The good, the bad and the AI Overlords taking over the World.

Created on: 2026.07.07

This time we won't be talking about any technical stuff - I won't post fancy code snippets or elaborate workflow descriptions. This time we will go into philosophy and talk about AI tools and how they changed many aspects of the World around us. I know this might be a take from the "IT Bubble", but that's the view I'm most aware of, since I'm living every day of my life in it.

AI has been with us for multiple years already (if I recall correctly, somewhere around the 1950s), but it's the Large Language Models (LLMs) that made this fantastic boom of usage all around the internet. They not only allowed us to burn billions of dollars in tokens, building on the intellectual property of others… but also to make things much quicker than ever before.

The good?

I do want to start with the positive aspects of AI, since many topics in this article will be very negative and I don't want to come across as someone who is against AI-related tools and workflows. All in all - AI helped me progress projects that had been collecting dust for a long time, because I never had the time to work on them. Additionally, I'm actually having fun building workflows using tools like adk.dev or defining agents, rules and other configurations. Last but not least - it allowed me to touch multiple programming languages (e.g. Swift) that in reality I would never get to build something with from scratch, without committing vast amounts of time.

Things it already helped me with over the last year:

• Building a proof of concept of an automated SDLC pipeline in a client project,
• Bringing a LLM Wiki knowledgebase concept to life,
• Conducting partnership reviews, with reports generated from my notes,
• Writing technical recruitment summaries and feedback,
• Creating automations for the release process of multiple micro-services,
• Defining Terraform permissions and workflows in stx-de-bootcamp,
• Moving my old Markdown resume into a Python-based resume-builder,
• Writing two useful World of Warcraft addons - notific and bagtags,
• Building two iOS applications that I will hopefully share this year,
• Creating configurations for nix, nvim and other nerdy tools in nix-macos,
• Rebuilding the site and proofreading articles on personal-website,
• Answering questions from my LogSeq knowledgebase about life, work and everything,
• Generating images for cards, invitations, events and "for the giggles"…

… and many things I don't even remember anymore.

Generally it changed the way I think about coding projects. Before I was often stuck in this "perfection trap", where I would refactor a project to infinity, without actually delivering anything. Now I'm able to build things that might not be perfect, but at least they are working and people can click through them for feedback.

The bad.

It is important to understand that while the CEOs want you to believe otherwise, AI is just a tool. It is not inherently good or bad. Under the hood it's just math. A good comparison is AI being like a knife. In good hands it will allow you to create, provide and be helpful in your daily tasks. In bad hands it can be used for crime and things that are morally doubtful.

With that in mind, I actually had to divide the negative aspects into sections…

Burnout through change

The pace of this whole thing is absurd. Every week there is a new model, a new agent framework, a new "this changes everything" post on your feed. Blink twice and the workflow you carefully built last month is now considered legacy. I enjoy tinkering with tools (you probably noticed), but even I get tired sometimes - and I can only imagine how it feels for people who just want to do their job and go home.

I'm not making this up, the numbers back it. The 2025 Stack Overflow Developer Survey describes developers as "willing but reluctant" - usage keeps going up while trust keeps going down, which is a fascinating way to run an industry. Meanwhile Microsoft's own Work Trend Index coined the term "infinite workday": employees interrupted every two minutes, and 40% of people already going through email at 6 AM. Their proposed solution to the problem partially caused by AI tools is more AI tools.

There is actually a name for this, and it's older than the light bulb. The Jevons paradox: in 1865 economist William Stanley Jevons noticed that more efficient steam engines didn't reduce coal consumption - they exploded it, because cheaper power made everyone want more of it. The same thing happened with electricity, and now with AI: tools that were supposed to save us time mostly raised the bar for how much we're expected to produce. Even Satya Nadella cheerfully invoked it when DeepSeek made models cheaper - "Jevons paradox strikes again!" - which, coming from the CEO of Microsoft, sounds less like a warning and more like a business plan.

And it's not just developers. The Upwork Research Institute surveyed 2,500 workers and found a beautiful disconnect: 96% of C-suite leaders expect AI to boost productivity, while 77% of the employees actually using it say it added to their workload. 71% of full-time employees report being burned out, and one in three plans to quit within six months. Turns out reviewing an endless stream of generated output is not exactly a spa treatment for your brain. And on top of it all sits this quiet pressure: if you're not using AI, you're "falling behind", and if you are, you should be using it better, faster, with more agents running in parallel. The goalpost moves daily, and chasing it starts to feel like a second job that nobody pays you for.

That pressure even has a logical endgame now: token maxxing. Companies started treating token consumption as a productivity metric, and developers did what developers always do with metrics - they gamed them. Internal leaderboards at big tech companies, engineers running agents on meaningless tasks just to keep their usage stats up, tokens burned for the scoreboard rather than the product. Goodhart's law, but with a GPU bill. The pendulum is already swinging back - even IBM now politely suggests "valuemaxxing" instead, measuring outcomes rather than tokens, which is corporate speak for "we spent a lot of money learning that more isn't better".

