Many of the smartest people working in artificial intelligence argue that these systems work best paired with human beings, letting each partner do what they do best. But dividing the tasks is often trickier than it sounds.
Why it matters: Figuring out where humans need to play a role — whether for their common sense, skills or ethics — is a key task facing society as AI systems grow increasingly powerful.
Some of these tensions are already apparent in the field of autonomous and semi-autonomous vehicles.
- Humans are highly flawed, prone to distraction, error and inebriation.
- But computer systems remain expensive and struggle to handle the array of conditions that crop up on the road.
Yet there's no simple method for combining the skills of both people and AI in the real-time decision-making driving demands. At a certain point, either the computer or the person has to be in control.
- There is a broad sense that fully autonomous cars could save lives lost to human error. But, with widespread availability of such technology still far off, the industry has been wrestling with how to incorporate some autonomy into cars while still relying on humans at the wheel.
- A key principle, says Avinash Balachandran, director of human interactive driving at Toyota Research Institute (TRI), is making sure that human beings are in control most of the time so they don't lose focus, with the computer either invisibly assisting or taking over only when absolutely necessary.
Context: Toyota opened the doors of its research center in Los Altos, Calif. to reporters for the first time in five years on Wednesday — in part to discuss the work it is doing to ensure humans and machines can complement one another.
The big picture: Society has been captivated by both the capabilities and glaring shortcomings in the latest generation of generative AI technologies, such as OpenAI's ChatGPT.
- Experts say that's a harbinger of what the next few years hold, and it is up to society to harness such systems' power, while ensuring humans are involved to limit the many potential missteps systems make on their own.
Between the lines: Getting humans and AI systems to work together well requires a sober and sophisticated understanding of where each excels and where they stumble. It also requires a mix of intentionality and hard work.
- "I love people who are optimistic about the future," says Gill Pratt, who runs the Toyota lab. "I am one of them, but we have to have our eyes open."
AI systems are really good at digesting vast amounts of data and increasingly good at presenting their findings in creative and even inventive ways.
- At the same time, such systems lack common sense, struggle to cope with novel situations, and can incorporate and automate existing bias, among other problems.
- "Data is fundamental but it has big gaping holes," said Adrien Gaidon, director of machine learning at TRI.
Humans, for all their flaws, are amazing at tasks that computers still struggle with. Even toddlers readily master tasks like understanding object permanence or identifying novel obstacles.
- Importantly, humans can learn by venturing out into the world and having new experiences, while AI systems are largely limited to the diet of digital information on which they are trained.
- But people are also often not at their best, thanks to boredom, distraction, exhaustion, moodiness or many other all-too-human conditions.
Of note: The argument for humans and AI working in tandem often presupposes the pairing of a trained computer system and a worker with years of hands-on experience.
- But as AI systems take on more tasks, over time there may be a dwindling number of humans with the needed expertise.
Be smart: Ideally, says Balachandran, computers can be trained to make humans better over time, but that often doesn't happen.
- He points to the fact that spell check has reduced errors but hasn't necessarily made humans better at spelling.
- If people are to remain at the wheel, he argues that the system should not only step in when needed, but also help humans become better drivers.
