In partnership with

Prefer another language? Prefere receber em português? ¿Prefieres recibirlo en español? Choose here / Escolha aqui / Elige aquí: language form

The title "AI will always find something to change", with two numbers: every AI review round brings 4 to 5 suggestions, and from second rounds only 2 to 3 were ever applied, across every post.

Let me start with the answer, so you don't have to wait for it: AI will always find something to change. Every time. That doesn't mean your work is bad. It means the AI is doing what it was built to do.

Now let me explain why I'm writing about this.

An optimist with one concern

People who know me well know I'm an optimist about almost everything in my life. That doesn't mean bad things don't happen to me. Everybody has problems. If you think someone doesn't, just get a little closer.

Over the years, I learned to keep a bad moment small, so it doesn't spread from work into my personal life. Decades ago, I couldn't do this. Today I do it every time.

But there is one thing I see growing, and it worries me. It comes from AI, or more exactly, from how some people use it: AI reviews used without thinking, and passed on to others as orders. It can hurt self-confidence, mental health, and make good professionals feel they are not good enough.

That is my view. If you disagree, reply and tell me. I would love to read your point of view.

The review that came back in minutes

In Tuesday's post, I showed everything AI does for me: it helps with these newsletters, reviews my documentation, my texts and my emails. It's very good.

But I see more and more people use AI in a superficial way, and then share the result as a voice of authority. Like: "Here it is. Do it."

I don't have numbers for the whole industry. This is what I see in my own work: in my documents, in our Mission Critical team blog, and in conversations with colleagues. So let's collect some numbers together. At the end of this post, there's a one-click poll. I'll share the results in a future post.

Here is an example. At work, I often write technical documents: a technical opinion, an assessment for a customer. Think of them as the blueprints of a house, but for computer systems. Sometimes, when I think it's very important to have another pair of eyes, I ask colleagues I trust to review one before it goes to the customer. Not every time: everybody is busy, and most of my documents don't need it. But when I ask, I choose people who I know have the technical skills. I add one more person. It helps me make sure I stayed on track and understood the need well.

Before it even reaches them, my document has gone through a lot. At work, I use Microsoft Copilot with several checks. Some are technical. Others check that the document follows the way Microsoft expects it to be written. Large companies have rules for publishing and writing. If you don't follow them, you can have a problem. One example: at the end of every post here, you'll see a note. It says these are my views, not Microsoft's, and that I'm a Microsoft employee. My AI checks cover all of that, on top of my own experience.

Then I noticed a strange trend. Some colleagues put my document into an AI tool and sent me the result: "This is what I found that needs to change."

It's easy to see when a text comes from AI. You see it in the formatting and in the way it's written. And sometimes the feedback came back so fast that nobody could have read even 3 or 4 pages, understood them and thought about them.

Some of those suggestions are good, and I apply them. But most of them I don't. Here's why.

The Ultimate Claude Code Guide to ship like Anthropic engineers

AI will write 90% of code by the end of 2026, and only 10% of developers will stay relevant.

The engineers in that 10% aren't smarter. They just know how to use AI.

We put together the exact playbook Anthropic engineers use to make sure you're in the top 10%.

Sign up for The Code and get access to:

  • The Ultimate Claude Code Guide 2026 — 50+ tips and tricks to code 5x faster

  • The Code newsletter — learn the latest AI tools, tips, and skills to code faster with AI in 5 minutes a day

Why the list never ends

When I review my own posts with AI, I write a draft and ask for a critique. I get a list of recommendations. I apply the ones I think are relevant and ask again. A new list comes. If I applied every single recommendation and asked again, there would be another list.

You get the idea. It has no end.

Why? Because AI really is wonderful, and it will always find something that could be done differently. Not necessarily something that adds value. Just something different.

Here is a real example, from this post. Before publishing, I asked for an AI critique of it. The first recommendation was to cut my opening, the part about being an optimist, and go straight to the AI problem. I made it shorter, but I kept it.

❝

A big note: I am taking the risk that the opening hook may be weaker. But that is not an error. It's my voice, not the AI's.

That's one of the reasons I use more than one AI tool, as I explained on Tuesday. When one tool gives me recommendations, the other one looks at them and tells me: "I agree with this one. This one isn't needed." A cross-check, like asking a second doctor for an opinion.

So, when a colleague comes back and says, "You need to change this," I choose my words carefully, because I know it can create conflict. I explain why I won't apply the others. A few times I heard something like, "But it's needed, and I would like you to change it." Almost an order.

Sometimes someone has to give in. A few times, I made the change, even against my own opinion. I don't need to win every argument. But if I applied every recommendation I received, I would lose the authority over my own text.

And for a professional in a large tech company, reputation matters. If you let this happen, people may start to look at you differently: "He wrote a document, I reviewed it and found many things to change." But that person didn't understand that AI will always suggest something. And your credibility can be hurt by people who don't understand this process.

When I am the reviewer

I'm also one of the people who review posts for our Mission Critical team blog at Microsoft, an internal blog. Depending on the content, we approve it or suggest changes.

I've said it internally: my suggestions are suggestions. My colleagues can apply them if they find them useful. If not, they don't need to.

The exception is something critical, based on Microsoft's publishing rules. Then I mention it and explain: "This is not me asking. This is a Microsoft rule. Here are our instructions." For their own good, because publishing something against the rules can cause them a problem.

