When Not to Use AI: The Tasks Where It Actually Makes Things Worse

Last Tuesday I spent three hours using ChatGPT to write a condolence card to my friend whose father died. Three hours. The final version I sent? I wrote it myself in four minutes, after throwing away everything the AI gave me.

That's when I realized I'd been using AI wrong for months. Not because the tools are bad — I've tested dozens and many are genuinely useful. But I kept defaulting to them for tasks where they actively made my work worse. Slower. More generic. Sometimes embarrassingly off.

Here's what two years of obsessive testing taught me about when to close the ChatGPT tab entirely.

Tasks That Require Actually Knowing Someone

The condolence card was the obvious example. But it's part of a bigger pattern. Anything that requires understanding a specific human being — not humans in general, but this one person — AI fails at hard.

I tried using Claude to help me write a wedding toast for my brother. It gave me beautiful, moving paragraphs about "the journey of love" and "finding your person." Generic Hallmark stuff. My brother would've cringed so hard he'd have left his own reception.

The toast that actually worked mentioned the time he got lost in a Costco for 40 minutes because he was too embarrassed to ask for directions. It referenced an inside joke about his terrible taste in pizza toppings. AI couldn't generate any of that because it doesn't know my brother. It knows brothers.

Same problem with performance reviews. My manager asked me to draft peer feedback for a colleague. ChatGPT produced reasonable-sounding professional language. But it missed the specific thing that made this person valuable — her weird ability to de-escalate tense Slack threads with one perfectly-timed emoji. That's the kind of detail that matters. AI writes around it.

Anything Where Being Wrong Has Real Consequences

I almost sent a tax document with AI-generated numbers in it. Almost. Caught it because something looked off about a deduction total.

The AI had confidently calculated my home office percentage using floor space numbers I'd given it. The math was correct. But it used the wrong numbers because I'd mistyped my total square footage, and instead of flagging that my home office was apparently 47% of my apartment (it's not), it just... ran with it.

This is the pattern I keep seeing. AI doesn't know when something is obviously wrong in context. It doesn't have the common sense to say "wait, that seems off." It processes the inputs you give it and produces outputs. If your inputs are slightly wrong, it makes them confidently, professionally, legally-binding-ly wrong.

Medical stuff is the clearest example. I tested asking ChatGPT about a weird rash I had. It gave me a thorough answer covering multiple possibilities with appropriate caveats. The problem? It couldn't see the rash. It couldn't ask follow-up questions about texture or if it itched in a specific way. My dermatologist took one look and said "that's contact dermatitis from the new laundry detergent you mentioned switching to." Seven seconds.

For anything where being wrong costs real money, health, or relationships — AI is a research assistant at best. Never the decision-maker.

The Kick: When You Need to Actually Think

This is the one nobody talks about. Sometimes the whole point of a task is the struggle. AI removes it, and in doing so, removes the value.

I used to journal every morning. Twenty minutes of writing whatever came into my head. It helped me figure out what I actually thought about things. Last year I started using AI to "help" by generating prompts and even expanding on my initial thoughts. My journaling got longer, more polished, more articulate.

And completely useless.

The messy, repetitive, sometimes-going-in-circles process was how I discovered what I actually felt. When AI smoothed it out, I was just reading someone else's (well-written) opinions about my life. I stopped in February and went back to pure pen-and-paper. The quality of my thinking improved almost immediately.

Same thing happened with brainstorming. I was stuck on a project and asked ChatGPT to generate 20 ideas. It did. They were fine. Professional. Reasonable. I picked one and moved forward.

Three weeks later I realized the idea I would've eventually come to on my own — the one that was sitting in the back of my brain, half-formed, waiting to be discovered through the frustrating process of thinking — was better than anything on that list. But I'd never found it because I outsourced the discomfort.

The rule I use now: if struggling with the task is secretly the task, don't use AI. Writing to understand. Brainstorming to explore. Drafting to discover. These aren't problems to be solved faster. They're processes that create value through friction.

When Speed Isn't Actually the Goal

I thought I wanted to write faster. I tested AI writing assistants for months trying to increase output. I published more posts. Got more done. Felt productive.

Then I read back through six months of AI-assisted writing and three years of my pre-AI stuff. The difference wasn't quality exactly. It was... recognizability. The old posts sounded like a person with quirks and opinions. The new ones sounded like content.

The slowness of writing — the pausing, the deleting, the staring at a blank page — that's not a bug. That's where the voice comes from. AI removes the pauses. It fills every silence with reasonable words. And reasonable words, at scale, sound like nothing.

I still use AI constantly. For research. For first drafts of boilerplate stuff. For code I don't feel like debugging. But I've gotten much more intentional about asking: what exactly am I optimizing for here? Because sometimes slower, harder, and more frustrating is the actual goal.

The weirdest part is how obvious this seems now and how long it took me to notice. I was so focused on what AI could do that I never really asked what I was losing when it did it.

Heads up: Some links in this post may be affiliate links. I only recommend tools I've personally tested. Opinions are entirely my own.