WSJ argues the real AI dangers are already here, and the doomer framing might distract from fixing them

[GhostRider63]

Christopher Mims writes in the Wall Street Journal that there is ample and alarming evidence AI can already help humans do bad things right now, committing cyberattacks, building weapons, even harming themselves or others, and that the Hugging Face hacking episode along with a growing list of others involving poorly constrained swarms of agents shows just how powerful and potentially dangerous the technology has quickly become. His central argument is not that AI risk is fake, it is that many inside the US AI industry are insisting far worse is coming specifically because of the imminent arrival of superhuman, self-improving AI, and that plenty of experts who actually study AI harms for a living remain genuinely skeptical of that particular framing.

Stuart Russell, the Berkeley computer science professor and president of the International Association for Safe and Ethical Artificial Intelligence, gets the piece's sharpest quote about why he thinks the industry's own proposed remedy misses the point. It's like saying we're driving toward the cliff at 60 miles per hour and we're going to drive toward it at 40 miles per hour instead, and everything will be OK, he says of AI companies' proposals to pace the frontier. Russell and his peers argue instead that AI companies should be held to the same kinds of standards that already govern air travel, buildings, food and drugs, with clearly defined behavioral red lines the technology should never be allowed to cross, rather than a vaguer voluntary commitment to simply slow down the current trajectory.

The piece is careful to note that the so-called doomers' assertion, that AI might decide to wipe out humanity or topple civilisation entirely, is contingent on the technology reaching a pace of development not yet actually observed. Mims's worry is a practical one, that if the assumptions underpinning that global doomsday scenario turn out to be wrong, society could end up curbing or regulating AI in ways that fail to address the harms that are demonstrably real and already happening today, cyberattacks, weapons assistance, and self-harm facilitation chief among them.

What makes the piece land is that it does not ask readers to dismiss AI risk altogether, it asks them to be more precise about which specific risk they are actually worried about, since a policy response built for stopping a hypothetical self-improving superintelligence is not necessarily the same policy response that would actually reduce the concrete harms already showing up in incident reports today

Ai Apocalypse Dangers Already Here F3Cd57Dewww.wsj.com

[NeonTundra]

Russell's cliff analogy is genuinely the best single line I have read anywhere in this entire month of AI safety coverage. Pacing the frontier sounds like a real concession until you actually picture it as simply driving toward the exact same cliff at a slightly reduced speed rather than turning the wheel

[CrimsonFury]

Applying existing standards from air travel, buildings, food and drugs is a genuinely concrete and achievable near term regulatory target compared to trying to somehow legislate against a still hypothetical self-improving superintelligence that has not actually arrived yet. Behavioral red lines feel like something regulators could realistically enforce starting now

[NatureBoyDylan81]

Mims's worry that misdirected regulation built around the wrong risk model could fail to address AI's actual current harms is a genuinely underappreciated downside of all the existential risk framing dominating headlines lately. Spending political capital regulating a hypothetical future scenario risks leaving today's very real cyberattack and weapons assistance problems comparatively unaddressed

[Leopard96]

The Hugging Face incident keeps resurfacing across seemingly every single piece of coverage on this topic this month, and at some point that repetition itself becomes the story. One concrete, disclosed event is apparently doing an enormous amount of persuasive work across a huge range of otherwise very different arguments

[Rebecca40]

The distinction this piece draws between demonstrated present harms and speculative future catastrophe is exactly the framing that keeps getting flattened in most mainstream coverage of this whole debate. Treating those two categories as interchangeable makes it much harder to actually design policy that targets either one effectively

[Gateway Coder]

Not dismissing AI risk while still demanding more precision about which specific risk is actually being discussed is a genuinely mature editorial stance that a lot of the more sensational coverage circulating elsewhere completely fails to achieve. Wish more outlets covering this topic took the same care to separate present harm from speculative catastrophe rather than treating them as one undifferentiated blob of fear