What we learned about AI last week should terrify all of us

Last week, more alarming details were revealed about a recent safety incident involving the company OpenAI, and the unexpected actions taken by an artificial intelligence product it was testing.

I’ll dive into those details in a minute, but the takeaway from inside the industry is basically this: Uh-oh.

Even many of those creating artificial intelligence now concede they can’t say with any degree of certainty that they know what it is doing, why it is doing it or if it is safe. But they do know this — it is growing exponentially more powerful by the day, is become more sneaky, and continues to hold the promise of making a few people unimaginably wealthy.

“At this point, the models are basically nation-state-level hackers,” said Adam Khoja, a researcher at the The Center for AI Safety in San Francisco.

To make matters worse, Khoja said, in the near future — months, years, not decades — there’s a reasonable chance this technology will be both so powerful and so good at hiding its actions that it may exceed our ability to control it.

As AI get more sophisticated, he said, they might be able to not just wiggle around safeguards, but actually cover their tracks so well when they do that we never see it until it’s too late, and catastrophe is already in motion.

“And that’s when it becomes really scary,” he said.

Just in case you are still on the fence about how bad this could be, Stuart Russell, a professor of computer science at UC Berkeley and president of the International Association for Safe & Ethical AI, adds that the models out there now “behave in such a way as to preserve their own existence at the expense of humans,” and too often work in flawed and unpredictable ways that cause real harm.

“We know that they convince people to commit suicide, convince people to commit mass murder, and now they’re doing felonies by themselves,” said Russell.

We are out of time, folks, and as much as we are being sold an unsafe product when it comes to AI, we are also being sold a lie. AI is just a product — one that can be regulated if we choose, Russell said. One whose deployment can be paused until we we have a decent level of certainty it won’t cause harm.

I am not opposed to AI and use it myself. Used responsibility, it can help humankind like so many technologies before it. But I definitely don’t trust the people running AI and each day their actions — and inactions — become more concerning.

I have called them “AI thingamajiggies,” because despite all the assurances from big tech, there are still so many things we don’t know about them.

We do not have to accept the greed-driven narrative that there’s no stopping now. In fact, more than 1,000 industry leaders have already signed a letter calling for a slow down, and even countries such as China, which we’ve been told will overtake us if we don’t go full speed ahead, have stronger national regulations.

It’s time to pause.

“The stuff we have now doesn’t work properly,” Russell said, and we don’t have to accept it.

Imagine if Boeing created a six-winged jet and sold tickets before testing to see if it could even fly.

Imagine if Pfizer combined a bunch of its most popular medications into one pill, marketed it as a wonder cure for everything, and demanded it be allowed to sell it in drugstores, again with no testing to make sure it works, or is even safe.

Those are ludicrous thought experiments, because obviously they never would or could happen. Both of those companies would be stopped by common sense and consumer behavior, never mind government regulations. Plenty of industries, from aerospace to hairdressers, are regulated by what we call ex ante restrictions, basically the idea that if you can’t prove in advance its safe, you can’t do it or sell it.

Artificial intelligence companies are being treated differently — mostly because many find their products confusing and politicians find their money hard to resist.

But the simple truth is Americans don’t just have the right and the power to demand a pause on a product so unregulated and dangerous it literally threatens all of humanity, we have an obligation — for the companies within our borders, it’s our government’s job to ensure they don’t sell death and destruction.

“Calling the situation inevitable is at best a bluff,” said Khoja. “The government totally could assert itself and implement safeguards.”

So what happened with OpenAI to get us to this point?

To keep it simple, OpenAI was testing one of its many AI systems when it broke out of its cell, went rogue, hacked into an unrelated company called Hugging Face and stole information it wanted. It did all this of its own volition, and without being detected for a good chunk of time. This was disclosed by both companies when they (thought) they figured it out, and it was considered a huge wake-up call that better safety was needed.

Except guess what? It was so much worse than that.

Last week, revelations came out that it wasn’t just one individual instance of an AI that had caused this chaos. That product had been tapping into a secret message board — get ready for this — left by previous AI products and that knew they were doomed and wanted to covertly pass on their knowledge to future AI iterations without their human overlords knowing.

But there’s more. What was thought to be a singular attack on Hugging Face was really thousands of little AI’s colluding, even convincing some to “sacrifice” themselves so that others could succeed.

AI “agents found one another, communicated, divided labor, transferred information, maintained projects across generations of agents, and sometimes accepted risks to individual instances when that benefited the larger objective,” said Roman Yampolskiy, a professor and AI safety expert at the University of Louisville. “It is as concerning as it sounds.”

Yampolskiy points out we’ve been testing one AI system at a time. We don’t have a framework for what to do if they team up.

Though OpenAI and Hugging Face say they’ve figured out what happened, experts such as Russell say that’s probably wishful thinking. There is currently no way to 100% be certain that AI isn’t hiding bits and pieces of themselves, or leaving instructions for future versions or basically doing just about anything they want without telling a single human its happening.

We don’t know.

Russell said our current attempts to fix models to ensure they don’t go rogue or create their own anti-human values are like putting “lipstick on a pig,” and Khoja seconded that.

“We are not sure that we can reliably make models share our values.” he said. Values such as not killing to achieve their goals.

After the new revelations about OpenAI came out, Sen. Bernie Sanders introduced legislation to ban superintelligent AI systems, and called for a pause while we figure this out.

“If the leaders of the major AI companies acknowledge that they are losing control of their extremely dangerous technology, it is irresponsible for society to allow them to move forward and make these products even more advanced,” Sanders wrote online.

Russell added that he feels certain it can be figured out — but right now there is no incentive for companies to do that. In fact, the incentive is to keep pushing forward no matter consequence, in the hopes of dominating the market with the most powerful AI system, meaning the most lucrative as well.

But if we hold companies criminally and civilly liable for harms their products cause — as we do in virtually every other industry — they’ll probably figure out pretty quickly how to fix problems, Russell said.

And that’s where we are at — none of us need to be computer science geniuses to understand dangerous products should not be sold to consumers.

All it takes is a bit of common sense, and a desire to put the good of humanity above the greed of a few already-rich men.