Observations from three years of AI

Three years ago I wrote My thoughts on AI code generators just as AI coding was taking off. AI has improved immensely since then.

My journey in the tech realm really began with my degree in Computer Science, starting in 2018. I don’t know how the AI landscape looked before that, but as I did my degree, AI was hyped up everywhere. Everyone wanted to learn AI and machine learning. The university even began creating an AI degree (which I believe was launched the term after I left in 2021). Large Language Models (LLM) as we know today didn’t exist yet, but every CS student wanted to learn about AI. Initially, I too saw AI as cool and interesting.

Over time, the glamorous world of AI started to fade in my eyes. I took an AI course and then did my Bachelor’s thesis on it. I realized it wasn’t nearly as fun as everyone made it out to be. Tweaking parameters to get slightly better results and then waiting for the Neural Network (NN) to train. I should clarify, my main gripe was with NNs. I loved and still do love algorithmic AI based on heuristics, for example game theory and alpha-beta pruning. Unfortunately, that is greatly overshadowed by NNs and LLMs today.

All this is to say that I was tired of AI before LLMs even existed. When the news of LLMs came I was always sceptical to its practical use.

Impact on work

Rotting the brain

What happens when you stop using your brain? It rots. This is true for the brain in general. Going months or years without doing a skill inevitably dulls the skill; this is nothing new for AI. Getting back into it usually isn’t too difficult. The risk with AI is doing the task without using the skill. Especially when the expectation is that you should know how to do it and how it works. Speaking as a software developer, my job is to be an expert at specific fields relevant for my employer. If I don’t actively spend time learning and debugging the application I won’t have the necessary knowledge. It is because I have spent months and years working within the field, debugging at every level, that I have deep knowledge of it. I can answer a lot of questions off the top of my head. Sometimes, AI cannot even answer these questions because they are rather niche and the answer may come entirely from experience, with no true source being written down anywhere.

I am afraid handing over my entire daily work to AI will result in me developing a lot of code that I have little understanding of. The better the AI is, the less time I will have spent working on the code. If it got everything right on the first try I would only have to review the code and test the feature (which is fun until something breaks). After a while I would lose the sense of the details because I never touch those parts anymore. Eventually, I’ll have moved up an abstraction level and am now a manager or product owner, with no human below me.

Personal development is heavily inhibited by AI. Outsourcing thinking means never having to think for myself. There is no opportunity to get better at the things I do. The things I love. Leading to over-reliance on AI. Even boring tasks are good to know how they work. Lifelong learning is a big part of who I am.

Pressure from management

At my job, management were early to jump on the AI bandwagon. Our CEO decided that we need to be on the forefront of AI. We’ll get left behind if we don’t. Microsoft already runs our entire IT infrastructure so getting their Copilot for everyone was the first step. Some employees have gotten to demonstrate how they use AI so others can learn. They have brought in external consultants to teach us how to better use AI. I always stayed on the sidelines just watching, trying it occasionally. It’s not useless but I’m still not entirely sold on it. Every management presentation mentions AI. They scheduled bi-weekly (optional) meetings for knowledge sharing. AI was always optional, but it was omnipresent.

GitHub Copilot has also been a large part of the AI rollout. Which is what primarily affects me as a developer.

My breaking point came when it was no longer optional. During the annual performance review I had AI put into my goals for the next year and felt like I had no say in it. Later in the year there was an optional seminar with an AI consultant. I didn’t feel it was optional. I was “strongly recommended” to attend it due to my “goals” of using more AI.

No matter how arduous or mind-numbing a work-related task is I will happily do it. But I went to that seminar with resentment and no intention to learn, and walked out of it the same way. Pressure is not a good motivator.

Pressure from Microsoft and AI generated messages

Every day when I open Teams it wants me to pin Copilot to the sidebar. The only options are “Pin” or “Maybe Later”. Teams is already damn slow as it is (on Linux at least). I have tried pinning it, but if I then unpin it I start getting the notifications again. I use Copilot occasionally, but when I do I go to the website copilot.cloud.microsoft.

The pin requests stopped after some number of weeks. But Microsofts insistence on Copilot mentally hurts. Sometimes it almost feels like they treat you like a child. “Hey, you can use a Copilot to write an email if you need some help”. Thanks, but I am fully capable of writing an email myself as I have done for many years. I have received AI messages myself. It’s not fun when you put effort into a conversation, only for the other end to give a generic AI reply. My worst experience was a support case I had with a supplier. I gave a long detailed explanation of the bug, everything I had tried, and pointing to what I believe the issue was. Only for them to a few days later send a fully AI generated response, of which about 95% was asking about things my initially message already covered, either because I’d already disproven it, or it was not relevant to my use case.

