Despite facial recognition technology’s potential, it faces mounting ethical questions and issues of bias.
To address those concerns, Microsoft recently released its Responsible AI Standard and made a number of changes, the most noteworthy of which is to retire the company’s emotional recognition AI technology.
Microsoft’s new policy contains a number of major announcements.
- New customers must apply for access to use facial recognition operations in Azure Face API, Computer Vision and Video Indexer, and existing customers have one year to apply and be approved for continued access to the facial recognition services.
- Microsoft’s policy of Limited Access adds use case and customer eligibility requirements to access the services.
- Facial detection capabilities—including detecting blur, exposure, glasses, head pose, landmarks, noise, occlusion, and facial bounding box—will remain generally available and do not require an application.
The centerpiece of the announcement is that the software giant “will retire facial analysis capabilities that purport to infer emotional states and identity attributes such as gender, age, smile, facial hair, hair, and makeup.”
Microsoft noted that “the inability to generalize the linkage between facial expression and emotional state across use cases, regions, and demographics…opens up a wide range of ways they can be misused—including subjecting people to stereotyping, discrimination, or unfair denial of services.”
Also read: AI Suffers from Bias—But It Doesn’t Have To
Moving Away from Facial Analysis
There are a number of reasons why major IT players have been moving away from facial recognition technologies, including limiting law enforcement access to the technology.
Automated facial analysis and facial recognition software have always generated controversy. Combine this with the often inherent societal biases of AI systems and the potential to exacerbate issues of bias intensifies. Many commercial facial analysis systems today inadvertently exhibit bias in categories such as race, age, culture, ethnicity and gender. Microsoft’s Responsible AI Standard implementation aims to help the company get ahead of potential issues of bias through its outlined Fairness Goals and Requirements.
Appropriate use controls
Regardless of Azure AI Custom Neural Voice’s boundless potential in entertainment, accessibility and education, it could also be greatly misused to deceive listeners by impersonating speakers. Microsoft’s Responsible AI program, plus the Sensitive Users review process essential to the Responsible AI Standard, reviewed its Facial Recognition and Custom Neural Voice technologies to develop a layered control framework. By limiting these technologies and implementing these controls, Microsoft hopes to safeguard the technologies and users from misuse while ensuring that their implementations are of value.
Lack of consensus on emotions
Microsoft’s decision to do away with public access to the emotion recognition and facial characteristics identification features of its AI is due to the lack of a distinct consensus on the definition of emotions. Experts from within and outside the company have pointed out the effect of this lack of consensus on emotion recognition technology products, as they generalize inferences across demographics, regions and use cases. This hinders the ability of the technology to provide appropriate solutions to the problems it aims to solve and ultimately impacts its trustworthiness.
The skepticism associated with the technology comes from its disputed efficacy and justification for its use. Human rights groups contend that emotion AI is discriminatory and manipulative. One study found that emotion AI constantly identified White subjects to have more positive emotions than Black subjects across two different facial recognition software platforms.
Intensifying privacy concerns
There is increasing scrutiny of facial recognition technologies and their unethical use for public surveillance and mass face detection without consent. Even though facial analysis collects generic data that is kept anonymous—such as Azure Face’s service that infers identity attributes like gender, hair, age, and more—anonymization does not alleviate ever-growing privacy concerns. Aside from consenting to such technologies, subjects may often harbor concerns about how the data collected by these technologies is stored, protected and used.
Facial Detection and Bias
Algorithmic bias sees machine learning algorithms portray the biases of either their creators or their input data. The large-scale usage of these models in our technology-dependent lives means that their use cases are at risk of adopting and proliferating mass-produced biases.
Facial detection technologies struggle to produce accurate results in use cases involving women, dark-skinned people and older adults, as it is common to find these technologies being trained by facial image datasets dominated by Caucasian subjects. Bias in facial analysis and facial recognition technologies yields real-life consequences, such as the following examples.
Regardless of the strides that facial detection technologies have taken, bias often yields inaccurate results. Studies show that face detection technologies generally perform better with lighter skin complexions. One study reports findings of the identification of lighter-skinned males having a maximum error rate of 0.8% compared to up to 34.7% for dark-skinned women.
The failures in recognizing the faces of dark-skinned people have led to instances where the technology has been used wrongly by law enforcement. In February 2019, a Black man was accused of not only shoplifting but also attempting to hit a police officer with a car even though he was forty miles away from the scene of the crime at the time. He spent 10 days in jail and his defense cost him $5,000.
Since the case was dismissed for lack of evidence in November 2019, the man is suing the authorities involved for false arrest, imprisonment and civil rights violation. In a similar case, another man was wrongfully arrested as a result of inaccuracy in facial recognition. Such inaccuracies raise concerns about how many wrongful arrests and convictions may have taken place.
Several vendors of the technology, such as IBM, Amazon, and Microsoft, are aware of such limitations in areas like law enforcement and the implication of the technology for racial injustice and have moved to prevent potential misuse of their software. Microsoft’s policy prohibits the use of its Azure Face by or for state police in the United States.
It is not uncommon to find facial analysis technology being used to assist in the evaluation of video interviews with job candidates. These tools influence recruiters’ hiring decisions using data they generate by analyzing facial expressions, movements, choice of words, and vocal tone. Such use cases are meant to lower hiring costs and increase efficiency by expediting the screening and recruitment of new hires.
However, failure to train such algorithms on datasets that are both large enough and diverse enough introduces bias. Such bias may deem certain people to be more suitable for employment than others. False positives or negatives may be the determinants of the employment of an unsuitable candidate as well as the rejection of the most suitable one. As long as they contain bias, the same results will likely be experienced in any similar context where the technology is used to make decisions based on people’s faces.
What’s Next for Facial Analysis?
All of this doesn’t mean that Microsoft is discarding its facial analysis and recognition technology entirely, as the company recognizes that these features and capabilities can yield value in controlled accessibility contexts. Microsoft’s biometric systems such as facial recognition will be limited to partners and customers of managed services. The availability of facial analysis will continue to be available to users until June 30, 2023, via the Limited Access arrangement.
Limited Access only applies to users working directly with the Microsoft accounts team. Microsoft has provided a list of approved Limited Access use cases here. Users have until then to submit applications for approval to continue using the technology. Such systems will also be limited to use cases that are deemed acceptable. Additionally, a code of conduct and guardrails will be used to ensure authorized users do not misuse the technology.
The Computer Vision and Video Indexer celebrity recognition features are also subject to Limited Access. Video Indexer’s face identification also falls under Limited. Customers will no longer have general access to facial recognition from these two services, in addition to Azure Face API.
As a result of its review, Microsoft announced, “We are undertaking responsible data collections to identify and mitigate disparities in the performance of the technology across demographic groups and assessing ways to present this information in a way that would be insightful and actionable for our customers.”
Read next: Best Machine Learning Software