My AI nagged me for 17 days. It was wrong the whole time.

Hey — and welcome to Sunday Signal.

For two weeks straight, my assistant opened every morning with the same nudge: post something, you've been silent 17 days.

It was very sure about that number. It was also completely wrong.

I hadn't gone quiet. I'd posted all week. The "17 days" wasn't counting posts at all. It was counting the days since an old plan expired, quietly adding one every morning. 15, 16, 17. A confident number, built on nothing, delivered with a straight face.

I believed it for a few days before I checked. That's the part worth sitting with.

1. People trust AI more at the exact moment it makes them worse
A study from three European universities gave people trivia that AI reliably fumbles. With AI help, their willingness to say "I don't know" fell from 44% to 3%. Their accuracy dropped from 27% to 9%. Their confidence climbed from 30% to 76%. Twice as sure, three times as wrong.

Why it matters for you: The danger isn't that the machine is wrong. It's that it hands you its confidence along with the answer, and you stop checking. The same trap my assistant walked me into.

2. The companies moving fastest are hitting the same wall
New research this week looked at more than 10,000 enterprise AI failures. The surprise: hallucinations were under 10% of them. The real damage came from agents confidently doing the wrong thing. In one survey, 57% of companies traced a confidently wrong answer back to missing context, and half had shipped an agent that passed every internal test and still broke in front of a customer.

Why it matters for you: Passing a test is not the same as being right in the real world. My agent "worked" perfectly. It was just fed a bad number. It's the same mistake these companies are making, only with more zeros on the end.

3. The models themselves are learning to doubt
Anthropic shipped Claude Opus 5 on Thursday, near its top-tier intelligence at about half the price. The line that stood out to me wasn't the price. It was that the model is now "much stronger at verifying its work and iterating carefully until it succeeds."

Why it matters for you: The frontier is teaching the machine to check itself before it answers. If the AI is learning to doubt its own first draft, that is a hint about the habit you want too.

AI doesn't lie to you. Lying takes intent. It agrees, fluently, in whatever direction you point it. And agreement in a confident voice is more dangerous than a lie, because you never think to argue with it.

I filmed a video about this very thing this week: your AI is a yes-man. Ask whether your idea is good and it will clap, then hand you a roadmap. Applause is not feedback. Agreement is not insight.

Two receipts this week, one lesson.

One of my trading bots had its first green week, up $22, every exit firing the way it was built to. A month ago I'd have posted that number and left it there. Now I say the honest half out loud: the edge is thin, around a 52% win rate, and one green week is a sample of one. The number is real. It just isn't proof yet.

The other receipt is the assistant. I spent a morning reading two weeks of its emails and followed that "17 days" down to the exact line of code that made it up. Nothing was broken. It was working perfectly, confidently reporting a world that no longer existed. I didn't need a smarter model. I needed to check the number.

Week 10 of building in the open. What it actually looked like:

Turn your AI from a yes-man into a second opinion. Five minutes.

I'm collecting the sharpest "argue against me" answers people get back. Reply and send me yours. The ones that changed someone's mind are the best.

Twelve Saturdays in. The arc so far: ship, constrain, build, compound, patch, subtract, ask, remember, price, staff, equip, and now doubt. Not doubt as in stop trusting the tools. Doubt as in check the number before you act on it, especially when the voice sounds certain.

The bigger lesson from building with these tools: they don't do the thinking for you, they remove the excuse not to start. I've shipped more in the last few months than in all the years I spent waiting to feel ready, and none of it required becoming a different person. Just a more willing one.

Long-form is coming. I've been quietly building the editing setup that makes it possible, and it's close. You'll be the first to see it.

If this helped, forward it to one person who's started leaning on AI a little too hard. That's how this grows.

Quietly, deliberately. One Saturday at a time.

— Kalpesh

P.S. Work with me. I build the kind of quietly reliable systems above for other people too.