To keep AI robots from killing humans, look at driverless trucks

• By Thomas Black / Bloomberg Opinion

• • Add TT as Preferred Source

  • 

  • 

  • 

The most bothersome aspect of the warning that artificial intelligence (AI) could wipe out humanity is that there are few details about how a rogue computer model could actually accomplish it.

The top scenarios, and it is by no means an exhaustive list, involve nuclear missiles, biological weapons and killer robots.

We have long been familiar with the first one. A maverick AI takes over the nuclear-launch codes and sends the world down the path to doomsday. (See the 1983 movie War Games.) The second involves AI tricking a human into releasing a deadly virus, which is a new twist on the old concerns of lab leaks by mistake (Outbreak in 1995), terrorist plots (12 Monkeys in 1995) or profit-making schemes (Mission: Impossible II).

While all of those are scary, there are a lot of obstacles before reaching Armageddon, including security agencies keeping close watch for illicit or dangerous biological research. The cybersecurity industry is more than capable of staving off an AI hacker, especially by adopting bespoke AI defenses and tapping into quantum computing for stronger cryptography.

Which brings us to killer AI robots.

This is where the thinking gets a bit speculative, because mobile robots that can perform multiple tasks do not exist now. Pilot programs are experimenting with humanoid robots that perform simple jobs in warehouses and factories, but their capabilities are limited. All the humanoid robots dancing or running marathons on TikTok are preprogrammed or are being controlled remotely by a human operator and are not acting independently. These robots are coming, just not yet.

Still, the idea of squads of killer machines eliminating civilization lives in popular imagination. Luckily, there is already a real-world example of how AI can be released into the world without murderous conspiracy.

The autonomous-vehicle industry is proving that AI can be tightly controlled and work safely around humans. Waymo LLC, Tesla Inc, Aurora Innovation Inc, Kodiak AI Inc and other self-driving companies are paving the way for how AI can be applied to mobile robots that can perform multiple tasks.

It turns out that driving is a good place to start deploying physical AI around humans, because there are already lots of established rules and protocols.

The computer that runs the AI that interprets inputs from vision sensors is on board the vehicle and does not require an Internet signal to operate. Different layers of AI are assigned tasks, such as foundational models that recognize real-world objects, large language models that understand commands, and action models that perform maneuvers depending on sensor data and rules of the road. The coding trail is transparent, so programmers can see how and why the system reached a decision. There is no “black box” AI in which it is impossible to know how the computer reached a conclusion.

As a last layer of defense, the onboard computer has a simpler model that provides a “sanity check” to prevent unsafe or unexpected maneuvers, said Don Burnette, chief executive officer of Kodiak, which has deployed driverless trucks in the Permian Basin oil patch and is poised to put them on interstate highways by the end of this year. In the unlikely event that the driverless system does not understand a scenario, the truck would slow down, move to a safe area and request help.

“We know systematically exactly how it works, and we’ve tested all the failure cases so that nothing can go wrong,” Burnette said in an interview. “If anything is out of the ordinary, it just refuses to execute.”

Autonomous vehicles have a focused job, just as the multitask robots would. The decisionmaking model should be on board the robot, which restricts its computing capabilities because of power and space, and limited to the tasks it needs to perform. The system does not need to know Shakespeare sonnets or the physics of orbital bodies or have the wherewithal to plan a revolt against humanity.

Mobile robots should be regulated by the standards now being worked on by the American National Standards Institute (ANSI) and the International Organization for Standardization (ISO) instead of government regulations, which cannot keep up with the technology. Companies work to obtain these certifications to show they meet industry standards to help win customers and contracts. The ANSI and ISO standards are reviewed and changed much more often than government regulations to keep up with innovation.

One standard should be that these powerful AI models need to be contained and focused on the specific assigned tasks. If a tool is dangerous, it should be treated that way, just as biology laboratories are careful to avoid the release of viruses. Models that can deviate from their instructions should not power the mobile robots that would operate around humans.

A so-called kill switch is complicated, or perhaps impossible, for an AI model that is integrated into the cloud and feeds off multiple interconnected data centers. On physical AI, a kill switch or emergency-stop button makes a lot of sense and should be visible and accessible in case something goes wrong.

Robots are tools to assist humans and do not need to be that smart. They should not accidentally hurt humans and certainly should never have the opportunity to hurt them intentionally. Autonomous vehicles are the blueprint for making sure robots do only what they are told.

Thomas Black is a Bloomberg Opinion columnist writing about the industrial and transportation sectors. He was previously a Bloomberg News reporter covering logistics, manufacturing and private aviation. This column reflects the personal views of the author and does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners.