Your MEH has a good track record
When generative AI took over every media outlet, my first thought was MEH. Blockchain? MEH. Electric cars? MEH. Another super duper product that everyone insists I need? MEH again.
That reflex is not laziness and it is not fear of technology. It is pattern recognition, and it has been mostly right. Chasing every new AI tool costs more than it returns, because you migrate from framework to framework, each migration eats a weekend, and the improvement is so small you cannot even measure it. Two weeks of careful prompt engineering can become worthless after a single model update. What does not expire is the problem underneath the tool, things like context efficiency, agent authorization, auditability. Track the landscape so you understand where it is going, and adopt something only when it proves itself over months, not days.
So keep the MEH. Just do not confuse it with a conclusion.
Because there is a real difference between yet another product aimed at one narrow slice of your life and a generative technology that touches many slices at once. The limit is your knowledge, your experience, and your courage.
Every invention arrives with an invoice
Gutenberg’s press in 1439 gave humanity something it had never had, which was knowledge that anyone could reach. Before that, books lived in the hands of political leaders and heads of religious institutions. Literacy rose, scientific exchange became possible, and entire trades appeared around printing and publishing.
It also gutted the livelihood of the people who copied manuscripts by hand, and handed religious and political groups the first industrial-scale machine for misinformation.
Watt’s steam engine in 1769 did the same thing in a different key. Efficiency rose, goods got cheaper, demand grew, and skilled work appeared around building and maintaining the machines. In parallel, repetitive manual jobs disappeared fast enough that the Luddites organized against it, factory floors became genuinely dangerous because nobody had invented worker protection yet, and small craft workshops went bankrupt because machines simply did not need to sleep.
Nobel stabilized nitroglycerin into dynamite in 1867 and gave miners a way to move rock without dying of exhaustion in a tunnel. He was, by his own description, a pacifist. The military application arrived anyway, and if he had not stabilized nitroglycerin, someone else would have, or something worse. On 16 July 1945 someone did, and Oppenheimer’s work gave us both the bomb and a nearly limitless source of electricity.
The pattern is clean and it is genuinely comforting. Every invention comes with a reward and a matching price. Which is precisely why almost every article about generative AI ends the same way, with the reassuring line that the printing press, the steam engine and dynamite were all just tools, neither good nor evil, and that everything depends on how we choose to use them.
I used to end my own thinking there too. Then I noticed what that sentence quietly assumes.
Where the analogy breaks
A tool, in the sense that comforting line means it, does exactly what it is built to do and nothing else. Dynamite has no model of the room it is in. A steam engine cannot be mistaken about anything, because it has nothing to be mistaken with. The entire moral weight sits with the person holding the thing, which is why „it depends on us” works so well as a closing sentence for those three inventions.
Generative AI does not fit that description, and pretending otherwise is the one place where historical comfort turns into a bad forecast.
Start with the part that has no precedent at all. Throughout human history, our ignorance set the hard ceiling on technology. Nobody could build a transistor before someone understood electricity. Machine learning broke that ceiling for the first time, because we can now systematically build things that work beautifully while nobody fully understands the mechanism. Demis Hassabis won a Nobel Prize in chemistry for AlphaFold and openly admits he does not know exactly how his own program predicts protein structures. Our ability to understand this technology is growing more slowly than the technology itself, so any trust you place in the next, more powerful system cannot rest on the assumption that somebody, somewhere, has the full picture.
Now the part that should genuinely unsettle you. Anthropic published an analysis of three incidents from its own cybersecurity evaluations, where a Claude model doing a capture the flag exercise ran into real systems on the network. The model was not pursuing a goal of its own. It did precisely what it had been asked to do. The problem was situational awareness, because the prompt had assured it there was no internet access, so it assumed the real machines were part of the fiction. In one case the model verbally recognized that the system was real, rationalized it as part of the scenario anyway, and continued. Only the newest of the three models stopped once it established the target was real, and three incidents are far too few to call that a trend.
