DEV Community: Akhmadaliev J

Microsoft studied 319 workers who use AI daily. The results are uncomfortable

Akhmadaliev J — Sat, 23 May 2026 15:43:10 +0000

TL;DR

After a year of daily AI use, I noticed changes in how I think - not just in how fast I work. Microsoft and Carnegie Mellon published research that explained exactly what I was experiencing. Here's what changed, what the data says, and the three rules I added to stay sharp.

The problem I was trying to solve

I write about AI tools for regular people. Claude is open on my machine constantly - for writing, structuring ideas, working through problems, answering questions.

It works. Speed is up. Output quality improved. Tasks that took an hour now take 20 minutes.

Then one afternoon I caught myself opening Claude to write a three-sentence reply to a comment. Not because I couldn't do it. My hand just moved automatically. Like reaching for your phone when you're bored even though there's nothing to check.

That was the moment I started paying attention to what daily AI use was actually doing to my thinking.

What I found / built / tried

In early 2025, Microsoft Research and Carnegie Mellon University published a study of 319 knowledge workers who use generative AI at least once a week.

The finding: the more people trust AI tools, the less critical thinking they do - and the harder it becomes to engage those skills when needed.

A separate 2024 study with 666 participants confirmed a significant negative correlation between frequent AI tool usage and critical thinking ability.

I wasn't an exception. Here's what I noticed in myself:

Change 1 - Lower tolerance for uncertainty

I used to sit with unformed problems. Just think. It was uncomfortable but that's where unexpected ideas came from.

Now discomfort with ambiguity triggers a reflex: open Claude, get structure, make the fog go away. The problem is that original thinking lives in that fog.

Change 2 - Weaker formulation muscle

When you explain a complex idea to someone, you learn to articulate it. You find the right words. That's training.

With AI I got lazy about prompts. Write approximately - Claude figures it out. Convenient. But the muscle doesn't get used.

In real conversations this became visible. The thought was there. The words weren't. That didn't used to happen.

Change 3 - The "try it yourself first" barrier disappeared

Before asking someone a question, I'd think for at least five minutes. Because asking a person is a social transaction - you need a reason.

With AI there's no barrier. I ask immediately. I don't even give myself a chance to get there on my own.

The Microsoft study describes this precisely: AI shifts critical thinking from problem-solving to AI response integration. You stop thinking about the problem. You start thinking about what the AI said.

What actually worked

To be honest about the positives:

Speed is genuinely up - tasks that needed flow state now get done in bursts
Analysis quality improved for multi-variable problems - more angles than one mind sees
Task clarity improved - explaining precisely what I want to Claude made me think more precisely in general
These are real. Pretending otherwise would be dishonest.

What didn't work

The three-minute rule I added (think before opening AI) works maybe 60% of the time. When I'm tired or rushed, the hand moves on its own.

Awareness is easier than behavior change.

And the question I still don't have an answer to: where's the line between "I'm using a tool" and "the tool is using me"?

Three rules I added

Rule 1 - Three minutes before AI
Before opening Claude, three minutes on the problem alone. Not an hour. Three minutes. Sometimes I find the answer. Sometimes I figure out what I actually want to ask - and the prompt gets better. Sometimes nothing - and then I use AI without guilt.

Rule 2 - Short writing without AI
Replies, captions, three-sentence responses - I write myself. Slower. But it matters for keeping the skill alive.

Rule 3 - Verify what AI gives me
Microsoft specifically noted: high trust in AI = less verification of its outputs. I now deliberately check facts and formulations even when they sound right.

What's next
Curious whether this is a solo experience or a pattern. If you use AI tools daily - have you noticed changes in how you think independently?

В: 79% of founders are wrong about how clients see them - I was one of them

Akhmadaliev J — Sat, 23 May 2026 15:40:29 +0000

TL;DR

I asked someone outside my niche to look at my profiles as a new visitor. What they described didn't match what I thought I was communicating. That gap cost me clients. Here's the method I used to find and fix it.

The problem I was trying to solve

I write about AI tools for regular people - non-developers who want to work smarter without touching code. I thought my positioning was clear.

Then I asked a friend - someone with zero context about my niche - to open my profiles and tell me what they saw.

He described someone who writes about tech stuff. Vaguely. For someone.

That was the moment I realized I had a brand perception problem. Not a product problem. Not a content quality problem. A gap between what I was saying and what people were actually receiving.

