You might have seen the diagram before. The one Vincent Driessen put up on his website a few years ago to explain the concept of a Git branching model.

A few days ago, Microsoft published a page on their Learn portal that contained a diagram that looked … familiar. But only until you looked at it more closely. That’s when you noticed that something clearly wasn’t quite right. Arrows pointing the wrong way. Arrows missing their heads. Others making no sense at all or missing completely. And then, in one of the text boxes, the two words “continvoucly morged”:

For a brief moment, “continvoucly morged“ became a meme. People online were having a good laugh – also because it felt so emblematic of the current moment: Careless AI-generated slop even slipping into official publications of multi-billion dollar companies.

But, as Vincent wrote on his blog, this whole story is also very saddening:

I think we can all relate to that. It is a feeling we all know all too well, especially if we care about our craft – be it art, music, writing, coding, or design. Most of the time, AI output doesn’t take the human work it was trained with to the next level, but only hashes and rehashes it and vomits it out again as something that looks valid, something that looks finished. Something that resembles understanding and sometimes feels indistinguishable from magic – but only at a quick, first glance of the untrained eye.

I guess by now, we’ve all played around a bit with a few AI tools out of curiosity, be it LLMs, text-to-image or text-to-speech generators, audio or video generation models, and more. At least for me, the result has always been the same. Besides from all environmental and copyright concerns, all those tools feel like they can be useful – but only if you need the type of content you can actually expect from them: recycled knowledge at a somewhat decent level that looks impressive at first, but that can and almost always will be incomplete, error-ridden, often uninspired and generic, and sometimes even comically wrong. And you never know when it happens.

That’s why LLMs and all their seemingly intelligent friends, regardless of how much you try to compensate for their flaws by bending over backwards (also called “prompt engineering”), are only really useful when you’re experienced and seasoned enough to spot the errors in the output. When you read an answer and stumble upon a piece of information that – as confidently as it is presented – seems either just a bit weird or also plain wrong.

Maybe that’s also why the output of LLMs always reads the most impressive and convincing in areas we have no clue about and thus are least equipped to judge it. Quantum physics? ChatGPT explains it so well! But CSS? Or JavaScript? Oh, look at all those funny little errors!

I’m not saying that AI can’t be useful. It can summarise. It can draft. It can suggest alternatives. It can translate and check your grammar. But you will always need someone who checks and reworks the output. AI output always needs curation. Human curation. Professional curation.

And that’s why “continvoucly morged” isn’t just a funny typo on some lazily ripped off diagram on a Microsoft portal. It’s also symptomatic of something deeper. Of a fundamental misunderstanding in the way many people see AI and what it can do at the workplace.

The first thing many people jumping on the AI bandwagon, often experimenting with ways to replace the work of skilled professional workers, seem to forget: before AI, organisations relied on the quiet labour of skilled professionals to compensate for weak process and poor decision-making. Designers who caught inconsistencies in the brief. Developers who corrected flawed requirements. Editors who smoothed rough edges before publication. Real humans with experience and a love for their craft who genuinely cared about what’s being created.

Care filled the gaps.

With AI, many executives seem to trust the machines more than the humans who used to do that invisible work. But who’s filling the process gaps now?

At the same time, the rise of AI tools seems to accelerate expectations around how quickly work should be done. You know the story. The loom didn’t free the workers but actually demanded higher output. In the same way, AI doesn’t buy us time. It compresses expectations. It demands higher velocity.

“Half a day for writing that summary? ChatGPT does it in ten seconds.”

“That app prototype? Claude Code can scaffold it in the blink of an eye.”

“That Git branching diagram? Just regenerate it.”

And then, just review it quickly. Just tweak it here and there. Just fix the weird bits and it’s done.

We are told to let the machine do the writing, the coding, the scaffolding – even the thinking. Our new role, we’re told, is oversight. Supervision. A light editorial pass. And it’s done.

This creates a serious problem: if you don’t own the work, if you didn’t think it through and understand it, cognitive debt accumulates. Simply put: if you create a piece of work yourself or together in a team, your brain learns and remembers important aspects of the work and your decisions. If you use an LLM to generate it, all that learning doesn’t happen and so it takes individuals and teams much longer to actually understand what’s going on and, for example, how a piece of overly complicated code actually works in case something needs to be changed. In the case of Vincent’s diagram, the person (or persons) in charge obviously did not understand the diagram in front of them. And, as Margaret-Anne Storey writes:

But all this also means that humans are turned into mere correction engines. We apply our experience not to shape ideas from first principles, but to smooth out the edges of thoughtless, generic AI output. We are handed something that looks finished and are being asked to make it actually finished.

The underlying assumption: AI does in seconds what expensive humans take hours to do. Finally!

But that assumption not only confuses speed with substance and creates more cognitive debt. It also reveals a fundamental misunderstanding about the nature of work and human craft.

Because writing isn’t the typing itself or the word count. Coding isn’t the syntax or the amount of lines written. Scaffolding an app isn’t the folder structure and the component files.

The artifacts aren’t the work. They’re the residue of the work. And you can’t get them without the real work.

The real work is the process, the ugly first drafts, the reframing, the wrestling, the decision to remove a feature instead of adding one, the decision to replace a certain word because it carries the right weight and captures what should be said perfectly. The real work is the rendering of intent. The real work is understanding the problem and the solution. The real work is being inspired by something and building on it. And sometimes, it’s just following that gut feeling you can’t explain. But there are no shortcuts.

If we reduce professionals to reviewers of machine output, we don’t eliminate “the work”. We don’t make work more easy and efficient. And we also don’t increase the speed and quality of our work. We devaluate it. We displace it. We create more cognitive debt in our organizations. And I’d even say we create less happy and less motivated professional workers. Workers that might even stop going the extra mile because they suddenly feel undervalued and disrespected.

Because the writing is the actual work. The coding is the actual work. The thinking is the actual work. The understanding is the actual work. The designing of a diagram is the actual work. The struggling for the best outcome is the actual work.

The care is the work.

The care creates the value.

AI doesn’t care. AI is just another tool.

~