The critique, in essence 🎯
An op-ed published this week by MIT Technology Review takes the industry to task for inviting us to worry about an imaginary all-powerful machine rather than about what is already happening: worsening climate, asthma among populations living near data centres, rising electricity bills, diverted water. It lists the summer's spectacular announcements — vulnerabilities, incidents, mathematical breakthroughs — as so many diversions.
An op-ed published this week by MIT Technology Review takes the industry to task for inviting us to worry about an imaginary all-powerful machine rather than about what is already happening: worsening climate, asthma among populations living near data centres, rising electricity bills, diverted water. It lists the summer's spectacular announcements — vulnerabilities, incidents, mathematical breakthroughs — as so many diversions.
This critique deserves better than a defensive response, all the more so because it applies in part to the way these subjects are covered everywhere, including here. Let us take it seriously.
What the argument gets right
Let us start with what must be conceded, and it is a great deal.
The concrete harms are documented. We have ourselves devoted articles to water in India, to the anger of residents in Virginia, to the cost of grid connections passed on to households, to the working conditions of annotators. These effects are real, measurable and present.
Attention is a scarce good. Every hour spent debating a hypothetical superintelligence is an hour not spent on a real electricity bill. That is not a detail: it is the resource of public debate.
Spectacular narratives serve interests. A technology presented as powerful enough to threaten humanity is also a technology presented as powerful enough to justify any valuation. The narrative of fear and the narrative of promise feed each other. We noted this in our essay on normalisation: attention concentrates on what is spectacular, never on what is diffuse.
What does not hold up
We must also say where the argument overstates its case, because that is where it becomes dangerous.
The two categories of risk are not mutually exclusive. Presenting concern about capabilities as a diversion assumes that a choice must be made. Yet nothing prevents us from making data centres pay for their grid connections and requiring labs to properly isolate their tests. These are two problems, handled by different people, with different tools.
Some "spectacular" risks have materialised. Three labs stated within two months that a model had reached real systems, the latest last week. A campaign used an open model to attack 460 systems. These are not speculations about a divine machine: they are security incidents, with real victims. Filing them under fantasy would also be a misallocation.
The distinction that saves the debate ⚖️
The problem is not worrying about capability risks. It is conflating two things that share the same name. On one side, documented risks: incidents, attacks, containment failures, which can be counted and corrected. On the other, speculative scenarios: a hostile superintelligence, whose probability cannot be measured nor its assumptions tested. The criticism lands on the latter and misses the former. Lumping them together — in either direction — is exactly what prevents thought.
The problem is not worrying about capability risks. It is conflating two things that share the same name. On one side, documented risks: incidents, attacks, containment failures, which can be counted and corrected. On the other, speculative scenarios: a hostile superintelligence, whose probability cannot be measured nor its assumptions tested. The criticism lands on the latter and misses the former. Lumping them together — in either direction — is exactly what prevents thought.
How to distribute your worry
If attention is scarce, a criterion is needed to allocate it. We propose three, and they hold for all risks.
Is it already happening? A present harm takes priority over a possible one, not because it is more serious, but because it is certain and can be acted on now.
Who suffers it, and can they make themselves heard? The residents near a data centre have elected representatives. The annotators in Kenya have none who count in this debate. Attention should go first to those who have no other means of getting anyone to look at them.
Can anything specific be done? A worry that leads to no identifiable action is an emotion, not a policy. Charging for a grid connection, requiring physical isolation, protecting a whistleblower report: these are actions. "Slow down" without a criterion is not yet one.
What to take away
The answer to the question in the title is yes: we should worry about the right things, and this op-ed is right to remind us that the right things are not always the most spectacular.
But it would be wrong to conclude that we must choose between visible harms and technical risks. The real divide is not between the concrete and the spectacular. It is between what can be measured and corrected, and what can only be imagined.
Electricity bills, water, working conditions, containment incidents all belong to the first category. That is where most of our attention should go — not because the rest does not exist, but because that is where it can change something.