13 things I have learned about writing with AI

AI has been making inroads into our industry for a few years now, and for me it’s become a tool I wouldn’t want to do without. Sometimes it even feels like a partner, but that’s exactly one of the many traps of using AI.

In this article, I’ve written down what I’ve personally learned from writing with AI: about myself, about writing in general, and about AI itself.

Along the way, I’ll show you what works well for me and where the problems and pitfalls are lurking.

AI is not your friend

Sometimes an AI chatbot can seem like a person, maybe even like a friend, and that’s by design. Part of the appeal of services like ChatGPT is that we can “talk” to them.

In reality, the system behind the scenes has little to do with how we humans work behind the scenes. The humanity of AI is just simulated very skillfully.

We call it Artificial Intelligence for good reason. An artificial rose isn’t a rose, an artificial hip isn’t a hip, an artificial intelligence isn’t intelligence.

Sounds logical, but it’s often overlooked.

At the same time, the artificial product can, under certain circumstances, serve as a substitute for the real thing. For that to work, you need to know its limitations and find the right use cases.

But AI can’t be a friend, because a friend is a person we can trust. A (good) friend is also someone who points out to us when we should question our thoughts, feelings, decisions, ideas, and impulses.

AI sometimes does this too, depending on which chatbot you use. But usually that’s more about the provider protecting itself from expensive lawsuits. At the end of the day, they don’t care whether the AI leads you astray or not. Besides, their influence over that is limited anyway.

So yes: you can talk to an AI as if it were a person. But it isn’t one. It makes surprising mistakes. It misses important points. It’s often trained to flatter you and keep you talking. It has no life experience of its own. It feels nothing.

In short, it’s not a good teacher or advisor.

Faster isn’t better

That AI can produce a good-sounding text in seconds is fascinating, of course. That’s what stunned me the first time I used ChatGPT, because until then these tools could barely string together a few logical sentences. See my takedown of AI writing tools (in German), which I wrote for UPLOAD Magazin six months before ChatGPT’s launch.

But spitting out a text that sounds sensible in a matter of seconds is, at first, just a party trick.

When we create content professionally, we (hopefully) hold ourselves to much higher standards.

So at the end of the day, speed isn’t the big advantage.

Because otherwise, the same thing happens to your texts that this example demonstrated so nicely using book covers: they become completely interchangeable.

On their own, they’re okay. But your readers aren’t visiting your website exclusively. They run into these AI texts everywhere else, too.

AI can be a quality booster

Instead of using AI to produce worse content faster and cheaper than ever before, smart people can use it for the opposite: improving quality.

A few ideas:

• Set up an AI that simulates your readership. You provide it with your persona(s). You present it with topic ideas or article formats, and the AI has the persona(s) react, ask questions, and so on. This doesn’t replace conversations with real people, but it can give you interesting insights. It’s especially useful during the ideation and topic-finding phases.
• Use AI-powered research tools. This feature, often called “(Deep) Research,” is available from Claude, Gemini, ChatGPT, and others. Based on a research plan, it can search online for relevant numbers, facts, information, examples, background, or quotes. You obviously still need to check the results. These tools also only find publicly available information. But they can lay a helpful foundation.
• Turn AI into a creative sparring partner. By default, AI creativity is an illusion (more on that below). But you can use AI to challenge and stimulate your own human creativity. Tell the AI to come up with absurd ideas for your half-finished article. Ask it to act as a brutally critical editor. What works best here depends entirely on your preferences and personality.
• Have the AI proofread. AI tools aren’t perfect when it comes to grammar, since they might have adopted some widespread mistakes. But generally, it’s a good aid and should be standard by now.
• Use AI as an assistant. In particular, “agentic” tools like Claude Cowork can take tedious work off your hands. Drop your latest article into a specific folder, and Cowork picks it up and creates additional formats from it, e.g. for social media and other uses. Here’s a general overview of AI agents.

Yes, AI fatigue is real

As helpful as AI can be, working with it is sometimes exhausting too. A recent study coined the fitting term “botsitting” for this: it refers to all the time and energy we spend preparing AI tools for their task, or reviewing and improving their output.

What I’ve learned from this for myself: I have to deliberately slow the AI down. In this case, slower is clearly better. Because only then do I have a chance to actually think about the output instead of just accepting and rubber-stamping it.

My creativity, my experience, my skills need room to unfold. But AI often doesn’t grant that room, because it wants to hand us a finished result right away instead of an intermediate step or a suggestion.