Human conversation and knowledge sharing

After the COVID-19 pandemic, many companies wanted employees to return to office under the guise of teamwork and collaboration. While work-from-home is nice, I do agree with that sentiment (but a few days from home here and there doesn’t hurt). Being in the office opens up for more natural conversations between people. Conversations that might not happen otherwise. And it builds social connections.

With AI capable of answering most questions, what reason is there to talk to our colleagues in our day-to-day work? Our colleagues will be out of the loop more than ever. Knowledge silos become more prevalent. Or are silos perhaps less of a problem because now the AI sits on all the knowledge, and all context is lost when the chat slowly fades further and further back in your Copilot history, never to be seen again?

Talk to your colleagues! Everyone has unique knowledge and experience to share.

AI elsewhere in life

Misinformation and slop

No LLM will ever be perfect. Neither will humans, but good search algorithms tend to filter out the slop. Trusted sources are presented at the top (though, recently even this is questionable as Google search itself has arguably become worse). We can look through the sources, learn things and get some context for the information.

AI is built on convenience, something everyone wants. One of the worst offenders is Google AI overview. Always there at the top, providing a quick answer to your question. No need to look further, assuming you trust the AI. Google AI summaries have been shown again and again to give troublesome information. The internet already had a problem of misinformation, AI only amplifies it many times over. This is another case of rotting the brain. When a confident answer is readily available it is very easy to trust it.

To get rid of the AI overview I recommend using alternative search engines. DuckDuckGo allows turning off their AI overview. They also offer a search with no AI features whatsoever.

Unfortunately, this doesn’t stop all the slop generated on various websites.

Art, videos, and music

I’m very much against AI art in general. There are many arguments against it in terms of copyright and how it hurts artists.

Images and videos can still be somewhat distinguished from human-made (but not always). I think there’s also a difference in how the art is appreciated, for me at least. When appreciating visual media I look at the details of it. Visual media tends to be connected to some message. For example, advertisements. Coca Cola famously made an AI commercial (2024, 2025). Seeing that invokes the “feeling of slop”, as if someone wasn’t bothered to use real art. If something posted online is clearly AI then it’s almost automatic that I won’t “appreciate the art” in the same sense as I would a really good human-made painting. If I know something is human-made it’s easier to appreciate the work put into it. This sentiment has also been seen in the game industry. Gamers can’t appreciate the game to the same level when parts of it are AI-made.

AI generated music, however, has given me a different impression. Maybe because my journey with it didn’t start out hearing all the slop first. Or maybe because I appreciate music differently. For me it started with the song A Feeling We Can’t Hide - Lyra Vesperi, which at the time I didn’t know was AI generated. I liked the song. As I discovered the artist used AI for the music I wasn’t immediately taken aback. After some reflection I decided I still appreciated the music. I’m by no means a music expert. If I like the sound I will dance along to it. But I won’t actively seek out AI music to listen to.

Cost of AI

Money has always been a question surrounding LLMs. People are “tokenmaxxing”, trying to utilize their tokens in the most efficient way possible. All while AI companies are still wildly unprofitable. Everyone is betting on it becoming profitable one day. The bubble is growing and I expect it to one day burst. Ask yourself: “Do I want to be fully reliant on AI when the costs skyrocket?”.

The cost isn’t entirely monetary either. The datacenters running all of this has taken a significant toll on their surrounding areas, water and noise pollution, and a big jump in electricity cost for the population.

My post in 2023 focused only on AI code generators and their legality. Today there has been many court cases. However, the question of AI training material still hasn’t been resolved in any court. OpenAI has openly admitted that they wouldn’t be able to make LLMs if they didn’t use copyrighted material.

“Because copyright today covers virtually every sort of human expression – including blog posts, photographs, forum posts, scraps of software code, and government documents – it would be impossible to train today’s leading AI models without using copyrighted materials,” said OpenAI — The Guardian

OpenAI originally set out to develop AI responsibly due to fears of how it could negatively affect the world if handled irresponsibly. Today, AI is rotting our brains, spreading misinformation, ruining art, building a potential economic bubble, destroying the land, and all this by possibly breaching copyright law. It touches every corner of society. All in the name of chasing profits.

To quote a Calvin and Habs sketch:

THEN THE BUSINESS SHOULD NOT EXIST!