The perils of creating a world more complex than we need
India, Nov. 20 -- In recent weeks, several everyday systems have broken. On October 11, hackers took control of the fridges in more than 700 French supermarkets and turned up the heat, then did the same across Europe. On November 12, Microsoft's email and Teams went down worldwide. The next day, the US Federal Aviation Administration stopped every flight in the country from taking off, only the second time since 9/11. On November 16, a software failure delayed United's flights, and on November 17, an internal update at Apple knocked out iCloud, Apple Pay and the App Store for hours.
These may be unrelated, although emergency patching of vulnerabilities detected by new AI models may be a factor. But together they point to one massive problem: We are building systems faster than we can understand them, run them or fix them when they break. Almost everything important now runs on dozens of systems, built by different companies over decades, and nobody understands the whole. Microsoft's own early report says a software bug, extra traffic after a network change and fewer working servers combined to bring it down; any one event alone would probably have been manageable. When such a system fails, it takes hours just to work out why, because the knowledge needed to fix it is spread across many organisations. The sociologist Charles Perrow warned in Normal Accidents (1984) that tightly linked systems whose parts interact in unexpected ways will produce serious accidents sooner or later. He was writing about nuclear plants, not supermarkets.
Then there is a question we don't ask: Why are supermarket fridges on the internet at all? Putting it online saves the supplier a service visit but gives every store a door that hackers can walk through. The same idiocy has reached our homes. When Amazon's cloud went down in 2025, owners of smart beds found them overheating or stuck upright. A bed that needs a data centre in Virginia to lie flat is not progress. We connect things because we can, because it earns companies data and subscription fees, and because nobody counts the cost of this dependency.
Two things make every failure worse. One is concentration. India knows this well: UPI runs a huge share of the country's everyday payments through a single operator, and in spring 2025 it failed four times in under three weeks, with another nationwide outage this February. Some failures started at the National Payments Corporation of India (NPCI), others at the banks behind it. When UPI stops, the vegetable seller, the auto driver and the large retailer all stop together.
The other is the loss of people who understand these systems. Years of running lean, layoffs, outsourcing and early retirement have cut costs. But the engineer who knows why a system was built the way it was is exactly who you need when it breaks, and that knowledge is rarely written down. Days before its outage, the FAA announced it would finally stop using floppy disks, still in use at more than 400 of its sites, by 2028. Short of air traffic controllers for years, it also cut its full-staffing target in May from 14,633 to 12,563, citing efficiency and automation. The systems grow more complicated while the number of people who can fix them shrinks.
As firms deploy more AI, the risks only grow. AI agents, software that books, buys and pays on its own, will make every problem above worse. Every agent sits on top of the systems that failed this month, so we grow more dependent on infrastructure we already struggle to keep running. Moreover, most AI runs on a handful of AI models and cloud providers, increasing vulnerability. As agents take over work, the people who knew how to run and repair these systems move on, and those who remain understand less of what the system is doing. We will be adding software that makes decisions few people can trace, on top of systems few people understand, run by fewer people who could fix them.
So what can we do? Design every service that matters to keep running, even if poorly, when its computers fail, and have staff who have practised for such situations. Anything that doesn't need the internet, from a fridge to a water pump, ideally shouldn't be connected or at least must be able to work without the internet. Boards must know which few vendors are mission critical and have contingency plans. Before any round of layoffs, management must be able to name who would fix the system at three in the morning, and retain them. And no agent should be switched on until someone has ensured that it stops and hands over to a person in the event of a glitch.
None of these problems comes from carelessness. It is a predictable consequence of many apparently sensible choices: A world more complex than we need, and more complex than we can understand, run, or repair. AI is about to add the next layer, running faster and more autonomously. Before we do, we should ask of every new system what happens on the day it stops working, and have an answer for that....
इस लेख के रीप्रिंट को खरीदने या इस प्रकाशन का पूरा फ़ीड प्राप्त करने के लिए, कृपया
हमे संपर्क करें.