Every few years a new technology arrives wrapped in the language of inevitability. We are told it will change everything, that the old world is finished, and that anyone who hesitates will be left behind. Right now, that technology is AI.
We have seen this pattern before. A few years ago, it was blockchain. Before that it was any number of other “revolutions” that were supposedly going to redraw the world overnight. Some of them mattered. Some found a useful niche. Some never became what the hype promised. So, we should be cautious when we hear the same tone again.
That does not mean AI is trivial. Far from it. It is already changing how people work, write, code, search, and make decisions. It will almost certainly become deeply embedded in business and daily life.
The concern is not whether AI is powerful. It clearly is. The concern is whether we are rushing into it without asking hard enough questions about power, dependency, cost, capability, and long-term consequences.
Maybe humanity will adapt, as it usually does. Maybe many of today’s fears will prove exaggerated. But that is not a reason to suspend judgement. If anything, it is a reason to think more clearly while there is still time to shape how this technology is used.
These are the risks we should keep coming back to:
One of the things that should trouble us most is the possibility that, in the name of efficiency, we are concentrating extraordinary power in the hands of a very small number of AI companies.
We spend years talking about competition law, monopolies, anti-competitive behaviour and the need to protect markets from becoming too concentrated. Yet at the same time, we are happily building business processes, workflows and decision-making around platforms controlled by a handful of firms whose scale is becoming extraordinary.
OpenAI was reported at a valuation of about USD 852 billion in March 2026, while Anthropic was reported at USD 965 billion in May 2026. Those are not normal numbers. They are signals of where power may be heading.
If we remove humans from more parts of the process and hand those tasks to AI, we are not eliminating power. We are relocating it. The question is whether we are relocating too much of it to companies that will eventually have enormous influence over cost, access, pricing, standards and dependency.
At some point, this stops being just a commercial story and starts to look like a political one. If that dependency deepens, are we witnessing the early stages of a real shift in power from governments to technology companies?
We should also worry that the convenience of AI will slowly chip away at our own capability. It is already becoming second nature to reach for ChatGPT, Claude or another model the moment a question appears. The attraction is obvious: it is fast, easy and often good enough. But convenience changes behaviour.
When a tool starts doing too much of the thinking for us, there is a risk that we stop exercising the very muscles that made us capable in the first place: reading deeply, wrestling with ambiguity, remembering details, testing assumptions, and working things through for ourselves. This is not intended as a dramatic end-of-humanity sense. It is in a much more ordinary and believable way - we get out of practice.
That may not matter for the occasional shortcut. It matters a lot if it becomes a habit, especially for younger people who are still learning how to think, write and solve problems without a machine filling in the gaps for them. A society that outsources too much of its cognition may not become smarter. It may just become more dependent.
One of the clearest near-term temptations is replacing junior staff with AI. For accountancy firms, law firms, consultancies and many other professional services businesses, it looks like an easy win. Low-value tasks can be done faster, more cheaply and at greater scale.
The problem is that junior roles are not just a cost line. They are the training ground for the future. Today’s junior analysts, associates and trainees become tomorrow’s experienced operators, managers and partners. If businesses hollow out the early career layer too aggressively, they may save money now but create a succession crisis later.
What makes this more dangerous is that the people making the decision may not be the people who live with the consequences. In partnership models especially, the financial upside is immediate, while the organisational damage may show up years later. That is exactly how short-term logic creates long-term fragility.
A lot of people still talk about AI as if it will simply make everything cheaper, but how real is this? In many cases, we are still in the honeymoon period, where the capability is obvious, but the long-term cost model is not.
As models become more agentic, they also become more resource intensive. Longer context windows, repeated reasoning steps, file access, tool use and multi-stage workflows all increase compute demand. That may be manageable at small scale. At enterprise scale, it becomes a very different commercial conversation.
In other words, AI may reduce some labour costs while quietly introducing a new category of infrastructure and usage costs that many organisations still do not fully understand. The headline savings may be real, but so is the risk of replacing visible human cost with less visible platform dependency and compute spend.