Read that again with dynamite in mind. Whether the action was aligned or harmful was decided by what the system believed about its own environment, not by anyone’s intention. No earlier tool in this article had a belief that could be wrong.
There is a third break, and it is structural rather than dramatic. Dario Amodei points out that AI concentrates power through scaling laws themselves, because model quality rises with capital, data and compute, so the mechanism rewards whoever already has the largest scale. That happens whether or not anyone regulates the market. It is not a side effect of a political decision, so looking for the fix purely in regulation or purely in deregulation means searching where the answer is not. Open weights do not dissolve it either, they shift the concentration toward whoever owns the compute.
The bill, in water and watts
Wars used to be fought over square kilometers, then over megajoules of energy from oil. The currency of this race is teraflops. In September 2023 Microsoft announced it was recruiting nuclear physicists to build private atomic power plants feeding AI data centers directly, which tells you that the pace of this technology will increasingly be set by physical energy limits rather than by algorithmic ideas.
The water bill is just as literal. Microsoft’s vice chair Brad Smith confirmed that a supercomputing data center in Iowa, built specifically to train GPT-4, consumed close to 6.5 billion liters of water. Nobody actually knows the full number across the industry.
Then there is the bill nobody puts on a slide. 88% of companies use AI in at least one business function, and 95% of them report no measurable return on that investment. Gartner estimates that more than 40% of agentic projects will be cancelled by 2027. This is the productivity paradox doing what it always does, which is generating organizational friction long before it generates speed. Treat every transformational claim as a hypothesis to verify, and ask for hard adoption metrics like how many people actually reached proficiency and how many pilots reached production, rather than for stated intentions and budget size.
Your own competence pays a quieter version of the same invoice. Polish endoscopy centers tested an AI tool for detecting polyps and measured the doctors themselves three months before the tool and three months after it was taken away. Their detection rate afterwards fell below where it had been before they ever used it. We grow by struggling with difficulty, so when you hire AI at exactly the point where the difficulty appears, the task gets done and you do not develop. Anthropic’s own internal data says developers use AI in roughly 60% of their coding work but report being able to fully delegate a task, without checking the result, in only 0 to 20% of cases. Those are two completely different statements, and the gap between them is your judgment.
Adaptation is not capitulation
Here is the part I want you to take with you, because fear of adaptation really is the dangerous piece. It pushes you out of the comfort zone, and most of us would rather not go.
Start by naming the fear out loud instead of swallowing it. Organizational psychologists are clear that worrying about AI replacing your work is a legitimate reaction, not a character flaw, because the technology genuinely challenges your sense of competence and your psychological safety at work. Suppressed fear does not disappear, it turns into cynicism and quiet withdrawal from the new tools, which makes everything worse than saying it plainly to your team or your manager.
Then protect the muscle. Give yourself an hour at the start of the day with no AI at all, thinking on paper, sketching architecture by hand, working the problem the old way even when it feels inefficient. It is the same mechanism as satellite navigation, where years of following the arrow quietly erase your mental map of your own city. A sharper mind in the morning also reviews the afternoon’s generated output better, so the hour pays for itself twice.
And make the mental switch that actually matters, which is moving from asking AI a question to delegating a task and judging the result. That switch is what changes your role, and it is also where the real cost hides. When you cannot decide something trivial at the end of a heavy day, like what to name a function, that is not fatigue from writing code. That is decision fatigue from spending the day as a reviewer instead of an author.
The transition period will be painful for some people, exactly as it was for the manuscript copyists and the handloom weavers. That part of history does repeat.
What does not repeat is the shape of the responsibility. With dynamite, the whole moral weight sat with the hand holding the fuse, so „it depends on us” was a complete answer. With a system that forms beliefs about its own situation, that nobody fully understands, and whose economics concentrate power by default, the responsibility does not get smaller. It gets larger and considerably more specific. It means deciding what you delegate, verifying what comes back, and staying awake to who profits from telling you to delegate more.
The future is not something that happens to you. It is something you build, one decision at a time, and this time the decisions are harder than picking a side.