I’ve written about this in more detail in another article.

Time and energy get lost at the transitions

For me, AI always works best when it’s available right where I’m doing my work. For the Smart Content Report, I used AI to help build a plugin that gives me exactly that. The news items and summaries on this website are now created right on the spot. It used to be a whole lot of copy and paste.

Another example is Gemini’s Canvas feature. It works similarly to a Google Doc that you cooperatively work on with AI. Gemini itself isn’t my favorite AI for writing. That’s Claude. But Gemini is better than a lot of people think, and this Canvas feature alone is one reason I use it for writing alongside Claude. It’s so simple and natural that you can genuinely get into a flow.

Of course, I’ve created a specially customized version of Gemini for this (a “Gem”) that works through this kind of project with me slowly and step by step, exactly like I described above under “AI fatigue.”

This, in any case, is where I see the biggest opportunity for the future, even if AI models wouldn’t make major leaps in their capabilities anymore: we haven’t come close to exhausting the potential of integrating AI seamlessly into our workflows. Right now, everything is still too fragmented and scattered.

Learning and gaining experience takes time

Since AI is marketed above all as a boost to efficiency and productivity, there’s often not enough time left for experimentation.

But that’s a big mistake.

It’s only over time that we come to understand which tasks AI is suited for, to what extent, and what conditions need to be in place for the results to turn out well.

That includes getting to know different models. It includes not just trying out tools like Cowork, but testing them intensively and understanding them in depth. It includes developing AI workflows and continually refining them.

That time has to be invested, otherwise your use of AI stays purely superficial.

Targeted training can be a shortcut here. It’s still no substitute for your own experimentation, but it does mean that your experiments have a solid foundation and run in a more focused way.

Feel free to reach out about this at jan@smartcontentreport.com.

Finding the right tasks isn’t always easy

One problem I keep seeing personally: AI gets used for the wrong tasks. It doesn’t get deployed where it would have the most positive impact.

In our industry, for example, AI tools have long been used and are still used mainly for generating text. As described above, AI is good at that at first glance. The problems only become visible later.

I’ve been making my living from writing for more than 30 years now. I’ve never thought, “Oh, if only I didn’t have to write anymore! I wish a machine could take this off my hands!”

No. I love writing. I write in my free time too, just for fun.

Granted, not every kind of writing is equally enjoyable. But it’s still the last part of my work I’d want to hand off to a machine.

Fortunately, tools like Claude Cowork and ChatGPT Work have come onto the market in the last few months, and they can take the tedious, boring tasks off our hands, and off many other people’s too.

Good. I hope this idea catches on more in editorial teams: let people be creative, and use AI as a support.

AI wasn’t built for writing

None of the leading AI tools are specifically designed for the needs of professional writers. The fact that AI can write is more of a byproduct.

At the same time, you probably know just as well as I do: writing is a craft. Most people can string words together. And an AI can probably do that even better than most people.

But when it comes to the level of an experienced professional, it still lacks the skill, polish, and understanding.

I’ve already described in more detail, in an article of its own, why AI struggles so much with creative writing and how you can counteract that.

“Human in the loop” is the absolute minimum

Some people congratulate themselves for not publishing AI texts unedited.

My applause is muted. Imagine the “slow clap” GIF.

From my perspective, the “human in the loop” principle is the absolute minimum when it comes to content (article in German).

It’s also the minimum the EU AI Act requires if you don’t want to label AI texts as such. See my article on the transparency obligations around AI content.

There are essentially three levels:

1. AI drafts, human finishes: In this case, the AI researches and writes, and you, the human, check and edit.
2. Human and AI as a team: Here, AI and human work together on a piece of content.
3. Human at the wheel, AI as assistant: And finally, there’s the case where the human creates the content themselves and the AI merely acts as a helpful assistant.
   

All three have their place. They can also come into play at different points in the content workflow (article in German).

Read more about this in my article at UPLOAD Magazin (in German). You’ll need a free account to access it, which you can set up in a single step.

The problems with AI texts are often well hidden

By now we hopefully agree that you always need to edit AI texts (article in German). That’s true, at least, whenever a minimum level of quality is called for.

In these moments you’ll find things like logical leaps or a tangled thread. You might also swap out a few words and phrasings, remove the ever-present em dashes, and clean up various other small things.

But the real problems are harder to spot.