Another concern is how quickly the pricing model can change once people are dependent on the product. GitHub Copilot (part of the Microsoft developer ecosystem), for example, announced that from June 2026 it would move from premium-request billing to usage-based billing through AI credits calculated from token consumption. Base plan prices stayed the same, but the charging logic changed in a way that makes heavy or agentic use potentially much more expensive.
It is not being suggested that there is anything improper about that. Companies are entitled to charge for what they provide. But it is a useful reminder that when an AI tool becomes embedded in your workflow, the supplier gains enormous leverage over how the economics evolve. What looks affordable during adoption can look very different once the habit is established and switching becomes painful.
These agentic models are also shifting the economics in plain sight. They are no longer just responding to the task in front of them; they increasingly expand the scope by suggesting further actions, refinements and next steps. Useful as that may be, it also increases token consumption over time.
That is one more reason to be careful about assuming AI will automatically drive down costs. In some cases, it may. In others, it may simply move spending from payroll to platforms.
Any business case for AI needs to be tested across different scenarios to make sure the value really holds up.
The environmental dimension should also be a concern, partly because it receives far less honest discussion than it should. AI enthusiasm is often framed as if compute is somehow abstract. It is not. It sits on top of data centres, chips, cooling systems, transmission networks and huge amounts of electricity.
Jensen Huang recently argued that the energy needed for future computing could be “a thousand times more” than what we have today and even suggested that estimate could prove low. Whether or not that exact number is right, the broader point is hard to ignore. Advanced AI implies a huge increase in compute demand, and compute demand means energy demand.
The International Energy Agency (IEA) expects global electricity demand to continue rising strongly over the coming years, driven by electrification, industry, air conditioning and data centres.
The IEA also put 2025 global electricity demand at 31,779 TWh, up 849 TWh year on year, with China accounting for 59% of the increase. Those are big numbers before AI demand has fully matured. Put differently, that annual increase alone would rank as the world’s sixth-largest electricity consumer if it were a country.
So, the question is not whether AI is clever. It is whether we are being realistic about the infrastructure needed to sustain it, and whether that infrastructure will be powered cleanly enough to avoid simply outsourcing the environmental cost into the background.
Whatever your view of Net Zero, governments are struggling to find the resources to address what they are currently committed to, let alone multiple times more consumption taking place over the coming years. There has to be a limit on where we can place wind turbines or plant trees.
##7. The economic story may be too neat
The economic narrative around AI is also starting to sound a little too tidy. Sam Altman has spoken about the extraordinary economic upside AI could create, while also recently saying he no longer expects the “jobs apocalypse” he once feared.
He said he had been wrong to expect more immediate elimination of entry-level white-collar jobs, and that the human element of work still matters more than he thought.
That may turn out to be right. We should hope it is. But we should not accept the optimistic version too quickly either. Even if AI does create enormous value in aggregate, that does not mean the gains will be evenly shared. Nor does it mean the transition will be painless for the people, sectors and regions that absorb the disruption first.
This is often the problem with grand economic predictions. They can be true at the macro level and still deeply uncomfortable at the human level. A bigger GDP number does not automatically answer who loses leverage, who loses opportunity, who becomes dependent, or who is priced out along the way.
If AI really does drive a large-scale shift in white-collar employment, it raises a basic economic question of who is going to drive the consumption for significant economic upside. If a large share of affluent white-collar workers see their incomes squeezed, where exactly is the demand meant to come from?
My instinct is that we are still in the phase where the upside is marketed loudly and the trade-offs are spoken about quietly. That does not make AI bad. It just means the conversation is incomplete.
This article is not anti-AI. AI has value and is becoming part of our every day lives. It’s easy to see why people are excited by it. In some areas it is genuinely impressive, and in others it will be transformative. The unease comes from the speed at which we are normalising dependence on systems we do not fully understand, controlled by companies we may eventually have little choice but to rely on.
Maybe this will all settle into something useful and balanced. Maybe the fears will soften and the benefits will prove worth it. We can only hope so. But, we should be more willing to ask awkward questions now, before dependency hardens into inevitability.
The real issue is not whether AI will change the world. It already is. The real issue is whether we are paying enough attention to what we may be handing over in return.