First of all, this is an image of me. OK, not quite an image of me, it’s a drawing of me. Yeah, not quite either… it’s a drawing of me when I finally hit the button to publish this post and left the room. And there is supposed to be a chair somewhere, and some nice background, lots of paper notes I can never completely get rid of, a laser cutter, and who knows what else. But I am not an AI, it could draw all of that in mere seconds, and it took me a few minutes to draw this much. So I’ll stick to that, it’s a schematic image of me in front of my desk when I’m actually not around. There, I own it.
It is also a good illustration of what I wanted to write about.
Everyone’s linkedin feed is different, but, for some reason, mine is full of people repeating the same theme over and over lately. It’s pretty simple, it seems to make sense, and it also seems to be offering hope. Basically, it’s this: “AI can do things faster, and, often, better, but we need humans to own the outcomes”.
That line offers hope to those who used to feel we would all be doomed, AI would take over, software developers (and not only) would not be needed, and it would be the end of the world.
So, do I disagree? Probably no, but there are nuances.
In the last year, I have worked a lot with AI. What does it actually mean, though? In my case, it means that I worked on a few things:
– Using AI for creating my own software, almost daily, for almost a year
– Using AI to find information, summarize it, argue with me, argue with AI, discuss, come to the conclusions
– Using AI to understand how AI works – prefill, decode, memory bandwidth… ask google, their AI will give you some useful details and will mix it with what does not necessarily apply, you’ll spend the next few hours sifting through that information. I’ve been there, I did that
– Setting up local AI on AI MAX 395 with llamacpp, making it work, but ultimately admitting my defeat since that huge 128GB memory does not translate into the ability to actually run Frontier-like models
– Creating my own version of AI chat application that has now evolved into something that I am actually starting to respect – configurable models, agents, swarms, prompts, tools, mcp support, scheduled tasks… Most importantly, I can tweak it as I wish and when I wish now
What does all of that have to do with where I started? It’s simple, the next statement comes from the personal experience, but there is also a “theory” behind it.
When someone says people are supposed to own the outcomes/decisions, it sounds very encouraging and seems to be promising job security, but there are still a few problems with that:
- Owning a decision when you don’t understand the details is useless
- Understanding the details you need to actually own the decisions takes time and effort
Just imagine: AI gives you some code and says it’s working. Of course you can simply decide to take it for granted, but, if you wanted to own the outcomes, you’d have to look at the code, perform some testing, run some scenarios, trace the logic. Only then you’d be able to say that, in your opinion, that code is, actually working (or not).
Or take my naive illustration above. I did it myself in the good old Paint, and I assure you I can explain why it worked out that way in all the details. I can even tell you how inconvenient it is to be drawing with a mouse. I can also tell you why I did it – simply because I had enough of those AI-generated images in my linked feed today, yesterday, the day before yesterday, and, in general, for quite some time now.
Take another example – those recent Open AI’s announcements about solving multiple math problems and the mathematicians community reaction to it:
“OpenAI fell short of those standards, particularly where the mathematicians emphasized the need for human understanding of a mathematical result.”
https://techcrunch.com/2026/10/08/openais-math-solutions-arent-meeting-the-fields-standards-yet/
All of this, in my opinion, comes down to a very simple fact: you cannot own the outcomes/decisions if you lack understanding of those outcomes were achieved and/or if you cannot validate the logic used to make the decisions. Which creates a bottleneck: you get AI to solve problems faster, and then you need to figure out what AI did and verify it, which means you need to do enough review, dig into the details, understand them, possibly suggest changes. Suddenly, there may be a lot more time you may need to spend on it than you thought you would ever need to.
I tried it. AI would offer an explanation which would look plausible until you ask a question that invalidates the explanation. AI would write a piece of software which would look as if it were working, but it would be missing important parts. AI would suddenly decide to drop a database after getting confused by all the migration code, and you’d be thanking yourself for having a copy of it.
If you wanted to still own it in such cases, the overall process could only become that much faster. At some point you’d be at your own max capacity so to speak trying to figure out whether AI’s statements are accurate, if the database was dropped and recreated correctly, if that software is doing what it needs to be doing, and, if not, where things went wrong.
And then what? You would either have to slow down and become the bottleneck, or keep going without being able to own those outcomes and decisions anymore. Neither of that is really encouraging.