your123 - stock.adobe.com
Microsoft's plan to remove AI emotion analysis from Azure and restrict how its customers use facial and voice recognition tools is the latest response by a tech giant to widespread criticism of the technologies' capacity for discrimination and bias.
The vendor framed the decision in several June 22 blog posts as a significant update to the "Responsible AI Standard" it first released publicly in early 2001 and an acknowledgment of the potential of AI systems for speech-to-text and facial recognition to "exacerbate societal biases and inequities."
Microsoft competitor Meta, Facebook's parent company, made a similar move in November 2021 by removing AI-based facial recognition capabilities from the social media platform, but not eliminating the technology entirely from its portfolio.
In explaining its new policy, Microsoft cited evidence that speech-to-text error rates for some Black and African American people are nearly double those among white users of the technology.
The vendor also highlighted the lack of scientific consensus on the definition of emotions and problems with AI facial emotional analysis related to privacy and making inferences across demographic groups and regions.
"It's confirmation that the ability to use facial recognition to do things like identifying emotion or other things, whether it's gender, age or racial background, are actually quite hard to do with a decent level of accuracy," said Merritt Maxim, an analyst at Forrester Research.
However, Maxim noted that, like Meta with Facebook, Microsoft -- while explaining in considerable detail the rationale for its new approach with some of its AI tools -- has not stopped developing facial and voice recognition technology.
Merritt MaximAnalyst, Forrester
"I think they are responding to consumer concerns, and that's generally a good thing. But it doesn't mean the end of facial recognition and therefore it also doesn't mean the end of potential abuses and or biases," he said. "But at least for the time being, Microsoft is kind of taking a time out on this and saying, 'Let's have a look at the technology.'"
Microsoft's new limited access policy on the AI technologies mainly affects cloud-based tools in the vendor's Azure Cognitive Services product line, including Face API, Computer Vision, Video Indexer, and Custom Neural Voice.
The vendor said it would remove from the Face API service capabilities that infer emotional states and identify attributes such as gender, age, smile, facial hair, hair and makeup.
Also, starting now, new users must apply to Microsoft to use facial recognition capabilities in its AI image tools, and the vendor will determine if the applications provide "high-value end-user and societal benefit." Existing users have a year to apply.
One high-profile user of the Face API is Uber, which has used the tool since 2020 to show passengers images of drivers' faces that drivers upload periodically and are compared to their photos on file.
Critics have accused the ride-hailing giant of using a racially biased algorithm for the facial recognition system, alleging that it does not accurately recognize people with darker skin.
Meanwhile, the Custom Neural Voice speech technology enables users to create a synthetic voice that sounds nearly identical to the original source, according to Microsoft. Users include AT&T with its in-store Bugs Bunny video experience and Progressive Insurance with its synthetic voice resembling that of fictional spokeswoman, "Flo."
Microsoft said it will restrict customer access to the service to ensure applications are acceptable and conform with Microsoft's transparency note and code of conduct for Custom Neural Voice.
The tech giant's retreat on the often controversial AI-based technologies, particularly facial analysis tools for emotion, is prudent, said Mike Gualtieri, another Forrester analyst.
"It's more an indication of how good facial recognition, or computer vision, AI has become that they are pulling back on some of it," Gualtieri said. "It works eerily well sometimes, but it also has flaws that produce inaccurate results that lead people to draw improper conclusions about a person."