This is a strange moment to be an active and enthusiastic AI user.
Between reports of unauthorized, "rogue" AIs escaping containment and hiding their tracks, and AI professionals warning us that the technology could kill us all, AI enthusiasm seems almost reckless.
That puts me in an odd position. I've spent much of the past few years writing and speaking about the promise of artificial intelligence and how people can use it in their work and lives. Recently, however, I've been getting fewer questions about how to use AI, and more about whether to fear it. People keep asking me: Are you worried about the AI apocalypse?
My answer: Yes, I'm concerned about the speed of AI development and the lack of meaningful regulation, and I'm disturbed to hear industry insiders laying odds on our collective extinction.
But is the prospect of AI destroying humanity keeping me up at night? Absolutely not.
For me, AI doom talk is fundamentally misleading and counterproductive -- it's speculative, self-serving, disproportionate and stripped of context. Worse, it can leave ordinary people feeling powerless at precisely the moment when we need far more people involved in deciding how AI is built, governed and used.
Just a guess
Start with the doomers' estimates, the idea that there is a 10% or 20% or higher chance that there will be an AI-triggered extinction or mass-casualty event. There is a reason that estimates of AI risk have been all over the place: We don't have the experience or data to meaningfully evaluate the odds of AI destruction. As Princeton computer scientists Arvind Narayanan and Sayash Kapoor have noted, when it comes to AI doom, "forecasters aren't drawing on any special knowledge, evidence, or models that make their hunches more credible than yours or ours or anyone else's."
All these headlines, in other words, are based on guesswork, and they tell us more about who's doing the guessing than about what's likely to happen. As many have observed, some people are drawn to AI precisely because they see it as enormously powerful. (And some of them aren't necessarily opposed to a future dominated by AI.)
What's more, these experts aren't innocent bystanders; they are, by and large, the same people creating this future. But that's a future we should all have the power to shape -- and indeed we can, once we apply some skepticism to these terror-inducing predictions.
For all that we hear about out-of-control AI, there are plenty of sensible and feasible ways to tame that risk. We've faced tech threats before, and negotiated our way to a safer world. Some have cited the International Atomic Energy Agency as a model for how we could collectively reduce our AI risks by agreeing to rules that AI developers around the world would have to follow, regardless of where they're based.
A multilateral agreement might begin with an extended moratorium on training new models, which is both resource-intensive and the source of the greatest risk. And it should certainly include rigorous standards that hold companies accountable for ensuring both the safety of AI models and the safety of AI users, backed by meaningful (as in well-funded) enforcement.
If we can slow down the pace of AI development while accelerating investments in AI safety, we give regulators (and citizens and business owners) a chance to catch up. We can drive years of scientific and productivity advancement on the strength of the AI models we have today. Giving the world a chance to catch its breath is how we bring more people to the table, so that the future of an AI-enabled world isn't determined by the handful of people at the forefront of the industry.
We're terrible at risk
Then there is another reason why I'm not so worried about the risk of extinction: Terrifying predictions make for an effective AI hype strategy, in part because humans are irrational when it comes to evaluating risk. We have a tendency to pay a lot of attention to sensational risks -- like stranger abductions -- rather than to the boring but higher probability events that can also cause harm or destruction.
And so we worry about train derailments, lightning strikes and killer bees, even though we're far more likely to die from choking, falling or drowning in a bathtub. Just think about the people who are so afraid of flying that they drive long distances rather than get on a plane, even though the odds of dying in a plane crash are incalculably small, while the lifetime odds of dying in a car crash are about 1 in 100.
Prophecies about AI destruction trigger an inner alarm that makes it difficult for us to weigh the pros and cons of something that is big and new and scary and different, like AI.
And that brings me to the final reason I don't worry about an AI apocalypse: My strong belief is that AI is a risk worth taking -- mainly because there are so many threats that are far more immediate and quantifiable, and that AI may help us address. I worry a lot less about AI killing humanity than about the risk that humanity will meet its doom in some other way.
We're already witnessing the impact of pandemics, antibiotic resistance and climate change: Experts say there is a 50% chance we'll experience another Covid-scale pandemic within the next 25 years. In that same period, drug-resistant bacteria are projected to kill more than 39 million people. Then there's the prospect of devastating displacements caused by weather-related disasters, the death of much of the earth's coral reefs, and large parts of the population being exposed to extreme heat.
Admittedly, none of that is as terrifying or sudden as an AI-triggered nuclear war, but I've been anxious about nuclear war for more than 40 years, and along with the risk of triggering nuclear war, AI offers some possibility of reducing that risk (through better monitoring and deterrence). When it comes to the other global risks that genuinely do keep me up at night, AI offers even more hope, thanks to advances in basic research that could lead to new vaccines or treatments for some of the diseases that we know kill millions of people a year.
Yes, it's early days on all these fronts, and AI companies have plenty of reason to exaggerate their promise or supposed breakthroughs. When it comes to clean-tech advances, in particular, it will take a pretty big leap forward for AI to offset its impact in driving up global energy usage. But it still feels more warranted and useful to focus on AI's emergent potential to reduce mass death than to get trapped in speculation about its hypothetical risks.
The real dangers
In fact, when it comes to the harm that AI will do, I'm more interested in the harms that are already visible and quantifiable, like cognitive offloading and our tendency to abdicate human judgment in the face of an AI's authoritative-sounding answer. Or the impact of AI on employment, as automation and AI-enabled surveillance become widespread. Or the way AI shores up authoritarianism and undermines democracy through the mass mining of citizen information, the explosion in AI-generated misinformation and the further erosion of independent media.
These are the harms we could -- and should -- address with the same measures we need to address fundamental AI risks. Whether you're worried about your child developing a delusional preoccupation with AI, or about a terrorist using AI to unleash a pandemic, the path forward is the same: We need to regulate the AI industry as a whole, and we need AI companies to commit the same kind of resources to safety that they are currently pouring into model development, marketing and data-center construction.
AI doomerism and AI resistance might look like opposites, but they share a similar assumption: that AI's trajectory is largely inevitable and our main role is either to fear it or opt out.
But we still have plenty of choices, and they're not between building a bunker or pulling the plug. We can take charge of AI, and use it to shape a better future -- one in which we not only avoid destruction by AI, but reduce our other catastrophic risks, too.