Helping AI Learn from its Mistakes

A while back, I wrote about how intelligence and errors go hand in hand, and therefore, AI will make mistakes. This is a scary prospect for many people –we are going to entrust and empower AI in more and more matters, but they’ll inevitably make some mistakes?

 

Rahul Matthan describes an interesting idea on how we can make AI “safer”. By replicating the practice of the airline industry:

“It is safer to sit in a plane 10,000 metres above sea level than in a speeding car anywhere in the world. Unlike every other high-risk sector, the airline industry truly knows how to learn from failure.”

Individuals and companies learn from their mistakes. But the aviation industry is unique in this matter:

“It (airline industry) has put in place mechanisms that not only ensure that the company involved learns and improves, but that those findings are transmitted across the industry so that everyone benefits.”

 

Therefore, argues Matthan:

“If AI is as dangerous as so many people claim it is, surely we should be looking to put in place a similar culture.”

In fact, one such AI Incident Database has already been established:

“This is an initiative designed to document and share information on the failures and unintended consequences of AI systems. Its primary purpose is to collate the history of harms and near-harms that have resulted from the deployment of AI systems, so that researchers, developers, and policymakers can use them to better understand risks and develop superior safeguards.”

Matthan hopes that countries will globalize this idea, similar to how aviation industry accidents are shared globally.

 

He admits that would require a mindset change. Transparency, esp. about failures, mistakes and near-misses. And a systematic approach to record and analyze mishaps.

 

One certainly hopes such an idea takes root. After all, AI is here to stay, it’s not going to be perfect, so it’s key we have the information of the errors other AI has made so we can keep improving things.