•
The thinking machines
I am impressed by what AI can do, but, I am increasingly concerned about what follows from it.
AI is one of the few technologies that can make me feel impressed and uneasy in the same conversation. The first models I tried were limited, but their capabilities have grown quickly. I can ask it to write code, explain a difficult subject, or generate an image and it will do it in seconds. Then I start thinking about the machines behind it, the electricity they use, and what happens when we build an entire economy around models.
This post is not meant to argue against AI. Rather, it is an expression of thoughts I’ve been having lately.
The price of progress
In my social circles, I am the tech guy people call when they want a good deal.
Recently, a friend asked whether a new phone could really cost nearly half their salary, while another wanted to know why the gaming console he had planned to buy had gone up 3x its starting price.
Neither of them care about GPUs, memory, or NAND flash. You see, they don’t care about AI. They just want a phone that works and a gaming console they can afford. I try to explain that AI companies are buying many of the same components used in the devices they want. So, when more supply is pulled towards data centers, the prices of the thing they want also go up.
If fewer people can afford to buy hardware, manufacturers will eventually have fewer reasons to make the products that enthusiasts and ordinary users want. Seeing that happen to some PC-component manufacturers does make me uneasy. Companies producing these goods will disappear and they will take those jobs and products with them.
The invisible bill
When I use a cloud-based model, I am paying for the amount of text that goes into the conversation and comes back out again. The provider measures that text in tokens and charges for it directly or through a subscription. Because the interaction resembles an ordinary conversation, the underlying transaction fades into the background.
The output is not guaranteed to be what I actually want, so I prompt it again and incur another cost. After a while, the mental model of “this costs real money” disappears. Spending 8,000 tokens does not feel like spending $15.
I’ve recently learned that this is called “payment decoupling” and it is not a new trick. Subscriptions, credits, points, monthly bundles, and in-app currencies have always separated us from the real cost of things. Conversing with a machine makes that distance feel abstract, because the product is a conversation (or a set of instructions running in the background).
Considering what these systems must cost to build and operate, the token price in the tools I use often feels subsidized. Maybe that is temporary, or maybe it is a way to build habits before prices change. I have seen many services begin offering something good for a agreeable price, but later become more expensive once people depend on them (yes, I am looking at you, streaming-services).
Privacy and ownership
Where did all the material used to train these systems come from, and what happens to the private material people put into them? I wonder whether training data includes pirated books, audio, video, and software, how much of the internet was scraped, and what each model is actually based on.
Books are being bought in bulk, scanned, and destroyed for AI training. Part of the reason is probably copyright. Nevertheless, if books are scanned and destroyed because the text is useful to a model, something of value for future generations has been lost, even if that means that the model becomes more capable.
You have might seen in the news that there is now an alleged AI-generated solution to the Navier–Stokes Millennium Prize Problem. Solving one of these problems can award you with a one-million-dollar prize.
The two mathematicians, Tristan Buckmaster and Levent Alpöge have independently been making progress on the problem. To make a long story short: They claim to have used AI to develop their solutions. Shortly thereafter, OpenAI published a solution to the problem.
It is alarming if someone feeds private work into a model and that model later produces something related. Also, If the model learns from previous sessions and then re-uses that data in later iterations, who owns the idea?
It saddens me that we’ve just spent a couple of years talking about privacy (think GDPR), and this is somehow where we end up.
The physical cost
There was an enormous push towards sustainability and decarbonization a few years ago. Corporate sustainability reports were full of images of green fields and wind turbines and tech companies had entire segments covering their sustainability progress.
Take Google, its 2026 Environmental Report covers the 2025 financial year and the first half of 2026. The report openly describes a tension between the growth of AI and environmental stewardship, while also presenting efficiency improvements and clean-energy expansion as ways to reduce the impact. Unsurprisingly, Google’s argument is that these improvements could eventually make AI’s environmental benefits greater than its costs. Rather than being negative, I actually applaud Google for being this open. The scale of the environmental impact this has is hard to ignore.
Apple presents a similar tension in a different way. But they focus more on the fact that most of their models run locally, rather than in the cloud. But, when they run in the cloud, they do so on 100% renewable energy.
Effort from both Google and Apple may be genuine, but they do not answer how much additional infrastructure is required when AI features become part of other software.
