The Smokescreen of AI Doomsday

Timnit Gebru makes a sharp point in WIRED (link below) about the strange way artificial intelligence is discussed. The more ordinary the technology becomes, the more extraordinary the language surrounding it seems to get. We are no longer merely talking about software that writes, codes, recognises images and increasingly makes decisions. We are talking about machine gods, extinction, the end of humanity and an intelligence that may one day escape the control of its creators.

Gebru suspects that this is not entirely accidental. The apocalyptic story directs attention towards a hypothetical future that nobody can inspect while some very real problems are sitting under our noses. Autonomous weapons are already being developed and deployed. Data centres are placing enormous demands on electricity grids. Companies can dress ordinary cost-cutting exercises in the glamorous language of "AI efficiency." Systems known to hallucinate and make mistakes are being pushed into increasingly consequential areas of human life.

There is an obvious political advantage in talking about the machine rather than the people who own it. If a bridge collapses, nobody asks whether the steel became malicious. We ask who designed the bridge, who approved it, whether somebody cut corners and who is legally responsible. Yet discussions of artificial intelligence increasingly attribute something resembling agency to the technology itself. The machine "decided," the model "wanted," the AI "deceived." Human decisions disappear behind an increasingly anthropomorphic vocabulary.

The extinction story also does something rather useful for the companies building the technology. A business selling unreliable software is merely a business. A business creating an intelligence that might transform civilisation or destroy humanity suddenly belongs in the room with presidents, defence chiefs and regulators.

There is an almost comical grandeur to it. The same industry can tell investors that it is building the most valuable technology in human history and tell governments that the technology is so dangerous that only the people already building it can explain how it should be controlled. Destiny attracts capital; doom attracts political importance.

That does not prove that the warnings are insincere. Human beings are perfectly capable of believing something that also happens to advance their interests. An AI executive can genuinely worry about catastrophic artificial intelligence while simultaneously benefiting from a public narrative that makes his company appear to be the Manhattan Project with stock options.

This is where Gebru's argument needs qualification. The choice is not necessarily between imaginary future dangers and real present ones. Nor does public attention operate like a fixed quantity in which every newspaper story about superintelligence means one fewer story about autonomous weapons, employment or electricity consumption. Before the spectacular warnings about AI began, most people were not carefully studying algorithmic discrimination or the liability problems created by autonomous systems. They were paying little attention to artificial intelligence at all.

Apocalyptic stories have always had one great advantage: people notice them. The fellow who clicks on a story asking whether AI will destroy humanity may eventually find himself asking a much more useful question about who is responsible when an automated military system makes a disastrous mistake. The science-fiction story can be the doorway into the engineering problem.

There is therefore a danger in treating the entire existential-risk debate as a confidence trick. Some of the scenarios being discussed are highly speculative, but the underlying trajectory is not. More capable systems are being given access to more consequential parts of society. They can already assist fraud, cyber operations, surveillance and weapons development. None of this requires consciousness, evil intentions or a computer plotting the extermination of Homo sapiens. A sufficiently powerful tool does not have to hate you to be dangerous.

That distinction gets lost whenever the argument divides into two camps. On one side are the AI doomers contemplating extinction; on the other are critics such as Gebru demanding attention to existing harms. But these are different questions rather than mutually exclusive answers.

The immediate problems are easier to investigate. Does an AI system work reliably? Who bears liability when it fails? What happens to an electricity network when data-centre demand expands faster than generating capacity? What military decisions should never be delegated to software? What happens when corporations use automation to eliminate workers while transferring the risks of faulty systems to everybody else? Those questions can be examined with evidence, engineering, economics and law.

The distant scenarios are necessarily more uncertain. Nobody can presently demonstrate that a future autonomous system will acquire hidden goals and attempt to destroy humanity. Such possibilities can be considered as tail risks without pretending that speculation has the evidentiary status of an existing engineering failure.

There is also a cost to crying apocalypse too often. If every new model is another step towards the machine god, and every year brings another warning that humanity has five years left, eventually the warnings become background noise. The boy who cried wolf problem does not disappear merely because wolves exist.

Gebru is therefore right to drag the discussion back towards human beings. Somebody builds these systems. Somebody decides where they are deployed. Somebody profits from them, and somebody should remain responsible when they fail. "The AI did it" must never become the technological equivalent of an act of God. But the ghost story cannot simply be wished away. Ironically, it may be one reason the public is paying attention in the first place.

The sensible response is not to choose between the ghost and the machine. It is to keep looking behind both of them at the engineers, executives, investors, soldiers and governments deciding what the technology will actually be permitted to do. The future machine god may never arrive. The institutions building it are already here.

https://www.wired.com/story/one-of-ais-fiercest-critics-says-all-the-doom-talk-is-meant-to-distract-us/