I don't simply tell people "change all of this because I think it's better." I feel it myself, with my posts and my technical documents. And the more AI we put on top of a text, the more polished it gets, and the further it moves from the human being who wrote it.

"Is it that bad?"

Now you can see how this can hurt self-confidence, and even create a dependence on AI. Someone writes a good text, sends it to AI or to colleagues, and gets recommendations. Then more. Then more. If you don't know that the list never ends, you may start to think: "Am I that bad? So many mistakes?"

There is a name for this feeling: imposter syndrome. It's when you feel you are not as good as people think, even when your work is good. It's real, and endless AI feedback can feed it.

I'll be honest with you: I felt it too. Reviewing technical documents came long before these newsletters. With the first documents I sent for AI review, my first reaction was: "So many errors. I need to rethink how I work and what I'm studying."

Then I read each item. Most of them were not errors. They were about how to describe and explain things.

That's also why I make the podcast debate about my posts: it helps me explain technical topics in different ways. I already know some analogies, and I created a few myself, but AI gives me many more. If I always explain something the same way, it gets repetitive. With different ways, I can explain it one way to one group today, another way to another group tomorrow, and see which one people understood best, like marketing teams do. Then I use the best one more often.

And I see this feeling in colleagues today. Some people who publish on our blog come to me and say: "I got your recommendation. Is it that bad?" They didn't read what I wrote with attention. They just saw the list of suggestions and thought they were doing something wrong.

So I explain again: "No. Read the beginning of my email: it's a suggestion. I didn't see anything wrong in what you wrote." Then I tell them what I learned, the same thing I'm telling you in this post.

Many of them don't know this, even working in technology. At Microsoft, they are a minority, to be honest. Outside tech companies, or in roles where technology isn't the main work, the number is much bigger. People open an email, read it fast, look at something else and miss a few important words. The reaction is "Is it that bad?" And it's not.

I'll confess one more thing. I check the critiques and the podcast debates on my phone. Every time I send a post for critique, I see this message: "Generating... Come back in a few minutes." If it takes more than five minutes, I start thinking: "Oh boy, this one is bad..."

NotebookLM on a phone, in the notebook Newsletter (WTP) Podcast. A new audio is "Generating... Come back in a few minutes", above the source wtp-post-0013 and the finished episodes for post 0012.

Nah. It's just a joke.

This Friday's tip: put a stop in the loop

If you share an AI review with someone, say clearly that it's a suggestion and that it came from AI. None of the people in my stories did this on purpose. Not one of them. But it can hurt, based on my own experience and on what people tell me about the emails I send.

If you use AI to review your own work, don't run it one, two, three, four, endless times. Put a stop in the loop. This is mine:

  1. One critique. I send the post for one review.

  2. I read it myself first. I already know which ones I think I should apply.

  3. A cross-check. A second AI tool looks at the recommendations, like a second doctor giving an opinion, and tells me which ones are valid.

  4. I apply only what I think is needed.

  5. One more round, only to learn. I send the new text for a second critique, to see what comes back.

I did exactly this with this post. Here is the first round, what I did with it, and what the second round said:

First round

Implemented?

Second round

Cut the opening about being an optimist

Partly: shorter, but kept

Same advice again: "aim the optimism at technology"

Explain terms with simple analogies

Yes: blueprints, second doctor

Same again, with new analogies: "a study buddy", "an over-the-shoulder editor"

Break up long sentences

Yes

Same again: split the same paragraph even more

Add a real example of a rejected suggestion

Yes: the opening, and this table

—

Turn "Your turn" into an action

Yes

Praised: "the ending works"

(Mine) Name the feeling: imposter syndrome

Yes

Praised

—

—

New: say that "more and more people" is my own experience, not industry data

Second round: four suggestions. Three were the first round again, in new words. I applied one.

Every round gives me 4 to 5 recommendations. And from all those second rounds, across every post I have published, I applied two or three recommendations in total. Only two or three.

That's the danger. And that's the stop.

Human vs AI: one click

This is the first poll in a series I'm calling Human vs AI, the same idea as our podcast debate. One question, one click. I'll only share the totals, never names.

Has an AI review ever made you doubt your own work?

Login or Subscribe to participate

Your turn

Next time you write an email or a document, try my five-step loop. Then reply and tell me: how many AI suggestions did you keep, and how many did you reject? And if an AI review ever made you doubt your own work, tell me that too. I read every reply.

See you next time.

Glossary

Term

What it means

AI (artificial intelligence)

Software that learns from a very large number of examples and can understand and write text, speech or images. In this post, it reviews documents and suggests changes.

Cross-check

Asking a second source, here a second AI tool, to check the result of the first one, so you don't depend on only one opinion.

Imposter syndrome

The feeling that you are not as good as people think, and that one day they will find out, even when your work is good.

Microsoft Copilot

Microsoft's AI assistant, used at work inside apps like Word, Outlook and Teams. It can review a document or draft a text.

NotebookLM

A Google AI tool. I use it to make the podcast debates and the critiques of my posts.

Podcast

An audio show you listen to on your phone or computer, one episode at a time.

Technical document

A document that explains a technical solution, an assessment or an opinion, often written for a customer. Like the blueprints of a house, but for computer systems.

Prefer to listen? Two AI hosts debate this post, one for it and one skeptic, on our podcast, We The People, in Tech or Not: The Debate: listen here

How I write: every post starts from my own notes and experience. I use AI tools to help draft and edit, and I review every word before it goes out.

Disclaimer: The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of Microsoft. The author is a Microsoft employee.