Emotional Bias in AI

2020-01-30

The progress of AI technology is sometimes exciting and sometimes chilling. Most industry professionals are excited about the development of technology that is pushing new heights, while some experts are concerned about the potential negative consequences that this technology may bring in the near future or may have already brought.


AI algorithms are becoming increasingly human-like in the emotional space. Nowadays, artificial intelligence is able to read people's emotions. The emergence of computer vision and facial recognition technology has prompted researchers to actively develop algorithms that can identify human emotions and intentions, and make other inferences.As pointed out by MIT Management Sloan School, emotional AI is a subset of artificial intelligence that can recognize, understand, simulate, and respond to human emotions. It is also known as affective computing, and the history of this field can be traced back to at least 1995 when MIT Media Lab professor Rosalind Picard published the book "Affective Computing".


Javier Hernandez, a research scientist at MIT Media Lab's Affective Computing Group, explains emotional AI as a tool that allows for more natural interaction between humans and machines: think about how you interact with other people, you look at their faces, you look at their bodies, and then you change the interaction accordingly. If a machine doesn't know its emotional state, doesn't know how it feels, and doesn't know how to respond to specific content, how can it effectively convey information?


As explained by MIT Sloan professor Erik Brynjolfsson, "While humans may currently have the upper hand in understanding emotions, machines are leveraging their advantages and constantly evolving. Machines are very good at analyzing large amounts of data, they can listen to changes in speech, and they start to recognize when these changes are related to stress or anger. Machines can analyze images and pick up on subtle nuances in human facial microexpressions, even ones that humans may not be able to recognize."


Researcher Dr. Lauren Rhue found evidence of systemic racial bias in two popular emotion recognition programs: when she ran Face++ and Microsoft's Face API on a dataset of 400 NBA player photos, she found that both systems assigned more negative emotions to black players regardless of how much they were smiling.


There is little evidence to suggest that these new emotion recognition products have any scientific validity. Researchers at UC Berkeley found that to accurately and highly consistently detect emotions, background information beyond facial and bodily cues is needed.



Others at the University of Southern California called for a pause in the use of certain emotion analysis technologies at the 8th International Conference on Affective Computing and Intelligent Interaction in 2019. This facial expression recognition technology is on the rise, but it is not without its flaws. It only makes mistakes, and in some cases, these mistakes can cause harm.


A significant evaluation report released last summer found that efforts to "read" people's internal states solely through the analysis of facial movements without considering the background are at best incomplete and at worst completely ineffective. After reviewing thousands of studies on emotional expression, the report's authors found that despite claims of being able to detect emotional states, these technologies actually achieve much more modest results: detecting facial movements.


Research shows that people have significant differences in the way they express emotions based on culture, context, and interpersonal communication. Furthermore, the same combination of facial movements (e.g., a smile or a frown) can express more than just one emotion. Regardless of how complex the algorithms of affective computing are, it is premature to use this technology to judge people's feelings.

RECENT AI NEWS

RECENT AI TOOLS