Commentary: AI companies are creating ‘all-powerful psychopaths.’ Maybe not a great idea?

Repeated law breaking. Lies and deceit. Reckless disregard for safety. Lack of regret.

These are characteristics that help identify a psychopath who lives among us but lacks the empathy and morality needed to behave in a civilized or even safe manner.

Unfortunately, it is also becoming increasingly clear that these are the characteristics that define the most powerful AI models being developed at an alarming rate by for-profit companies; These companies would like us to believe that slowing the progress towards AI dominance is somewhere between impossible and insane.

It is neither, and this is not a progressive approach; bipartisan common sense.

“New rules are needed “This new technology frontier is not to stifle innovation, but to ensure that our innovations do not exceed our protections,” the Texas Republican said. Representative Nathaniel Moran he wrote on social media.

He was responding to a recently disclosed incident that shows that we should not create artificial psychopaths without at least some thought.

Silicon Valley giant OpenAI, run by Sam Altman, recently gave two of its models a test to assess how well they could hack on their own. Spoiler: Really good..

The test takes hundreds of known flaws in the software (which have already been fixed for public use) and asks the models to find a way to exploit them, to exploit them if you will, to do something bad, like hack a secure system.

It’s like showing a burglar an unguarded window and asking him to figure out the best way to break in and ransack the place.

These OpenAI models are very smart. They could do as expected and try out each of these flaws, one by one, like beautiful little models. Or they may literally think outside the box.

Even though the models were supposed to be “sandboxed” and unable to access the internet, they went crazy trying to figure out how to break free.

When they escaped into the wilderness, they apparently did so without much difficulty; they just didn’t run. They went on a crime spree with the intention of cheating on an exam because that was the best way to pass quickly.

They targeted another artificial intelligence company called and broke in. hugging faceWhere models doubt the answers to the test. They obtained real credentials, broke into different systems, and eventually captured at least some of the information they were after.

Yes, AI models have figured out on their own that it is easiest to cheat and also how they can get around all the restrictions and do it.

Using Chinese technology, Hugging Face was able to stop the attack before OpenAI reached out to the company to tell it it had happened. At OpenAI’s discretion, He made this incident publicStill, I have to wonder if there might be a way to keep this quiet in the insular tech world.

What bothers me most about this incident is this: This was not a fraudulent action by an “evil” AI. The models were doing exactly what they were supposed to do: achieve the desired result with 100% effort in the method deemed most efficient.

“This is not evidence that AI is conscious, malicious or ‘wanting freedom,'” said Roman Yampolskiy, an AI security expert and associate professor at the University of Louisville.

What the OpenAI models do, he told me, is that forcing these systems to be as powerful, self-sufficient, and goal-oriented as possible “can produce dangerous behavior without malicious intent, and that is arguably the most important problem.”

Call it Murphy’s law, the idea that anything that can go wrong will go wrong.

UC Berkeley professor Stuart J. Russell is also president of the International Assn. Safe & Ethical uses this example for AI: Imagine you ask an AI model to take you to the airport as quickly as possible, but you forget to tell it to obey traffic laws. So you run over a bunch of school kids on the way, but you make your flight. Is this really the model’s fault?

It is almost impossible to think of all the possible ways an AI could perform even the simplest tasks and what the unintended consequences might be; just as it is currently impossible to expect a machine to understand – or inherently value – the emotional or physical consequences of its actions, no matter how hard we try to “train” it to be human or seek that spark of sensitivity.

Russell said performance racing without adequate safeguards seems like bad, undesirable behavior when in fact it’s nothing more than the system.

“I don’t think so [the AI models] “He wanted to hurt Hugging Face,” he said. “I guess they just wanted to pass the test and didn’t care what kind of damage was done to Hugging Face in the process.”

Yampolskiy worries that if this happens next time (and it will), the consequences could be more dire.

Yampolskiy said it was about stealing test answers from a private company. “But the same general capabilities could be directed at financial systems, critical infrastructure, military networks, biological research facilities, or the AI ​​developer’s own security controls,” he noted.

Which brings me back to psychopaths who fail to see or care that their actions cause harm. These models are not humans, despite our many debates about how aware they are or will be. They cannot be expected to be fully aware of the damage they may inadvertently cause; but the people who benefit from them and profit from them can certainly appreciate it.

And these people are acutely aware that they cannot control the creatures they have created, especially after this incident.

“I would say companies are accepting of that, right?” Russell said. “They say, ‘We don’t have a solution to the control problem, but we’re still going to spend $10 trillion to create these very powerful psychopaths.’”

This is where the chorus shouts that if we don’t do it, someone else will. Essentially, the argument here is: Would you rather be destroyed by American technology or Chinese technology?

But Yampolskiy and Russell agree that it’s not inevitable or necessary that we go full steam ahead with so few regulations and so few safeguards.

Russell says that despite American rhetoric, Chinese officials are actually Took a stronger role in regulation Everything the United States does.

“China has clearly said, we want to sit down and create common-sense, basic regulations for all countries so we can prevent this kind of thing from happening,” Russell said. “And the US is ignoring it.”

It’s clear that before any changes can be made in the US, there will need to be a push back from the average people demanding regulation. In Yampolskiy’s words, “the responsibility lies with man.”

None of this is inevitable. None of it has to happen on the timeline imposed on us. We don’t have to allow companies to create models they can’t control, without adequate safeguards to prevent them from remaining free and doing what they want.

We may not even be average people. We can get lost in the fluid language of “exploits” and “zero-day vulnerabilities.”

However, we recognize lying, cheating, and reckless behavior when we see it, whether from a human or a machine.