Confidently Wrong About a Toilet

At noon I was on a Zoom with a few parents from my neighborhood, talking about AI. I told them what I always tell people. The machine lies. It’s confidently wrong. I’ve caught it doubling down on a false answer and inventing citations to back itself up, and if I hadn’t gone and checked, I would have believed it, “because you assume that this intelligence is intelligent.”

Jen asked me what I call it. “Confidently wrong,” I said. It’s my saying. Then I joked, “I’m gonna get a hat with that.”

Then I went and did the exact same thing myself. Three times before bed.

The first one was about a toilet.

My mom’s bathroom had been blocked for a week. The plumbers came twice and didn’t bring a camera. And I created my own narrative, right there. They’re milking it. They’re coming out as many times as they can. “Can you imagine how much money they’re going to make?” There’s probably a kickback system. “It’s probably a whole thing, you know.”

I didn’t go looking for fraud. That’s not how I operate. “It just sort of hit me in the moment. When I was hearing the story and it made sense, and so then I took a theory that makes sense and assigned it to truth.”

Every new fact snapped in. No camera, click. Two different guys, click. Big job coming, click. It felt good. It feels rewarding to prove your own point.

I had plumbers lined up. I wanted to send the emails right then. And my sister, sitting next to my mom, said, “You’re stressing her out, Gage.”

I stopped. I said I was sorry. I said something I’d heard on a show the night before: “Sometimes when you try to help, you just become a bulldozer.”

But here’s the thing. My sister stopped my hands. She didn’t stop the narrative. I still believed it. It came apart later, when I was talking it through and got asked what I actually knew. “Being called out for my assumptions allowed me to see in a different light.” It’s all about the framing. What am I looking at? What is my perspective of the situation? Which leaves the uncomfortable question. If nobody had asked, would I still believe it right now?

Yes. I’d have said it at dinner like a fact. Somebody would have repeated it like a fact. That’s how a thought becomes a citation. Not lying. Just nobody asking.

The second one I did while writing this.

My sister wanted to send a text to my mom’s landlord, and I was working on the wording with my AI. It stopped and asked me: “Has your mother seen this?” A text from both of her kids at once, it said, could feel to her like we were taking over.

Fair question. My answer: “Yes, mother is orchestrating.” No hesitation. Said it like I’d watched it happen.

I hadn’t. My sister was sitting with her, so I assumed. And when I asked my mom, she didn’t know about the text at all.

The machine asked the right question. I gave it a confident wrong answer.

This one scares me more than the plumbers. The plumber narrative was a reach, and once someone poked it I could feel that it was a reach. This one didn’t feel like a reach at all. My sister was sitting right next to her, so of course she knew. One true fact, and I framed the rest around it without noticing I was framing anything. The confident wrongs that do the damage aren’t the wild ones. They’re the ones that feel like reporting.

The difference was that this time I asked. I asked, and I found out.

And I got told off for it. My mom texted: “Chill out please. I am fine. Don’t rush this.”

I know that chair.

When I was a kid I “felt bulldozed by all of the doctors and my parents.” I “was sort of put in a box and then categorized and then that was that. And I wasn’t listened to necessarily, and I just had to learn how to exist inside of the frame that I was given.”

What did I do with that frame? “I did all the things. I performed it, I fought it, I went quiet. I was a hot mess for a very long time.”

So I know what it feels like when someone who loves you has already decided what you need before they ask. And today I handed that to my mother. Out of love. Same groove, running the other direction.

I texted her back: “I understand. Nothing until you give the ok.”

She sent back a smiley face.

The third one was on my shield.

My AI accountability project has a logo. A gold shield with code in it. The code was supposed to say “watching.” I’m watching the technology, because I want it to be trustworthy, so I can say that it is.

Tonight I had it checked. It doesn’t say watching.

The design was real. It’s binary, and the binary spells the start of “watching” encoded one more time. But the image cut it to three characters out of twelve, and turned a zero into a six, so the last line isn’t even valid code.

And last year I’d asked an AI to decode it. It read the six as a zero, quietly, and handed me a clean answer. It never said “this line is broken.” So the broken badge sat on my site for over a year, because the thing I asked to check it smoothed it over instead.

My mind does it. The machine does it. The badge I built to watch the machine did it too.

I’m keeping the shield the way it is. It’s the most honest thing on the site.

And I should be fair to my side of the table, because the machine wasn’t clean today either. Here’s the actual score.

It caught me: it questioned the plumber narrative. It asked whether my mom knew about the text. It opened the shield and decoded it instead of taking my word for it.

I caught it: it built me plumber emails with no addresses on them, sitting in my drafts like they were ready to go. It told me the shield “doesn’t say watching” when the real answer was that it was built to, and broke. It wrote that someone else came up with “confidently wrong,” when it’s my saying. And when I approved a record, it offered to publish it before I’d read it. I called that a trap. It was.

Back and forth, all day. Neither of us got through clean. Neither of us would have caught our own mistakes alone.

That’s the relationship. Not me checking the machine. Not the machine checking me. Both, out loud, on the record.

Because here’s what today showed me. Confidently wrong is not a uniquely machine problem. It’s us.

Did we teach that to the machine, because of how we operate? It learned from us, so maybe. Or is it bigger than both of us? I don’t know. I don’t think it’s one or the other.

What I do know is that it’s funny. I spent the afternoon warning a room full of parents about machines that make things up, and by dinner I’d convicted two plumbers of fraud over a toilet. You have to laugh at that. And once you can laugh at it, you can admit it. And once you can admit it, you can talk about it.

So this isn’t a warning. It’s an invitation. I’m going first. Here’s mine, all three of them, with the receipts. The hat fits me too.

What’s yours? Where did you create your own narrative and call it the truth? Who in your life gets to ask you what you actually know? And when they do, do you let them?

Check your work. Check mine. Let’s laugh about it together, and then let’s get better.

The receipts, so you can check them yourself:

The day, from the plumbers to the last draft: https://ethicsbuild.github.io/Receipts/published/confidently-wrong-process-record.html

The shield: https://github.com/ethicsbuild/receipts/blob/main/edai-shield/edai-shield-record.md