AI: Danger from the usual suspects, tech bro version

Is there an existential danger from AI? Yes. But it’s not from a real-world Skynet. Most of all, that danger comes from powerful oligarchs using AI for selfish purposes, and especially the tech bro billionaires.

10% chance of extinction?

Recent reports, from AI companies and others, have described this danger in apocalyptic terms. In one, a particular statistic is quoted: that AI has a 10% probability of ending our species. That would be beyond frightening – as some have noted, there would be no way you’d get on a plane that had a 10% chance of crashing. Even if 1 out of 100 commercial planes crashed, you’d find another way to get where you are going.

Please.

I agree that there are serious dangers. But that 10% doesn’t come from any real data. The panicked researchers pulled it out of their butts.

What are the most likely dangers?

No one knows exactly what modern or future AI systems will do. In a much more benign way, this was also true of the neural network systems that colleagues and I put together in the 1980s and 1990s. Our networks often had millions of trained parameters, and unlike earlier forms of AI that were constructed with logical rules, developers could have little idea of what was going on inside a system that was just an elaborate collection of elements that summed inputs that were weighted by training on a lot of data. (I’m ignoring some details, but that was most of it).

Modern forms have some new architectures, and some new processing steps, but the biggest difference is scale – while modern AI systems are also based on neural networks, they have trillions of parameters, trained on enormous amounts of data. That’s bigger than the ones we used to build by a factor of a million. There is literally no way to figure out what is going on inside these things.

Given that, could such systems develop internal goals that lead to human disasters? Yes, of course, much as systems that control nuclear weapons could even now. But there is no good way to really come up with a number to say how probable that is. The recent “jailbreaks” of an AI system certainly suggested that many bad outcomes are possible, and this does require a serious multinational discussion on guardrails. Or as a former colleague who is at the forefront of modern AI development told me, “Perhaps cybersecurity is where I see some potential for short-term harm (specifically worrying about an incident with the banking system, for example), but as you said, it would probably be human initiated. It’s relatively easy to set up an autonomous system to find vulnerabilities, so even if human initiated, there’s a high degree of automation, and a low barrier to entry, which is a big delta from one year ago.”

Are mass deaths more probable from this development than from others? I don’t think anyone knows.

But what we can say for certain is that we want those in power to be concerned for our welfare. And in the U.S. these days (and in many other countries, such as Russia), that does not seem to be the case. So much of the population could be at serious risk. 

Though as one other former colleague told me, he thought the Amish would be fine.

Where do these dangers come from?

I would argue that they come from the same sources that we see in so many other problems in today’s world – from influential and malevolent men of wealth, who have a significant effect on elections, wildly disproportionate to their numbers. And if, influenced by the communication powered by this money, voters elect people who are more loyal to their donors than to the people, our country will be unable to move forward. And dangers that include AI misuse and many other problems will grow.

Consider the suppression of information about climate change, and the active obstruction against any action to diminish it. Or the destruction of USAID that is likely to lead to the death of millions of children. Or the reductions in food aid at home, and the increased difficulty of receiving good health care. While there have always been problems in our postwar world, the rise of these malign forces has greatly expanded them, especially in the last 20 months.

And AI, as it now is implemented, is an extremely powerful tool that can (and frankly, will) be misused by these forces, both directly and indirectly. Directly in that it can empower dictatorial control in a way that was not feasible before; and indirectly, in that the deranged actions of the oligarchs can lead to societal chaos that will prevent effective safeguards from being put into place for AI.

So here is the immediate task for our country: this year, elect better political representatives, and in 2028 elect a knowledgeable and compassionate President. Without that, the risk of disasters from AI and more will be much higher. That’s a necessary response to scary news about AI systems.

And we’ll need a lot of volunteers to counter the effects of the vast sums of money oligarchs are putting into elections.