Yesterday, the White House’s Office of Science and Technology Policy (OSTP) announced an effort to create an Artificial Intelligence (AI) bill of rights. They will engage the American public for feedback in two public listening sessions and six public events in the coming weeks, “…to make sure new and emerging data-driven technologies abide by the enduring values of American democracy.”
The listening session topics range from consumer rights, social welfare, equal opportunities, civic justice and more. Personally, I am excited to hear about this development, but am cautiously optimistic given the lack of a succinct government voice in this space for many years.
AI continues to further integrate into the fabric of our day-to-day lives. I can’t think of one day recently where I haven’t asked my Alexa echo show for the weather, unlocked my iPhone using Face ID, or my Surface Book with Windows Hello, or used an Instagram and Snapchat photo filter. I have autocorrect AI to thank for helping me how to spell pterodactyl. Heck, my Netflix streaming experience is based on AI analysis of how much I love documentaries and Tom Cruise suspense thrillers.
When you couple this effort from OSTP with the recently introduced European Commission (EC) Artificial Intelligence regulation, these are a huge steps in the right direction toward instituting guidelines to protect citizens rather than simply trusting that technology companies will self-regulate and introduce ethical practices on their own.
However, the EC’s AI regulations are far from perfect. It categorizes AI into four areas based on the “risk factor” that is can present to EU citizens: minimal/no risk, limited risk, high risk, and unacceptable. The continued criticism for this type of categorization is that the broad scoping can lead to loopholes with drastic societal implications. Today, the Human Rights Watch published a 28-page report criticizing the EC AI regulations because its’ “narrow safeguards neglect how existing inequities and failures adequately protect rights – such as the digital divide, social security cuts, and discrimination in the labor market – shape the design of automated systems, and become embedded by them.” In other words, it has left gaps in the way that AI is used for social services and employment which can cause discrimination. One example provided by the report includes using facial recognition to verify identities for welfare applicants in Ireland and compare past applicants’ in order to prevent fraud. It is unclear why this process has implemented when other forms of identity verification (such as passport verification or proof of address) should be enough.
The US National Institute for Science and Technology (NIST) in the US conducted a large scale study on facial recognition algorithms in 2019 and determined that they are less accurate for Asian and African American faces compared to Caucasians. This report was cited in the Human Rights Watch’ report and has been used previously as proof of racial bias embedded into AI. The bias introduction has primarily been attributed to a lack of diverse programmers’ perspectives and imbalanced training datasets in the machine learning models. i.e. not enough diverse people of color included in face datasets.
The issue of bias broadens out and can have negative implications on consumers, healthcare, job hiring practices, advertisements, social media and more, which I may dive deeper into other posts.
As we see the ethical and societal dilemmas stack up in the US and globally, the lack of comprehensive US federal regulation has been a critical gap. High profile advocates for regulation, including Brad Smith, VP and chief legal counsel at Microsoft, have touted direct government regulations for many years. Brad Smith most recently wrote about the topic in his book, Tools and Weapons, where he dives deeper into why government regulations are essential as AI development advances.
There have been past efforts to publish recommendations by federal agencies, such as the published guidance from the Office of Management and Budget as of 2020, FTC guidance on commercial implications on AI usage, and the State Department’s recent efforts to counteract the harmful impacts of AI in malicious situations (see State Department AI landing page).
In 2020, Sen. Edward Markey (D-Mass.) and Sen. Jeff Merkely (D-Ore.) led the introduction of a moratorium to ban government use of facial recognition and other biometric technology for state and local law enforcement entities and would “effectively strip federal [financial] support.” The bill also cited the report from NIST which continues to serve as proof that “biometric surveillance systems” are unreliable, not regulated, and should banned from use against US citizens. It remains unclear whether it would actually pass the Senate to law. Similar moratoriums have been considered in the European Parliament, but they also remain divided on the topic.
The lack of federal regulations has left many state and local governments to introduce their own regulations to protect constituents and consumers from AI-biometric data. My home state of Illinois recently passed the Biometric Information Privacy Act, which states that customers have to give consent before companies could use biometric data. There are many other states that have introduced, enacted, or failed to pass laws with similar language.
The inherent bias and lack of ethical oversight has been well-documented by many data scientists and people of color, notably AI ethics researcher Timnit Gebru, who documented her departure from Google last year after raising ethical implications of language models. Gebru continues to advocate for companies to slow down AI development.
This area continues to be a complex, yet fascinating topic. Lawmakers need to catch up and truly start to understand the implications that AI, when used maliciously or without diverse consideration, can have to the societal fabric. Let’s hope the US can get there soon rather than later.