Let me tell you about the time I almost ruined my wife's baby shower.
It was June 14, 2025. My wife was about seven months along, and I was trying to do something thoughtful — design the baby shower invites myself using ChatGPT. I had a vision. I fed it all the details: the date, the venue, the vibe. The first output was pretty good. I checked it carefully, made some notes, and started iterating. A little tweak to the colours here. A small wording change there. Each time I reviewed, I focused on the specific thing I'd just asked it to change.
You can probably see where this is going.
By the fifth or sixth version, I had stopped reading the whole thing. I trusted it. I knew what I was tweaking, and I assumed everything else was sitting exactly where I'd left it.
It wasn't.
The date had quietly changed to something else entirely. Wrong day. Wrong everything.
My saving grace? I always share a final draft with my wife before anything goes out. She caught it immediately. We laughed. We fixed it. Nobody showed up on the wrong day.
But here's the thought that's stuck with me ever since: what if I hadn't shown her?
The Trap Has a Name: Iteration Fatigue
This wasn't carelessness. I'm a software engineer. I know better than most that you don't ship without checking.
But here's what happens when you iterate on anything — not just AI outputs — your brain starts taking shortcuts. When you change one small thing, your mind quietly fills in the rest as safe. I only changed the font. The date is fine. The venue is fine. I would have noticed if something else changed.
AI doesn't share that assumption.
Every time you send a new prompt, the model is making a fresh attempt at the whole thing. It has no memory of what you approved in the last version. It has no loyalty to the details you already signed off on. It is not tracking what changed. That's your job — and it's easy to forget that in the flow of a back-and-forth session.
Your Eager, Clueless Intern
I've always described AI as a fast, eager, occasionally clueless digital intern. And this is exactly where that analogy earns its keep.
If you handed a draft to an intern and said "just change the heading font," would you forward whatever came back directly to a client without reading it? Of course not. Because the intern might have accidentally deleted a paragraph. Might have changed a number. Might have introduced a typo while reformatting.
You'd glance over the whole thing. Takes thirty seconds. Saves enormous embarrassment.
AI is that intern — except faster, more confident, and completely unbothered by mistakes. It will hand you a document with a wrong date, a made-up statistic, or a quietly altered detail, and it will do so with the same calm, polished tone it uses when it gets everything exactly right.
It has no skin in the game. No consequences. No embarrassment. It won't flag its own mistakes.
Your name is the only one that goes on the final product. That's not a burden. That's just the deal.
The One Habit That Saves You
I'm not going to give you a ten-step checklist. That defeats the whole point.
Just one rule: do one final read-through as if you're seeing it for the first time.
Not a skim. Not a "I'll just check the parts I changed." The whole thing. Top to bottom. Like a stranger is reading it for the first time — because sometimes, someone is.
And for anything going out to more than a handful of people — a team email, a client document, an announcement — add a second pair of eyes before it leaves your hands. Not because you're incompetent. Because a fresh perspective catches what familiarity misses. My wife was my second pair of eyes that day, and she wasn't even trying to proofread. She just read it fresh.
That's the human in the loop. That's you.
A Prompt Worth Keeping
Before you do your own final check, try asking AI to audit itself first. It won't catch everything — but it sometimes catches more than you'd expect.
Copy and paste this at the end of any AI-generated draft before you review it:
"Before I review this, please re-read it carefully and flag anything that seems inconsistent, factually uncertain, or that may have changed from my earlier instructions. List any concerns, even small ones."
It won't be perfect. But it gives you a head start — and it puts you in the right mindset to read critically rather than just confirm what you're hoping to see.
A Few Honest Warnings
Since we're being real with each other:
AI will not tell you when it quietly changes something. There's no notification. No asterisk. No "hey, I also tweaked the date while I was in there." You will only know if you look.
Confidence is not accuracy. The tone of an AI response tells you nothing about whether the content is correct. It sounds equally assured whether it's right or wrong.
Longer sessions carry more risk. The more you iterate in a single conversation, the easier it is to stop reading the full output. Set a personal rule: any final version gets a full read, no matter how many rounds it took to get there.
You Were Always the Last Line
Here's the thing I want you to take away from all of this — and I mean it as genuinely good news.
You are not the last line of defence because AI is broken. AI is genuinely useful. I use it every single day. It saves me real time and real mental energy.
You are the last line of defence because you are the only one in the room who actually cares about the outcome. You're the one who knows your wife will be reading that invite. You're the one who knows your client's data team will fact-check that stat. You're the one who understands the consequences of getting it wrong.
AI doesn't have any of that context. It doesn't feel the weight of it. It just has your prompt and its best guess.
So use it. Lean on it. Let it do the heavy lifting.
Then read the whole thing before it goes out.
Have you ever caught an AI mistake before it caused real damage? Or has one slipped through on you? Hit reply — I'd love to hear your story.
