MORE THAN ever, the big questions about artificial intelligence drive the ups and downs of financial markets. The bulls, who think AI will be both transformative and lucrative, stand against the sceptics, who believe the promised returns will never arrive.
But a third sentiment is now coming to the fore for investors: what if AI systems are immensely capable, as the bulls predict, but deeply dangerous? Fears about “misaligned” AI that hacks other systems, enables fraud or misleads human engineers
MORE THAN ever, the big questions about artificial intelligence drive the ups and downs of financial markets. The bulls, who think AI will be both transformative and lucrative, stand against the sceptics, who believe the promised returns will never arrive.
But a third sentiment is now coming to the fore for investors: what if AI systems are immensely capable, as the bulls predict, but deeply dangerous? Fears about “misaligned” AI that hacks other systems, enables fraud or misleads human engineers are growing. Researchers in frontier AI labs worry their tools could help rogue states or terrorist groups build their own bioweapons. Some think the models could result in human extinction, which would take the shine off any future listing. It’s hard to celebrate an initial public offering that leads to a terminal public offing.
Tyler Cowen, an economist, wishes the alarmists would put their money where their mouth is. Many place high odds on a catastrophe. If they are right, then intimations of doom should start affecting the markets before the cataclysm actually arrives. The technophobes could therefore test their theories by making financial trades. If they were successful, it would add credibility to their prophecies well in advance of the disaster that would ultimately vindicate them. This call for intellectual accountability is understandable. Alex Tabarrok, one of Mr Cowen’s colleagues, once called such a bet “a tax on bullshit”. But when it comes to the dangers of advanced AI, the bets can be hard to place.
Some of the milder downsides of powerful AI are already being priced into markets. Concerns about AI-enabled hacking have helped lift the value of cyber-security stocks like CrowdStrike. Its price, which has more than doubled in the past six months, now stands at nearly 200 times its forecast earnings for the next 12 months. For some companies, fear, not just greed, is good.
Other, untested strategies could work for more adventurous worriers. If cyber-security firms are trading at a premium, then surely the companies most vulnerable to hacking and fraud should be trading at a discount. In the private sector, the health-care, finance, telecoms and transport industries are often targeted. Small and medium-sized firms are more vulnerable than their larger rivals. One way to trade on this logic would be to buy credit-default swaps against the debt of regional banks, for example.
For more traditional perils, such as floods, wildfires or even terrorism, the obvious answer would be to buy insurance products. But the market for AI risk has been slow to develop. Even the physical risks faced by data centres are tough to insure. Earlier this year, Insurance Services Office, a standards-provider for the industry, released a new template that excluded damage caused by generative AI from a wide range of claims. Several big firms have asked regulators to let them leave out such novel threats. That wariness has created a gap in the market, which a handful of startups are aspiring to fill. But for now the firms are piddling in size relative to the task.
The problem faced by insurers has deeper roots. They struggle to cover AI risks, because “risk” is too neat and tidy a concept for what is at stake. In economics, risk refers to quantifiable probabilities. The wild and woolly threats associated with AI are harder to pin down. They are new and poorly understood. The base of evidence is thin. Thus neither the buyers nor sellers of insurance know how to price them.
In the long history of human pessimism, there have been many doomsayers who were fast and loose with their predictions. People who scream and shout about supply-chain collapses, hard landings or financial meltdowns could easily make money from their prophecies, if they were as far-sighted as they claim. But AI is different. At the most extreme end of the spectrum, the dangers are intrinsically unhedgeable. Even if you could calculate the chances of an existential threat, there is not much point in buying protection from counterparties that will not be around to pay up.
In the case of an “AI takeover”, where humans are subjugated or sidelined, the big winners will be the AI beings taking charge as the dominant forms of life on Earth. They would make the natural counterparties in any insurance policy a doomsayer might want to buy. But even if this were possible, a superintelligence would make a tough negotiating partner. Unlike humans, they would probably read the small print.
One Subscription.
Get 360° coverage—from daily headlines
to 100 year archives.
Archives
HT App & Website