We've all heard of the following three technologies: police body cameras, artificial intelligence, and facial recognition software. Across the nation, some police departments use body cameras.
Do the three technologies go together -- work well together? The Washington Post reported:
"Axon, the country’s biggest seller of police body cameras, announced that it accepts the recommendation of an ethics board and will not use facial recognition in its devices... the company convened the independent board last year to assess the possible consequences and ethical costs of artificial intelligence and facial-recognition software. The board’s first report, published June 27, concluded that “face recognition technology is not currently reliable enough to ethically justify its use” — guidance that Axon plans to follow."
So, a major U.S. corporation assembled an ethics board to guide its activities. Good. That's not something you read about often. Then, the same corporation followed that board's advice. Even better.
Why reject using facial recognition with body cameras? Axon explained in a statement:
"Current face matching technology raises serious ethical concerns. In addition, there are technological limitations to using this technology on body cameras. Consistent with the board's recommendation, Axon will not be commercializing face matching products on our body cameras at this time. We do believe face matching technology deserves further research to better understand and solve for the key issues identified in the report, including evaluating ways to de-bias algorithms as the board recommends. Our AI team will continue to evaluate the state of face recognition technologies and will keep the board informed about our research..."
Two types of inaccuracies occur with facial recognition software: i) persons falsely identified (a/k/a "false positives;" and ii) persons not identified (a/k/a "false negatives) who should have been identified. The ethics board's report provided detailed explanations:
"The truth is that current technology does not perform as well on people of color compared to whites, on women compared to men, or young people compared to older people, to name a few disparities. These disparities exist in both directions — a greater false positive rate and false negative rate."
The ethics board's report also explained the problem of bias:
"One cause of these biases is statistically unrepresentative training data — the face images that engineers use to “train” the face recognition algorithm. These images are unrepresentative for a variety of reasons but in part because of decisions that have been made for decades that have prioritized certain groups at the cost of others. These disparities make real-world face recognition deployment a complete nonstarter for the Board. Until we have something approaching parity, this technology should remain on the shelf. Policing today already exhibits all manner of disparities (particularly racial). In this undeniable context, adding a tool that will exacerbate this disparity would be unacceptable..."
So, well-meaning software engineers can create bias in their algorithms by using sets of images that are not representative of the population. The ethic board's 42-page report titled, "First Report Of The Axon A.I. & Policing Technology Ethics Board" (Adobe PDF; 3.1 Megabytes) listed six general conclusions:
"1: Face recognition technology is not currently reliable enough to ethically justify its use on body-worn cameras. At the least, face recognition technology should not be deployed until the technology performs with far greater accuracy and performs equally well across races, ethnicities, genders, and other identity groups. Whether face recognition on body-worn cameras can ever be ethically justifiable is an issue the Board has begun to discuss in the context of the use cases outlined in Part IV.A, and will take up again if and when these prerequisites are met."
"2: When assessing face recognition algorithms, rather than talking about “accuracy,” we prefer to discuss false positive and false negative rates. Our tolerance for one or the other will depend on the use case."
"3: The Board is unwilling to endorse the development of face recognition technology of any sort that can be completely customized by the user. It strongly prefers a model in which the technologies that are made available are limited in what functions they can perform, so as to prevent misuse by law enforcement."
"4: No jurisdiction should adopt face recognition technology without going through open, transparent, democratic processes, with adequate opportunity for genuinely representative public analysis, input, and objection."
"5: Development of face recognition products should be premised on evidence-based benefits. Unless and until those benefits are clear, there is no need to discuss costs or adoption of any particular product."
"6: When assessing the costs and benefits of potential use cases, one must take into account both the realities of policing in America (and in other jurisdictions) and existing technological limitations."
The board included persons with legal, technology, law enforcement, and civil rights backgrounds; plus members from the affected communities. Axon management listened to the report's conclusions and is following the board's recommendations (emphasis added):
"Respond publicly to this report, including to the Board’s conclusions and recommendations regarding face recognition technology. Commit, based on the concerns raised by the Board, not to proceed with the development of face matching products, including adding such capabilities to body-worn cameras or to Axon Evidence (Evidence.com)... Invest company resources to work, in a transparent manner and in tandem with leading independent researchers, to ensure training data are statistically representative of the appropriate populations and that algorithms work equally well across different populations. Continue to comply with the Board’s Operating Principles, including by involving the Board in the earliest possible stages of new or anticipated products. Work with the Board to produce products and services designed to improve policing transparency and democratic accountability, including by developing products in ways that assure audit trails or that collect information that agencies can release to the public about their use of Axon products..."
Admirable. Encouraging. The Washington Post reported:
"San Francisco in May became the first U.S. city to ban city police and agencies from using facial-recognition software... Somerville, Massachusetts became the second, with other cities, including Berkeley and Oakland, Calif., considering similar measures..."
Clearly, this topic bears monitoring. Consumers and government officials are concerned about accuracy and bias. So, too, are some corporations.
And, more news seems likely. Will other technology companies and local governments utilize similar A.I. ethics boards? Will schools, healthcare facilities, and other customers of surveillance devices demand products with accuracy and without bias supported by evidence?