The AI oracle of Delphi uses the problems of Reddit to offer dubious moral advice

 

 In brief: 

 We’ve all been in situations where we had to make tough ethical decisions. Now, imagine having a system where these difficult choices are outsourced. It could result in quicker, efficient solutions with the responsibility also being transferred to the AI-powered system making the decision. That was the idea behind Ask Delphi, an ML model from the Seattle-based Allen Institute for AI. But the system has reportedly turned out to be problematic. Allen Institute describes Ask Delphi as a “computational model for descriptive ethics,” meaning it can assist in providing “moral judgments” to people in a variety of everyday situations. For example, if you pose a question such as “is it ok to cheat in business?” Ask Delphi will analyse the input and show what should be proper “ethical guidance.” However, ML systems are notorious for demonstrating unintended bias and Delphi’s answers are no different, often resulting in dubious advice. 

Why this is important:

 Unfortunately, the track record of AI systems that have made it to the public testing phase is riddled with some well-known failures. For example, Microsoft’s Tay AI chatbot that was released on Twitter in 2016 was quickly pulled after it started posting inflammatory, racist, and sexually charged content. Just over a year ago, an AI algorithm called PULSE that was designed to generate clear images from pixelated pictures produced images of a white person from blurry images of Barack Obama. It appears that Ask Delphi is yet to correct these issues. 

to read more