Sign In

Communications of the ACM

ACM TechNews

People's Notions About AI Are Terrible, an MIT Study Asks Whether They Can Be Helped


View as: Print Mobile App Share: Send by email Share on reddit Share on StumbleUpon Share on Hacker News Share on Tweeter Share on Facebook
The painting ingested tons of artwork samples from artists through the ages to become tuned to produce art of a certain style.

'Edmond de Belamy,' produced by the art group Obvious and auctioned at Christie's in 2018 for $432,500, relied on generative adversarial network algorithms developed over years by various parties.

Credit: MIT

Researchers at the Massachusetts Institute of Technology (MIT), Vassar College, and the Max Planck Institute for Human Development in Germany surveyed several hundred people to determine whether they view artificial intelligence (AI) like an agent or a tool.

The study focused on a painting produced using a generative adversarial network that sold at auction in 2018.

Study participants were asked to rate how much each party that took part in the creation of the painting should be given credit, and how much they agreed with various statements about the fictional algorithm that implied agency.

The researchers found a significant correlation between the extent to which participants agreed with statements about the algorithm's agency, and how much credit they gave to different parties.

Said MIT's Ziv Epstein, "By tweaking the language, to up-play the agent-ness of the AI ... we can manipulate the allocation of money and responsibility."

From "People's Notions About AI Are Terrible, an MIT Study Asks Whether They Can Be Helped"

ZDNet (09/18/20) Tiernan Ray
View Full Article

 

Abstracts Copyright © 2020 SmithBucklin, Washington, DC, USA


 

No entries found