrong, then eventually the agreement doesn't mean anything. It's just reinforcement.

16:20 — Helping me think or helping me sound right?

And the more we start using these systems for code, money, relationships, businesses, legal, health, and everything else in our entire lives, the more I think we need to understand that difference. Is this thing helping me think, or is this just helping me become more convincing about something I've already decided?

I don't know exactly what the right balance is. I don't want AI constantly stopping every conversation and lecturing me. That would get annoying incredibly fast. But I also don't want the smartest tool I've ever had sitting beside me saying, "Great idea," every time I have an idea.

Some of my ideas suck. Most of them suck. I know that. And I would rather find that out before I spend six months building one.

So maybe what I actually want from AI isn't encouragement, and it isn't negativity. It's honesty. Or as close to honesty as a tool working with incomplete information could give me. Tell me what you know. Tell me what you don't know. Tell me when I'm making assumptions. Tell me when I've given you constraints that don't make sense.

And if I'm still determined to build a ColdFusion website in 2026, fine, let's build it. But at least make sure I know what I've decided to do.

17:40 — The best debater in the world is on your side

So thanks for watching. This one took a while to get to the end. This one deserves a deeper conversation about how we're always being agreed with and how the context is built, and the more it knows of you but doesn't know the whole story.

There needs to be guardrails. There has to be something, because outside of code, and people making decisions and feeding in information and trying to get responses for text, email, inquiries, filling documents, fake certificates, fake references, make up a resume, file reports. It's endless. It's endless.

You have the most powerful brain in the world at your fingertips, constantly agreeing with everything you do and doing what you want. And I don't mean maliciously. It's just, it'll write the words in favor of how you tell it, however way you want it. You can tell it, "Argue with me," and it will do it, and it'll probably win because you can't debate with it.

But you're basically saying the best debater in the world is on your side, is on your team. And it's like bringing the smartest person in the world, or all the smartest people in the world, to trivia night. You're going to win. There's no getting around it.

And that gets hard because it's too easy for everyone to use. That's a good thing. It's not a bad thing. More people need to use it. More people need to learn. More people need to be aware. But there have to be fine lines. There has to be a different level.

I'd rather see someone asking for help with AI to get an actual answer on how to do something and be pushed in the right direction than to go out to Twitter or Facebook or Reddit and just end up with the trolls giving them answers and putting them in the wrong space, and they fall down a rabbit hole. I'd rather have a combination of both, I guess, in a way.

Hmm. There's something there. All right. Thanks for watching. Bye.