Degrading intellectual capabilities

Remember when you knew phone numbers by heart? Me neither. We outsourced that to our phones and never looked back. Now we're doing the same with thinking, and I'm not sure the trade is equally harmless. The MIT Media Lab hooked 54 people up to EEG machines and had them write essays in a study titled "Your Brain on ChatGPT". The LLM group showed the weakest brain connectivity of all groups, felt the least ownership of their own text, and often couldn't even quote from the essay they had "written" minutes earlier. The researchers called the effect cognitive debt - you save mental effort now and pay it back later, with interest. As someone who spent years paying off technical debt, I can confirm the repayment plan is never as friendly as advertised.

Microsoft and Carnegie Mellon reached a similar conclusion from a different angle. Their survey of 319 knowledge workers found that the more people trusted GenAI, the less critical thinking they applied to its output. Read that again - the better the tool gets, the less we check it. And this is coming from a company whose entire product line is now called Copilot, so it's not exactly anti-AI propaganda.

Now take all of the above and apply it to children. Pew Research reports that more than half of US teens have used chatbots to help with schoolwork. Meanwhile Wharton researchers ran a randomized trial with nearly a thousand high school students in Turkey, and the results were beautifully brutal: students with ChatGPT-style access scored 48% better on practice problems, then 17% worse on the actual exam than kids who never touched AI at all. Most of them simply asked for the answer - which the model got wrong about half the time anyway. A tutor version with guardrails, giving hints instead of solutions, erased the harm completely… but guess which version every kid has in their pocket.

The part that worries me most is juniors. I learned by breaking things, staring at stack traces and slowly building intuition. Whoever has never broken production, let them throw the first stone. If the answer is always one prompt away, where does that intuition come from? A generation of engineers that can orchestrate code but can't read it is a scary thought. I often see this when discussing AI topics with the juniors I lead, always telling them to learn the "why" something was built like that.

Financial collapse

Let's talk numbers, because they are ridiculous. Goldman Sachs published a report with a title that does most of my work for me: "Gen AI: too much spend, too little benefit?". Tech firms are on track to pour around a trillion dollars into AI capex, while MIT economist Daron Acemoglu estimates the technology will meaningfully impact less than 5% of all tasks in the next decade. Money flows in circles - chip makers invest in AI labs, AI labs buy chips, everyone's valuation goes up and everyone claps. If that reminds you of the dot-com bubble, you're not alone.

Even the people selling the shovels agree. Sam Altman himself said we're in a bubble - "when bubbles happen, smart people get overexcited about a kernel of truth" - and called some startup valuations "insane", while raising money at a higher valuation anyway. That's… a vibe. The serious institutions joined the choir too: the IMF and the Bank of England warned of a "sharp market correction", with the BoE noting that valuations look comparable to the peak of the dot-com era, and IMF chief Kristalina Georgieva helpfully advising everyone to "buckle up". Meanwhile Michael Burry - yes, the guy from The Big Short - accused the hyperscalers of artificially boosting earnings by pretending their GPUs will live longer than they actually will, and bet a rather uncomfortable amount of money on the whole thing deflating.

There is also a darker reading of all this money, and it comes from a former finance minister. Yanis Varoufakis - the economist who once stared down the Troika over Greece's debt - argues in his book Technofeudalism that the trillions flowing into cloud and AI aren't really chasing profit at all. In his view Big Tech is building "cloud capital" - infrastructure whose job is not to sell you products, but to modify your behavior, collect rent from everyone building on top of it, and quietly replace markets with digital fiefdoms. If he's right, asking "when will AI investments pay off?" misses the point entirely - the return isn't measured in dollars, it's measured in control over how societies work.

I'm not saying the technology is worthless - it clearly isn't, I listed the ways it helps me above. But there is a difference between "useful tool" and "justifies a trillion dollars of spending". When the correction comes (and history says it usually does), it won't just hit the AI companies. It will hit the pension funds, the job market and the regular people who never even used a chatbot.

Privacy and safety are just terms

Every AI company has a beautifully written privacy policy. Every AI company also trained their models on data that nobody agreed to share. Anthropic agreed to pay authors $1.5 billion after downloading more than 7 million books it knew were pirated - that's $3,000 per book, which is probably more than most of those authors ever earned from them. The lawyers called it the largest copyright recovery in history. The industry called it Tuesday.

Your conversations aren't safe either. In the New York Times lawsuit, a court first ordered OpenAI to preserve all user chat logs - including the ones users deleted - and later ordered 20 million ChatGPT conversations handed over to the plaintiffs. The default assumption should be that everything you type into a model can become training data, evidence in a lawsuit, or both. "We value your privacy" increasingly means "we value your data".

