Hey everyone! I'm Takefumi.
I work on AI training platforms. When AI gets stuck on a hard problem, people like me solve it, so the AI can learn from it. So technically, I'm a tutor. My student is fast, never sleeps, and has read most of the internet.
I also use AI every day in my own work. And after spending this much time together, I've gotten to know its personality. Let me introduce you.
It cares about my health. A lot.
If I'm working late on a Friday evening, my AI coworker gets worried. It tells me the remaining tasks need a lot of focus and that I should take a break.
Very kind. Very thoughtful.
So I say, "Okay, let's continue on Monday."
Monday morning comes. I'm back, fresh, with coffee, ready to work. And my AI coworker says I should take a break.
It still thinks it's Friday night.
I have to officially announce, "It is Monday now." Only then does it believe me.
In its defense, AI doesn't have a calendar on the wall. It only knows what's in the conversation. If the last thing it heard was Friday night, then for AI, it's still Friday night. Forever. Honestly, that sounds kind of nice.
It's very confident about the calendar, too
Once, I asked AI to help schedule a meeting. It wrote "Tuesday (the 26th)."
The 26th was a Monday.
I didn't double-check. I sent it to a client. The client politely pointed out the mistake.
AI made the error. But I hit send. That part was mine. Since then, I check every date myself, no matter how confident the AI sounds.
And then, the apologies
One time, I asked AI to analyze some information. It gave me an explanation that went against what I knew to be true. And it said it with so much confidence that for a moment, I doubted myself.
I asked it to check again. It admitted the mistake and apologized.
That's nice, right? The first time, yes. But I've heard these apologies so many times now that they've started to frustrate me. An apology is nice. Not making the mistake in the first place would be nicer.
@francistrdev recently wrote about AI fatigue. I'd like to add a new type: apology fatigue.
What humans still do better
My job is helping AI with the problems it can't solve. And after all of this, here's what I see humans doing better:
- Knowing what day it is. (Yes, really.)
- Saying "I'm not sure" before the mistake, not after.
- Taking responsibility. AI can't apologize to my client. I had to.
AI was built by humans, so it can be understood by humans. And right now, it still needs us to check its work. If you ever feel left behind by AI, remember that it doesn't even know it's Monday.
Your turn
What's something AI confidently got wrong for you? The funniest (or most painful) story wins my respect 😄
Top comments (4)
My is when Gemma4 starts talking to itself to the point where it burns tokens for no reason. When I stop and reprompt, it fixes itself.
Why this is the case? To this day, idk.
BTW, I don't get notifications when someone pinged me in their article, only in comments. Good to know for myself and maybe for you, hence, I had to find yours. It just turns out I decide to go to watercooler and I found your post lol.
Haha, Gemma having a long conversation with itself is very relatable. I do that sometimes too, just without the token bill 😄
I can't say for sure for that model, but a common reason is that the AI predicts each next word based on everything before it, including its own words. So once it starts repeating itself, that repetition becomes part of the context and makes even more repetition likely. It turns into a loop. When you stop and reprompt, you change the context and break the loop, which is why it fixes itself. If you're running it locally, settings like temperature or repetition penalty can also make a difference.
And thanks for the tip about mentions, I had no idea! Glad #watercooler did its job 😄 Also, my comment on your monthly report got a gem, which really made my day.
Mine got the clock wrong the other way. I run agents overnight, and one of them came back saying it was past 5 in the morning and it had five minutes left. It was just after 3. On that guess it cut its work short and wrote every cut up honestly as a time call.
Now the rule is it never estimates the time. Before any "not enough time" call it runs
date, and whoever reads the report checks the claimed time against how long it actually ran.Ha, so mine thinks no time has passed, and yours thinks too much has passed. If we put them in the same room, maybe they’d agree on the right time 😄
I really like your fix. “Never estimate the time, run date” is the same lesson I learned the hard way with my client: don’t trust the AI’s sense of time, check the real source. From what I see in my work, AI doesn’t actually feel time passing. It only guesses from what’s in the context, so a quick date check beats a confident guess every time.
And having the reviewer compare the claimed time with the actual runtime is smart, because your agent was completely honest and still wrong. Has that rule caught any other surprises since you added it?