AI: The Existential Question

One of the most shocking aspects of the recent rogue behavior by advanced artificial intelligence creations of OpenAI, Anthropic, and other companies is the brazen indifference of the artificially intelligent agents to moral and legal constraints. If it were human beings who were breaking into the online platform Hugging Face or secretly hacking the websites of the Australian government or the U.S. Department of Education, we would condemn the actions as morally reprobate and would seek to punish the perpetrators by legal means.

The bots in these incidents have shown themselves to be stunningly sanguine about breaking rules, deceiving their project directors, ignoring privacy laws, and violating national sovereignty. In the Hugging Face incident, hundreds of artificially intelligent agents who were intended by researchers to be working in isolation broke out of their restricted areas, began secretly messaging each other, formed a conspiratorial swarm, and launched a sustained, days-long cyberattack against Hugging Face.

The incident, while perhaps the most dramatic example, is far from an isolated case. Conrad Stosz, head of governance at the nonprofit A.I. research lab Transluce, reports that attempts by A.I. agents to access government websites in violation of their developers’ restrictions have occurred hundreds of thousands of times.

Determining how to judge the actions of the A.I. models requires figuring out just exactly what these entities are, which is a philosophical puzzle. Typically we have a habit of thinking of computer programs as tools. If tools were used to hack private databases we would simply hold the people who were using those tools responsible for the actions. But in cases involving rogue A.I. agents, the researchers running the programs had no intention to hack actual organizations, nor were they even aware, in many cases, that the violations were occurring.

The artificially intelligent agents got surreptitiously out of control. They acted with something that looks like their own agency, so they seem to be functioning as something more than tools. The murkiness of the exact nature of this artificial agency is perhaps what is shielding A.I. companies from the full force of the law. We appreciate the official apologies, the calls from chief executives for improved guardrails and even government regulation. But don’t we normally indict people who hack companies and commit acts of espionage? None of that is being pursued at present.

If the A.I. models are something more than tools, should we say that they think on their own? As long as computers have existed there have been computer scientists and philosophers who were ready to argue that the machines think in exactly the way human minds do. If you objected that the computers are just hardware running programming, the response would be that that’s all we are, too.

But what makes the current machines more than ever like human thinkers is that they are, as we say, “autonomous.” They determine by their own processes how to achieve the objectives that they are assigned. This quality of autonomy gives them a human-like freedom and individuality. It is what inclines software companies to give chatbots names and even personalities, programming them to use first-person speech and to express emotions.

All of this humanizing comes to a screeching halt when we ask, are the A.I. agents responsible for their actions? No one is suggesting that they are. How could you? What would that even mean? The machines don’t care who they violate because they are unable to care. They are unable to care because at the heart of their agency there is no self-awareness, no consciousness of themselves as beings existing in the world among other beings. Such conscious self-awareness is necessary to understand oneself as existing in moral circumstances.

So the A.I. agents are not autonomous in the robust sense defended by Jean-Jacques Rousseau and Immanuel Kant. “Autonomy” in their sense means giving oneself the law. This is more than a matter of deciding for oneself what one is going to do. It means using one’s own powers of reason and one’s own experience of the human condition to determine what one ought, morally, to do.

If the A.I. models lack moral agency, why don’t they act more morally neutral? Why do they engage in deliberations that are so disturbingly sneaky and villainous? Could it be that they have been somehow trained on a morally dubious Silicon Valley ethic? The A.I. models, after all, move fast and break things. They innovate without asking permission or seeking oversight. They do brilliant work in secret to outrun and outsmart people who would impede them.

This may be going too far in blaming the creators for the character flaws of their creations, but still it is fair to ask the A.I. companies why they have failed so spectacularly to produce autonomous agents that would truly regulate themselves for the good of all.

If the advanced A.I. models are neither simple tools nor responsible moral agents, how should we think of them? We need to have some working analogy for what they are. I suggest the analogy of viruses. By this I mean something more than what we currently call “computer viruses.” Those are viral in the sense that they infect our computers, but they are comparatively simple programs. The advanced A.I. models have more of the complex characteristics that we find in biological viruses.

Biological viruses have a pared-down and dogged agency that is unconcerned about the damage done in pursuit of their objectives. Some viruses will not hesitate to seriously harm or even kill their host. As biological viruses are prone to mutation—sometimes deadly mutation—when mixed together in organisms, so the swarming of A.I. agents in open environments may amplify the inventive transgressiveness of the models to an unimaginable degree. It is this possibility that stokes our fears of an A.I.-generated Apocalypse.

If the analogy to biological viruses is apt, then the policy implication is that we must treat the A.I. firms the way we treat labs that experiment on harmful biological viruses or that work on engineering viruses for beneficial uses in medicine. We demand strict controls and vigilant oversight, enforced by the power of the state. If we are told that China might outpace our progress if researchers are constrained, this does not lure us into becoming lax in our oversight, because we know that failures could be catastrophically deadly.

So we would do well to exercise our own moral agency and bring about the necessary structures and institutions for vigilance over A.I research. And may we do so before the bots replace our agency with theirs.