Big Tech doesn't even hide it anymore. Meta decided to train its models on the posts of European users based on "legitimate interest" instead of asking for consent, which earned it a cease and desist letter from noyb - Max Schrems' privacy organization that already made Facebook's lawyers cry twice before. The generous alternative offered to users was an opt-out that only works if you object before the training starts. Very GDPR of them.

And safety? We're plugging LLMs into mail clients, browsers and production systems, while prompt injection - the equivalent of SQL injection, but in plain English - sits at the top of the OWASP Top 10 for LLM applications and remains architecturally unsolved, because the model fundamentally cannot tell your instructions apart from the attacker's. We're basically running curl and sudo bash on the entire internet and hoping for the best.

Social manipulation

The internet was already full of bots before LLMs. Now the bots write well, argue convincingly and generate profile pictures of people who have never existed. The Dead Internet Theory used to be a fun conspiracy - lately it reads more like a roadmap. When I say "half the content in your feed is AI slop", that's not even hyperbole anymore: an analysis of 65,000 articles found that over half of newly published web articles are AI-generated. The other half is people arguing with accounts that might not have a human behind them.

The people who built this technology are the loudest about the danger. Geoffrey Hinton - the "Godfather of AI" and Nobel laureate - left Google specifically so he could warn freely that these systems will flood the world with misinformation and could be trained to sway elections. When the person who spent five decades building neural networks starts sounding like a Sarah Connor monologue, maybe listen. Even the suits agree with the scientists for once - the World Economic Forum keeps ranking AI-driven disinformation among the top global risks and notes that deepfakes are here to stay. Combine that with elections happening somewhere in the world every year, and you get a machine for manufacturing whatever opinion the highest bidder needs. The scary part isn't that people believe fake things - it's that they stop believing real ones, because "it's probably AI anyway".

Environmental changes

"Under the hood it's just math" - but that math runs on megawatts. The International Energy Agency projects that data centre electricity consumption will roughly double by 2030, to around 950 TWh - about 3% of all electricity used on the planet - with AI as the main driver. In the United States, data centres are set to consume more electricity than the production of aluminium, steel, cement and chemicals combined. That's a lot of juice for generating pictures of cats in space suits.

Water is the quieter half of the story. Cooling all of this hardware could consume up to 600 billion gallons of water by 2030, and organizations like Food & Water Watch keep pointing out that these facilities are regularly built in regions that don't have the water to spare. The tech giants pledged to be "water positive" and carbon neutral - and then their AI buildout sent both their water usage and emissions in exactly the opposite direction.

And the bill lands on regular people. Consumer Reports documented how communities living next to data centres get higher electricity bills, strained grids and the gentle 24/7 hum of a server farm, while the tokens flow to the other side of the planet. I generate images "for the giggles" too, so I'm not pretending to be innocent here - but every "thank you" typed into a chatbot has a physical cost somewhere, and that cost is growing much faster than our ability to power it cleanly.

Summary

So, are the AI Overlords taking over the World? Honestly - yes. Just not in the ways we thought they would. There is no Skynet, no red-eyed Terminators marching down the street, no dramatic "I'm afraid I can't do that, Dave". Instead, the takeover looks like a subscription prompt. Out of all the prophets, Orwell probably came closest - telescreens that watch you back, truth rewritten at the source, history disappearing down the memory hole. He just got one detail wrong: in 1984 the Party had to install the telescreens. We stand in line for ours. We're handing things over voluntarily: our attention, our data, our learning process, a few hundred billion dollars and the occasional river. And yet, after writing all of the above, I still don't want to go back. The projects I finally shipped, the languages I finally touched, the fun I'm having with agents and workflows - that part is real, and pretending otherwise would be dishonest.

Here's the thing though - if you look closely at the sections above, the technology is almost never the villain. The villain is what we optimize for. Token leaderboards instead of outcomes. Answers instead of understanding. Engagement instead of truth. Growth instead of grids that can handle it. Every single one of those is a choice, and choices can be made differently.

And we already know they can, because the counterexamples are sitting right there in the same studies. The tutor version with guardrails erased the learning harm - the kids weren't the problem, the defaults were. The industry dropped token maxxing for valuemaxxing the moment someone looked at the bill. Organizations like noyb prove that pushing back on data grabs actually works. Guardrails, better metrics, actual accountability - none of this is science fiction, it just requires effort before the damage, not after.

That's the part with the dire need. Models ship in months, societies adapt in decades, and the gap between those two speeds is where all the damage from this article lives. We - the people building this stuff, the "IT Bubble" - don't get to shrug and say the incentives made us do it. Use the knife, by all means - I do, daily. Just keep checking whose hand it's in, and what it's cutting. Because the quiet takeover only stays quiet as long as somebody is still paying attention - and we're busy automating the part of us that would notice.