It is being deployed faster than anyone can judge whether it is safe

The argument

A second argument concerns pace. Opponents observe that the ordinary way societies establish whether something is safe is slow: trials, inspection, published evidence, adversarial review, liability for failure, and revision after harm. Each of these mechanisms assumes that the thing being assessed stays still long enough to be assessed. Opponents argue that this technology does not, because capability changes between the start of an evaluation and its publication, and because systems are released to the public as a means of development rather than after it. The mechanism opponents emphasise is irreversibility. A medicine can be withdrawn and a vehicle recalled, but a system released at scale has already been copied, built upon and woven into hiring, credit, policing, education and the public information supply by the time any judgement about it could be reached. Opponents argue that the costs of an error therefore fall on people who neither chose the exposure nor could have known about it, which is the standard case for caution and the one circumstance in which this technology is developed without it. Opponents argue that this is why the dispute cannot be settled by pointing to benefits. They accept that there are real gains and argue that the question is who bears the risk of being wrong. A technology whose harms appear only after wide deployment, and whose developers are rewarded for deploying before the harms are known, systematically transfers that risk from those making the decision to everyone else. On this account artificial intelligence is not good for humanity as it is actually being built and released, whatever a well governed version of it might have been.

Premises

Counter-arguments

Proponents argue that the objection describes a problem of regulatory design, not a verdict on the technology. Staged release, auditing requirements, incident reporting and liability rules exist for other fast moving products and can be applied here, and some already are. They add that caution has its own costs which the argument does not count: withheld medical and educational benefit falls on identifiable people too, and a standard requiring certainty before deployment would have blocked most of the technologies now relied on. The risk of acting must be weighed against the risk of waiting.

Rejecting the premises

[Rejecting P1] Safety assessment routinely handles moving targets through staged release, monitoring and incident reporting, so stability is not the precondition the premise claims. [Rejecting P2] Public release as a means of development is a commercial choice that regulation can forbid, which makes it a feature of the industry rather than of the technology. [Rejecting P3] Harms from withheld benefit also fall on people who did not choose them, so the asymmetry the premise relies on does not by itself favour delay. [Rejecting C] The premises support regulating the manner of deployment, not the conclusion that the technology is bad for humanity.