It saddens me that the environment is never our top-priority. As soon as new opportunities emerged, tech-companies rushed towards the gold mine without even thinking of the environmetal impact this might have.
What we lose together
One of the things that has kept us alive for a long time is our ability to collaborate. We demonstrate things, communicate, and solve problems together. It is not always “efficient”, but it often produces something better than any of us would have managed alone. A colleague who asks a difficult question can reveal a weakness in an idea, while someone who has solved a similar problem can provide context that was never included in a training set.
I do see the positive sides of collaborating with AI. I mostly use it for brainstorming, and it can produce many possible solutions in a short period of time. When I work in a large codebase, it has sometimes pointed out things that I missed. That can be genuinely useful, especially when I am stuck or looking at a problem from only one angle. The suggestions are not always right, though, so they still need to be checked.
AI does not bring emotions, thoughts, or an understanding of the business into the conversation. It can suggest a solution, but it does not understand why the problem matters to the people who are dealing with it.
Yes, a conversation with another person can take longer, but that does not mean the time was “wasted”. The person you are talking with may challenge an assumption, remember something from a previous situation, or explain why an apparently good solution will not work in the real world.
Children and education
The “AI” of my generation was social media, and good things came out of it, but so did many bad things: spreading intimate images, putting information online that people did not want exposed, and normalizing an almost permanent breach of privacy.
When I went to school, we did not stare at screens all day. We used pen and paper. By the time I went to university, devices had become much larger part of every day learning, and I could feel the difference. I felt that I absorbed more when reading slowly, wrote things down and focused on one thing at a time. At university, a large part of learning became filtering away the noise introduced by the massive amounts of digital information available to me.
As a parent, I wonder what happens to a child’s curiosity when a machine can summarize and explain every answer. What will my role be when I am no longer the person with all the answers, or even the person who can provide the most nuanced explanation? That makes my role more important, not less. We always need to question an answer, look for its source, and decide whether it makes sense instead of accepting the first provided response.
Looking at my own son, I see him having more fun with a large spoon than with some of the toys we bought him. He does not need a sophisticated toy to have a good time or learn something new. Sometimes the simple thing is enough, and I worry that we will add technology before we have given children enough time to discover things on their own.
The generative models are now good enough to produce convincing images and videos of things that never happened. With tools like that becoming easily accessible, what will people generate? Unfortunately, I think history will repeat itself - women will be exploited in ways we cannot yet imagine. That makes me so damn sad.
Images and videos have historically been something we trust. Going forward, we are going to need better ways to establish what is real and what is not. Technology alone will not solve the social part of the problem. People still need to pause, ask who made something, and decide whether it deserves to be believed.
Critical thinking should be a subject on its own in schools.
The people selling the future
Sam Altman, Dario Amodei, Jensen Huang, and Elon Musk are some of the public faces of AI today.
Lately, some of these leaders have been talking about existential risks from advanced AI and the need to slow down the progress of frontier models. I find myself asking whether that concern is sincere, or whether it also creates a reason to raise more money and continue a massive build-out of data centers.
I get the icks, and it is difficult to trust warnings about danger when the same people warning us are also asking us to buy more access to the thing they are warning us about.
I often see people compare the AI boom to the transition from horses to cars, but, I think the analogy misses its point. If the faces of AI are to be believed, this technology has the potential to disrupt far more than a single industry like transportation. Its impact could reach across huge parts of the economy, and many people are likely to fall through the cracks. Valid questions that people have about AI and the future have low chances of being addressed. Their strategy seems to be that every criticism can be dismissed as fear, while every promise is inevitable to happen.
AI as a tool
An axe can build something useful, chop wood, or cut off your limbs. When you wield a tool, you direct its action and apply your own skills, judgment, and effort to the work. AI is different: you give it a request and it does the work on your behalf.
People often describe AI as just another tool and I am not sure that is the right conclusion. The risk, as I see it, is that we adapt people to technology instead of adapting technology to people. We may start designing our work around what the model can produce, rather than around what helps us do the work we care about.
I am not arguing against AI. What I am arguing against is the idea that we should accept it as inevitable, pretend it is free, ignore the harm, and put the technology ahead of the people who have to live with it.
Maybe I am wrong, and I hope I am wrong about some of this. But I would rather ask uncomfortable questions now rather than later.
← Go back