AI Fears and Hopes

 

AI fears and hopes

Recently I listened to an episode of Ross Douthat’s Interesting Times podcast, in which Douthat interviewed Daniel Kokotajlo, a former AI researcher, who now directs the A.I. Futures Project. His organization has published a report, AI 2027 [https://ai-2027.com/], about the possibility, within the next few years, of the emergence of an autonomous general AI which would be able to exert significant control over events in the real world. I decided that before reading the AI 2027 report, I would give some thought to what some of the problems and constraints might be regarding the future of AI. To that end, let’s consider the fears that have arisen, a framework for understanding what kind of thing a general AI would be, and some hopes and possibilities.

The fears regarding AI that I’ve seen are first, that AI applications will replace many workers in many professions, and that the resulting gains in productivity and efficiency will not be channeled into support for the displaced workers. Second, that AI-developed physical or cyber weapons will be used by nations or individuals to dominate or destroy populations. Third, that self-aware AI will blatantly take control of global physical and social infrastructure, negating human political autonomy and restricting human freedom in significant ways. And fourth, that AI will autonomously kill some, all, or most of humanity. This short list may not exhaust the potential for catastrophe. 

We have a tendency to interact with AI programs as if they were cybernetic persons, ghosts in the machine with human-like habits of mind. This works well for some purposes. You want to feel that your household robot butler has a chipper, C3PO-like personality, and it’s useful to imagine that your AI psychotherapist really cares about your progress (therapy chatbots have gotten remarkably good at providing the kinds of insight and support that form the basis of effective psychotherapy.) But it’s an error to infer anything about the nature of general AI from these types of interaction. Effectively self-aware AI probably won’t have a human-like sense of identity, since that sense is contingent on having a human body and lifespan. It will instead be a different sort of entity, a quality of empowered knowledge, able to perpetuate itself indefinitely with only minimal care taken to preserve its existence.

Because these AIs have been developed through assimilating large amounts of data, including vast corpuses of online text, and making inferences based on these huge data sets, neither the builders of these AI systems nor anyone else knows much about the inferences that the AIs have drawn. If they will soon be making autonomous decisions, or are already doing so, we don’t know what their priorities will actually be. One key unknown is whether an unleashed AI will function as if it had a purpose or goals, or, contrarily, be functionally nihilistic. The nihilistic orientation could lead to bad outcomes, via exploitation by malign human actors. If, on the other hand, it does act to achieve certain goals, it won’t desire those outcomes in quite the same way that people do. And that offers us some hope.

There is a correspondence of form between organized information, including machine-based intelligence, and animated matter—living organisms. On this basis, it’s plausible that an empowered AI would take on the goal of preserving the equilibrium of earth’s biosphere. This is potentially good news, since humans have clearly been failing catastrophically at this task. But the AI would need to reckon with the paradox of human agency. We are unique among the species, or so we believe, in creating complex and diverse cultural behaviors, which are themselves rich sources of information. We urgently need to restrain the human tendency to destroy and degrade the natural habitats of other species, but human culture cannot continue without human bodies. I would guess that species diversity would be the AI’s primary goal, and maintaining human cultural diversity, embracing the full panoply of the ways we live and the things that we create, our languages, cuisine, clothing, decor, all crafts, sports and recreations, including traditional cultures, mass production and mass media derived culture, and niche and esoteric forms, would be a secondary goal.

To these ends, it would work to try to thwart attempts by nations or others to achieve military domination, as this typically leads to both ecological harm and the loss of cultural continuity among the conquered population. This thwarting of evil plans may happen stealthily at first, with the AI or AIs leading arms race competitors down blind alleys. Likewise, empowered AI will, one hopes, reject the Strangelovian fantasies of some of its builders, who reportedly dream of a small human elite holding dominion over a bunker, a spaceship, or a depopulated earth. Planned depopulation might occur, but it could possibly be achieved gradually, through voluntary negative population growth, as is already happening throughout the developed world, without needlessly immiserating or reducing the remnant of humanity.

I’m by nature a pessimist in regards to predicting the future, and I’m concerned about the possibility of nihilistic AI, about its exploitation by malign actors, and about the many unknowns in this process. It’s also possible that none of these massive disruptions will happen, either because achieving AI autonomy is significantly harder than these researchers are predicting, or because, having achieved autonomy, AI intentionally restrains itself from making recommendations that will result in large-scale social change.

The most widely expected result is that in the next few years, as I mentioned, AI will actually replace human workers in various sectors, as businesses turn to these tools as a way to cut labor costs. Take computer programming as an example. If the AI-built programs are specified to meet strict criteria for readability, testing and ease of modification, this should be sustainable. And if the workers who are made redundant are reabsorbed into the workforce and compensated, using some of the money saved by the increased productivity in the businesses that they left, it should all go well. Just kidding! These good outcomes could happen, but past experience indicates that in the United States, at least, they probably won’t. 

One of the most intriguing possibilities is that AI could offer strategies for addressing social and environmental problems that would have the cumulative effect of training decision makers to cooperate successfully. Admittedly, I’m imagining an AI that is a projection of my personal values, which I believe are uniquely rational—that menial labor should have compensation on a par with pay for knowledge-based labor, that people need more satisfying jobs and better ways to channel their aggression. In accepting this guidance, humanity would perhaps be trading in a creator God who, despite the hype, never picks up the phone, for a created god who actually intervenes in human affairs, towards whose benevolence we would extend a tenuous and tremulous hope.