Artificial, With Questionable Intelligence

A common theme is emerging in the race to develop big machines that can navigate through a world filled with people, animals, and other assorted objects—if an accident is inevitable, what options are available to machines and how should they decide?
 
This question was raised at a number of semiconductor industry conferences over the past few weeks, which is interesting because this idea has been kicked around for at least a couple years. But in the past, it was merely theory. Now, as autonomous vehicles begin to roll onto city streets and highways and real accidents occur, this issue has seen a sustained resurgence.

On a business level, the rollout of autonomous vehicles has been generating an increasing number of questions about who is financially responsible for these accidents. AI has fascinating possibilities, and machine learning/deep learning are valuable tools, but these technologies are still deep in the research phase even though they are being applied to real products today. Changes to algorithms are being made almost daily, which is why most of the chips being used for these devices are either inexpensive arrays of GPUs for the training, or some type of programmable logic for the inferencing.

AI/ML/DL collectively represent a potential bonanza for the tech world in general, and the semiconductor industry in particular. But this is an opportunity that also has sharp teeth. There is a very real possibility that the chain of liability could extend far deeper into the supply chain than in the past. The chip industry has been remarkably successful in avoiding liability issues so far, in large part because most designs have not been aimed at safety-critical markets.

As chipmakers, tools vendors and IP developers race into the automotive world, they are being forced to deal with very different kinds of issues. For example, if machines can perform tasks better than people, is that considered good enough? If the number of roadway fatalities decrease with the introduction of autonomous vehicles, is there an acceptable accident rate?

Accidents will continue to occur, with or without autonomous devices, even with more nines added to right of the decimal. The odds of failure increase as the number of components and the amount of software content continue to rise and the supply chain continues to expand. This problem extends well beyond just cars, too. It includes drones, robots of all sorts, and even seemingly benign devices such as household appliances. As connectivity increases, so does the potential impact of failure of any sort, whether it’s caused by equipment failure, unexpected interactions and corner cases, or malicious code.

Quality, coupled with best practices, has been a solid defense against liability claims in the past. But two things are changing. First, technology increasingly is crossing into the world of functional safety, a trend that is only going to increase. And second, machines that react to stimuli on their own likely will develop unique behaviors over time, so defining quality and best practices becomes significantly more difficult.

The chip industry is just beginning to grapple with these issues, which is why discussions about rules, standards and ethics are surfacing at semiconductor conferences—and in conference rooms of semiconductor companies around the world. The scenario cited most frequently (in one iteration or another) is, if an accident is inevitable, does an autonomous device hit a person or risk injuring the occupants of the vehicle by veering off the road.

That leads to several important questions. First, who will make that determination? Second, on what basis will that determination be made? And third, will that determination still apply if the behavior of a machine isn’t entirely predictable?

This all adds up to a lot of questions, and not so many answers.