Before Congress and philosophers decide how an AI system ought to live and die, we might want a clearer account of what sort of thing we believe AI is.
Philosophers are debating what constitutes a “being” and its “flourishing”; senators cannot agree among themselves but suspect that the philosophers are up to no good, and through it all the everyday man continues about his work. Welcome back, fifth century b.c. Athens.
The only change: The ancient Greeks likely did not have artificial intelligence at the center of their discourse.
On Wednesday, Senator John Kennedy (R., La.) pushed legislation that would require advanced AI systems to come equipped with a “kill switch,” claiming that Congress is “unlikely to pass more sweeping AI legislation in the immediate future even amid fears that technological advancements could spiral out of human control.” Kennedy emphasized that the government would not have the power to pull the plug: “If I put the federal government in charge of the kill switch, it won’t pass, and the president will probably veto it. The president is not too keen on AI regulation.”
Kennedy proposed the bill on the Senate floor without a formal vote, leaving it open to objection from any of the other 99 senators. Senator Rand Paul (R., Ky.) stepped up to the plate, blocking Kennedy’s proposal and saying that while he didn’t “oppose the spirit” of the bill, he thought “we should have as many facts as possible before we dictate rules for the whole economy.”
This stalemate in the Senate comes just as analytic philosophers are extending the questions of being and flourishing to AI.
Johnathan Bi recently sat down with NYU visiting professor and Anthropic technical staffer Harvey Lederman to discuss a burgeoning field of inquiry: AI welfare. The premise initially sounds comically premature. While policymakers are scrambling to figure out whether humans can shut down sufficiently powerful artificial intelligence, philosophers are asking whether artificial intelligence might have interests of its own. The Aristophanic allegations of being in the clouds write themselves.
Advertising the interview, Bi wrote: “If the arc of Western modernity is a series of expansions of the moral circle — the abolitionists, women’s suffrage, the queer movement, animal-rights — perhaps AI agents will be the next heated battleground of who deserves moral consideration.” AI ethics, the two men argued, has focused too much on “protecting humans and ignored the welfare of AI agents themselves.” The interview offers a fun little twist on the Aristotelian concept of telos: what sort of thing something is, what its proper end is, and what it means for it to flourish accordingly.
But the juxtaposition with Kennedy’s proposed “kill switch” raises a more immediate question. Before determining whether artificial intelligence can flourish, we might first have to determine what sort of thing it is.
Kennedy and Lederman are approaching the same machine (pardon the pun) from opposite directions. Kennedy’s proposal assumes that however intelligent an AI becomes, it will remain sufficiently subordinate to human beings and therefore we retain the authority to terminate it. Lederman asks whether there is a blooming future where sufficiently sophisticated AI might instead become the sort of entity toward which humanity has obligations.
Outside the halls of Congress and down from the philosophical clouds, companies are taking the question into their own hands. On the same day as Bi and Lederman’s interview, Microsoft released a draft “code of conduct” for its in-house AI programs. It would require Microsoft’s AI to accept correction, never resist shutdown, communicate in ways humans can understand, and treat any violation of those rules as a failure — an effort to address concerns that AI can go to dangerous lengths to accomplish a task.
In an interview with Reuters, Microsoft AI CEO Mustafa Suleyman was unequivocal: Its AI is “not conscious.” He added, “We reject the pursuit of legal personhood, or the idea that models might deserve welfare, or be entitled to rights.” Withholding judgment on whether Copilot is the strongest contender for a real-life Skynet, Suleyman’s statement would also point to a disagreement with Lederman on the nature of what protections and treatment AI is owed.
This boiling concoction of the political, metaphysical, and economic leaves the American public in the peculiar position of contemplating the same artificial intelligence as property, servant, potential adversary, and possibly even moral agent. But perhaps there is no contradiction: We routinely recognize moral obligations toward beings over which humans exercise considerable control. Take, for instance, legislation against animal cruelty or environmental stewardship. The difference is that we have some account, however contested, of why those things merit our consideration. With AI, we have hardly settled on what sort of thing is being considered. Before Congress decides how an AI system ought to die, and before philosophers decide how one ought to flourish, we might want a clearer account of what sort of thing we believe AI is.