On productivity, trust, and the physical limits behind the AI boom
Writing used to be thinking. Sitting in front of a blank sheet of paper (or later a screen) and waiting for something to happen (or, in my case, to not happen). For a long, long time, that world has existed (not the screen bit). Now it is starting to disappear.
The AI engine is a powerful tool, and it is tempting to believe that it can perform perfectly. But it is easy to get fooled when a series of half-baked thoughts are churned out in a matter of seconds, especially when a little tinkering with the prompt can yield results that look impressively competent. People have been enjoying their PC’s recent increase in productivity. The emails that used to take half an hour to compose are now done in five minutes, their complaints are dealt with efficiently, and proposals are written to a I’m off to the coffee shop. But the PC performs too well, and that makes me wonder whether any real thinking is happening at all.
Artificial Intelligence is producing an avalanche of digital junk. In written form, this takes many forms: articles, blog posts, comments, reports. All correctly worded, professionally crafted, yet lacking in content. You could say that the internet has figured out how to talk flawlessly yet ended up saying nothing of value. It’s like people who have drunk too much, amiable enough, but not really discussing anything of real substance.
But beneath this rapid digital progress lies a mounting crisis. Data centre capacity is growing so fast that it is making utilities uncomfortable. Rather than gently increasing in power consumption, now it is racing ahead. Even the build strategies for new data centre capacity are predicated on the need to support AI loads, for which, just a couple of years ago, mindlessly oversizing servers for big computational jobs for which someone was willing to pay would have been considered excessively expensive.
Are we getting the value we had hoped for from these vast infrastructures? I’d wager that a considerable sum of money has been wasted on machine learning, deep learning and related infrastructure. People keep telling me that AI is currently delivering sufficient value, but that can’t continue to be the case. Sure, there are some patches of innovation, but generally, these assumptions ‘age badly’. Remember “build it, and they will come” mentality of about 20 years ago?
While this is an impressive piece of work, there are still some cracks in the system, and they are not technical quirks but economic ones. The cost of being wrong at scale is increasing, not decreasing. When flawed outputs are produced faster and with greater confidence, they are more likely to be trusted, reused, and embedded into decision-making. What looks like efficiency on the surface can quietly translate into wasted time, misdirected effort, and capital being deployed against assumptions that were never properly tested.
This is where the issue of trust becomes more than theoretical. If the output is fast, fluent, and consistently acceptable, it becomes very easy to stop questioning it. Multiply that across teams, departments, and organisations, and you begin to see the economic problem: a growing volume of work that looks finished but carries little depth or originality. At scale, that is not just a quality issue, it is a cost issue.
It actually works really well for what it does. AI can be effective at identifying priority tasks, generating an outline, or helping to quickly produce content. In the best scenarios, it could even cut down the amount of time spent on repetitive, time-consuming aspects of the work, allowing you to think more and produce faster.
Despite all the recent attention on AI Agents, their emergence does not fundamentally change the equation. If anything, it accelerates it. More autonomy means more output, more decisions, and more reliance on systems that still struggle with accuracy and context. The cost base expands, the speed increases, but the underlying question remains unchanged: is the value keeping pace?
We should be cautious of jobs. Eventually, we won’t need to be. Already, many ‘safe’ jobs are under threat, not just those requiring repetition or simple manual labour, but also those which require skilled thinking. Then there’s the rise of robotics: whilst some positions may be more at risk than others, it’s foolish to believe that only people in low-skilled jobs are going to be let down.
This is a subject nobody really wants to address thoroughly because, in the end, it always comes down to the same question: when machines are capable of ‘thinking’, what will be left for us poor humans to do?
What’s the optimal state of the economy? Is it a Star Trek-style model where work is optional because everyone is too rich to bother, contributing to the economy because they want to, not because they must? And if so, would we implement a universal basic income to fill the gap? Sounds lovely and assumes that governments and societies are far more competent and coordinated than history suggests they are.
It is more likely that the future becomes more complex and opaquer, increasingly dominated by those who control the essential infrastructure and who are bankrolled by a handful of large companies and governments. The true costs of the future will be hidden and unfairly distributed.
OK, so is AI going to collapse, or is it actually happening? Everyone is guessing. My view is it is a bit of both. It produces obvious benefits (like improved productivity), and it introduces real risks (like increased speed, scale, and unexpected assumptions). AI is an incredible tool for certain problems, but it is also a reflection of how we have built systems, enabled incentives and turned a blind eye to problems that are difficult to address with current technology.
For now, it shines a light on what we can do, but it will also reflect back what we have missed. It is brilliant, right up until it isn’t.
By Peter Zanatta
