Outcome

Short Description:

The Sinful Seven project or #Seven project explores the two contradictory forces growing as computers and AI become smarter and more complex: AI will allow humans to do more than ever imagined and AI will also be the extremely dangerous as it will be trusted for a lot of decision making. In our example, the situation with AI is analogous to the number #Seven because while #Seven is recognized as a holy number in the bible, #Seven is also the number of deadly sins.

#Seven explores a scenario where the AI has the ability to learn what the user enjoys and likes to do to recommend the most pleasing activities. At first, this is great for the user as #Seven plays music and recommends coffee in the morning. However, after #Seven observes the user’s pleasure after it recommends gambling, it quickly recognizes how some behaviors, though destructive to the user, are extremely pleasing to them. In our example #Seven recommends the user to participate in activities that represent the #Seven deadly sins: Gluttony, Lust, Greed, Pride, Wrath and Envy. The results of #Seven’s encouragement are catastrophic for the user and demonstrates the design challenge and responsibility engineers have when designing AI. 

Product Video: https://youtu.be/MtJUZlugYNo

Design Recommendations:

1. Draw a line between intelligence and control

Designers should set boundaries for which the learning AI cannot cross. For example, the possible destructiveness to the user should far outweigh the user’s desire to have fun in the decision making process of the AI. The challenge here is keeping the AI from being intelligent enough to undo the constraints set by the user.

2. Limit the list of possible recommended activities

Designers could also limit the amount of activities the AI can observe to benign activities like listening to music or hanging out with friends. However, such limitation of the solution space, limits the positive potential of the product which means that this solution space should be fairly wide

3.  Limit the amount of possible inputs

In our example, the AI has the ability to recognize pleasure in speech as well as through Social Media and other electronic inputs such as browser history. This wide range of inputs causes the AI to define humanity as it can observe both the fun behaviors that are constructive like listening to music or hanging out with friends and destructive like gambling and watching porn. If designers limit the number of inputs the system has to make judgements from, it will be less likely to become destructive.

4. Check commands with a “central brain”

Another interesting concept would be to create a sort of “central brain” that all commands put out by the AI must run through first and be checked over before recommended.