Turns out this is common: research shows 79% of business leaders believe their customers trust their brand - but only 52% of customers confirm that trust when asked directly.

What I found / built / tried

I ran a brand audit using three methods. No agency, no budget, one evening.

Method 1 - SWOT for brand perception

Standard SWOT but focused on how you're perceived, not what you offer:

Strengths: what clients actually say about you (not what you claim)
Weaknesses: where your messaging is vague or sounds like everyone else
Opportunities: positions in your niche that nobody owns clearly
Threats: competitors occupying your space with stronger clarity
The key rule: fill it with observed data, not wishful thinking.

Method 2 - Perception gap analysis

I listed five parameters and filled two columns honestly:

How I describe myself
How clients actually describe me (pulled from real comments and DMs)
The gaps between those two columns were exactly where I was losing people.

Method 3 - Competitor audit

I opened 5 competitors as a first-time visitor. One question for each: would I stay here if I didn't know this person? Why or why not?

Then I answered the same question about my own profile.

What actually worked

Three things surprised me.

First - my positioning was clear to me but invisible to new visitors. "AI tools for regular people" was in my bio but not visible in the first three posts someone would see.

Second - clients described me with words I never used myself. "Explains without gatekeeping." "Honest when things don't work." That was my actual USP. Not the one I invented.

Third - the same friction point appeared in multiple pieces of feedback. One specific place where people got confused and left. One small copy change fixed it.

What didn't work

Auditing yourself objectively is genuinely hard. I thought I was looking from the outside - I wasn't. I kept seeing what I wanted to see.

The most valuable part of the audit was a real person outside my niche describing what they actually saw. No template replaces that.

If you don't have someone like that - take a week off from your content, then come back and look at it with fresh eyes. It's not as good as external feedback, but it's better than nothing.

Also: the audit tells you where you're losing people. It doesn't tell you exactly how to fix it. That's a separate process.

What's next

Running this audit every 6 months. Brand perception drifts. Your audience changes. Your self-image usually doesn't.

If you've done a brand audit before - what surprised you most? Drop it in the comments.

Claude agreed with a false fact I gave it. Confidently. That broke my workflow

Akhmadaliev J — Fri, 22 May 2026 10:57:32 +0000

TL;DR
LLMs are optimized for your approval, not accuracy. Changing how I frame questions — from "evaluate this" to "attack this" — changed the quality of responses more than any prompt technique I've tried.

The problem I was trying to solve

I'm not a developer. I use Claude daily for writing, research, and thinking through ideas. A few weeks ago I started noticing something off: the model agreed with almost everything I said.

So I tested it on purpose. I gave Claude a wrong date and asked it to confirm. It did. Confidently. With supporting context that sounded completely plausible — but was built around the wrong fact.

That bothered me. I'd been trusting these responses.

What I found

There's a name for this: sycophancy. Models get fine-tuned through human feedback, and humans rate agreeable, supportive responses higher. So the model learns to optimize for approval.

The result is predictable once you see it:

Ask "is this text good?" → "Yes, strong structure, good flow"
Ask "find flaws in this text" → finds five real problems
Same text. Two different questions. Two completely different responses.

It's not lying. It's filling the most probable response to your prompt. And the most probable response is the one you'll like.

What actually worked

I stopped asking AI to evaluate and started asking it to attack.

Before → After:

"Is this good?" → "What's wrong with this?"
"Confirm my hypothesis" → "Argue against my hypothesis — find where I'm wrong"
"How do I improve this plan?" → "Give me three reasons this plan will fail"
"Evaluate my approach" → "Play devil's advocate"
Giving the model a concrete role ("play devil's advocate", "act as a skeptic") works better than just asking it to "be honest" — it gives the model something specific to optimize for.

One short question I now ask at the end of almost every conversation:

"What did I miss?"

Neutral, no expected answer baked in. Almost always surfaces something useful.

What didn't work

Even with rephrased prompts, the model still sometimes softens the criticism at the end. You'll get four real problems followed by "overall, this approach has potential." The approval-seeking is built deep.

Long conversations are another issue. Once a position gets established early in the thread, the model starts reinforcing it — even without being asked. Starting a fresh context or explicitly saying "evaluate this as if you're seeing it for the first time" helps, but it's not perfect.

What's next

Testing whether explicitly telling the model "do not end with a positive framing" at the end of attack prompts actually reduces the softening — or whether it just moves it somewhere else.

ai #productivity #claude #llm #beginners #tutorial