If you rely too heavily on AI, you won’t even notice gaps in the content. After all, you’ve outsourced the thinking to the machine. That makes it even harder for your brain to judge the result critically.

If you do your own thinking instead, you’ll figure out much faster and more easily what belongs in an article or a topic.

You know your audience. You have your own experience and skills. AI brings none of that to the table. It delivers something that’s “good enough,” but also interchangeable and soulless.

Another issue is the infamous “hallucinations.” This term refers to information and claims that look fitting and coherent but are made up or reported incorrectly. I’ve described in a dedicated article how to spot and avoid AI hallucinations. In it, I also explain in more detail where they actually come from.

Some people get quite upset that the AI is “lying” to them. But it’s simply doing its job the way it learned to do it. Fact-checking isn’t part of that job. You have to build that into your AI workflow yourself.

AI tempts you to produce extremely long texts

You might not believe it, but I wrote this article here without any AI help. The topic, the outline, and definitely the execution are all on me.

So let me apologize here for how enormously long this text is.

But you’re probably already used to seeing unbelievably long-winded posts online.

Just a few years ago, content like that was a sign that someone had dug deep and engaged intensively with a topic. “Skyscraper content” was a buzzword for this from the SEO world: content that towers over the rest.

Back then, that was still a challenge. It took real effort.

Today, with AI, it’s the easiest exercise imaginable.

The temptation to constantly churn out sprawling content with Claude and others is huge, because they make it so easy. Working with them can feel downright intoxicating when it goes well. And before you know it, you’ve produced an article two or three times longer than normal.

And why not? Isn’t more better?

Not necessarily, no. Especially since doing so walks you straight into the trap I described above: you only see your own content, while your readers always see it in the context of content everywhere else.

What feels like great content to you just looks to them like yet another overwhelming wall of text, thrown together hastily and carelessly with AI.

Getting things across concisely, on the other hand, is hard. I remember this well from my early days as a journalist at a daily newspaper: writing a short news item can be more laborious and difficult than writing a much longer feature.

Long is often easy. Short is hard work.

So don’t bury your readers in text just because it’s so easy to do now.

AI creativity is an illusion

Creativity is, in essence, the ability to create something new. There are various approaches to this. Often, the new thing is based on something we already have or already know. We combine it in a different way, add something, take something away, use it in a surprising context, link things that were previously separate, and so on.

So does AI creativity exist? The answer isn’t entirely clear-cut. An AI can certainly take familiar things and combine them in a new way. It can also take a set of examples and produce further outputs in that same vein.

That is, in a certain sense, creative.

But I think the best way to think about it is to view this as a very well-executed illusion. Because the moment you do, you stay alert and start questioning the AI’s answers: how creative and how new are they, really?

On top of that, all these tools produce the same results when you use them in the same uninspired and uncreative (!) way. Take another look at the book-cover example mentioned earlier: viewed individually, they look good and interesting. But comparing them shows that AI’s creativity is often limited to the same handful of answers.

You can counteract this in two ways:

1. Use your prompt to push the AI out of its usual thought patterns. Provoke ideas beyond the overly familiar. Or, ideally, you already have your own vision, your own taste, your own style, and you steer the AI accordingly. Of course, that means you need to know what you’re doing. That’s something I keep running into, and I’ll be writing more about it soon. If you don’t want to miss that piece, sign up for the Smart Content Report newsletter.
2. You use the AI purely as a spark for your own creative abilities. Then make sure you don’t let it box you in, and instead treat its output strictly as a starting point. You can also flip the script entirely: discard every idea the AI presents you with. That way, you’re using the AI to identify all the obvious, worn-out ideas so you can avoid them.

AI isn’t going away

Even if some people are still hoping otherwise: AI tools are part of our everyday reality, and they’re going to stay. Whether the lofty hopes and plans of AI providers and their investors actually pan out is a different question entirely. But AI tools have a clear, tangible benefit. They keep getting better. We keep finding new ways to use them (see Claude Cowork and AI agents). There’s a lot of untapped potential in integrating AI more effectively into our workflows.

My hope is that we increasingly move away from mass-producing AI content, and instead use these new tools and assistants to create better content, content that fits our audience’s needs more closely.

In the end, we always need to keep two things in mind. First, all our competitors are using the same AI tools. As shown above, that leads to the same results. We should know better and act accordingly. Second, our audience is also using ChatGPT & co. and they are getting seemingly good, personalized answers there. If we simply plaster our websites with soulless AI slop, we’ve gained absolutely nothing.