The saga of algorithmic bias and the APRA has taken a new turn. The privacy bill would impose racial and gender quotas on artificial intelligence algorithms. Two weeks ago, I profiled this effort in a detailed article about the Warlock conspiracy.
A lot has happened since then. Most importantly, the publicity surrounding its quota provisions forced APRA’s drafters to retreat. A new discussion draft was released that removed much of the quota-driven language. A day later, the House Commerce Subcommittee marked up the new draft. It was really more of a non-marking; member after member insisted that the new draft needed further revisions, then withdrew the amendments pending further discussion. With this vague endorsement, the subcommittee sent APRA to the full committee.
Still, it is good news that APRA has now omitted the original disparate impact and quota provisions. There was no explanation for the change, but it was clear that few in Congress wanted to see quotas mandated in algorithmic decisions.
That said, there is reason to worry that drafters still want to sneak algorithmic quotas into most algorithms without having to defend them. The new version of APRA has four provisions on algorithmic discrimination. First, the bill prohibits the use of data in a manner that “discriminates or otherwise results in equal access to goods and services” based on various protected characteristics. Second. 113(a)(1). This promising start was immediately undercut by a second provision, which allowed discrimination in data collection, either through “self-testing” to prevent or mitigate unlawful discrimination or by broadening the scope of applicants or clients. ID. At (a)(2). Article 3 requires users to assess the algorithm’s “likelihood of causing harm, including to a person or group of people on the basis of a protected characteristic”. ID. (b)(1)(B)(ix). Finally, in the assessment, users must provide details of the steps they have taken to mitigate such harm “to individuals or groups.” ID.
The requirement for self-assessment clearly motivates algorithm designers to be fair not only to individuals but also to population groups. Algorithmic harm must be assessed and mitigated not only on an individual basis but also on a group basis. Judging a person based on his or her group identity sounds a lot like discrimination, but APRA ensures there is no liability for such judgments; it defines discrimination to exclude measures taken to expand the client or applicant pool.
Thus, despite its vague wording, APRA could easily be interpreted as requiring algorithms to avoid harm to protected groups, an interpretation that quickly led to quotas being the best way to avoid harm to groups. Of course, agency regulators will have no difficulty providing guidance to achieve this outcome. They need only state that an algorithm harms “a group of people” if it does not ensure that “a group of people” gets a proportionate share in the distribution of jobs, goods and services. Even a private company that likes quotas because they are a cheap way to avoid accusations of bias can impose them and then invoke two statutory defenses—that its self-assessment requires adjustments to achieve group justice, and that the adjustments are not protected by law Influence.
In short, while the new APRA is not as stunningly mandatory as its predecessor, it may still encourage adjustments to the algorithm to achieve proportional representation, even at the expense of accuracy.
this is a big problem. It far exceeds quotas in academic admissions and employment. It will integrate “group fairness” into various decision-making algorithms—from bail decisions and health care to Uber rides, facial recognition, and more. What’s more, because it’s difficult to determine how machine learning algorithms achieve their exceptionally accurate results, the designers of these algorithms may be able to sneak race or gender factors into their subjects without telling them, or even their users. of products.
This process is already well underway—even in healthcare, sacrificing algorithm accuracy for proportionate results can be a matter of life or death. A recent paper on bias in health care algorithms published by the Harvard School of Public Health suggests that algorithm designers protect “certain groups by inserting an artificial standard into the algorithm that overemphasizes these groups and deemphasizes others.” ”.
In fact, experts on algorithmic bias often recommend such crude interventions that confer artificial advantages through race and gender. As a result, McKinsey Global Institute recommends that designers impose so-called “fairness constraints” on their products to force algorithms to achieve proportional outcomes. Methods it considers valuable include “post-processing techniques” [that] Transform the model after making some of its predictions to satisfy fairness constraints. Satisfying fairness constraints, model accuracy decreases. In each case, the designer’s social justice perspective is hidden by the fundamental characteristics of machine learning; the machine produces results that are rewarded by the trainer. As a result, this is what the machine will produce.
If you want to know how far these limitations are from reality, take a look at the original text-to-image results generated by Google Gemini. When asked for photos of German soldiers in the 1940s, Gemini’s training required it to provide photos of black and Asian Nazis. 
The consequences of bringing this kind of political correctness into health care decisions can be devastating and harder to detect.
That is why we cannot afford APRA’s quota-pushing approach. The answer is not to simply remove these regulations, but to directly address the problem of hidden quotas. APRA should be amended to clarify the underlying principles that require special justification for identity-based algorithm adjustments. They should be a last resort, used only when actual discrimination has been shown to have distorted the outcome of the algorithm and other remedies are insufficient. Never use them when significant bias can be eliminated by improving the accuracy of the algorithm. For example, facial recognition software from ten or fifteen years ago had difficulty accurately identifying ethnic minorities and dark-skinned people. But now, better lighting, cameras, software and training gear can go a long way to overcoming these difficulties. Such improvements in algorithm accuracy are more likely to be viewed as fair than forcing identity-based solutions.
It is also important that the introduction of race, gender and other protected characteristics into the design or training of algorithms should be open and transparent. Controversial “group justice” measures should never be hidden from the public, algorithm users, or individuals affected by these measures.
With these factors in mind, I’ve made a very rough cut of how APRA might be modified to ensure it doesn’t encourage widespread implementation of algorithmic quotas:
“(a) Except as provided in subsection (b), race, ethnicity, national origin, religion, sex, or other protected characteristics shall not be used to modify, train, prompt, reward, or otherwise design a covered algorithm— —
(1) Affect the results of the algorithm or
(2) Produce a specific allocation of results based in whole or in part on race, ethnicity, national origin, religion, or sex.
(b) Covered algorithms may only be modified, trained, prompted, rewarded, or designed as described in subsection (a):
(1) Correct, to the extent necessary, one or more proven discriminatory practices that directly affect the data on which the algorithm is based, and
(2) Whether the algorithm is designed to ensure that whenever a modified algorithm is used, any party adversely affected by the modification can be identified and notified.
(c) An algorithm modified under subsection (b) shall not be used to assist in any decision unless the parties adversely affected by the modification are identified and notified. Any party receiving notice may challenge whether the algorithm complies with subsection (b). “
It’s unclear to me whether such a provision would survive in the Democratic Senate and Republican House. But the makeup of Congress could change dramatically in a matter of months. Furthermore, regulating artificial intelligence is not just a concern for the federal government.
Left-leaning state legislatures have taken the lead in passing laws targeting AI bias; last year, the Brennan Center identified seven jurisdictions that have proposed or enacted laws addressing AI discrimination. Of course, the Biden administration is pursuing a number of anti-bias measures. Many of these legal measures, along with a broader push for ethics guidelines targeting AI bias, will have the same quota-driven impact as APRA.
Conservative lawmakers have been slow to respond to enthusiasm for AI regulation; their silence guarantees that their constituents will be governed by algorithms written to blue state regulatory standards. If conservative legislators don’t want to introduce invisible quotas, they will need to pass their own laws to limit algorithmic race and gender discrimination and require transparency when it comes to modifying algorithms using race, gender and similar characteristics. So even if APRA is never revised or passed, the language above, or some more subtle version of it, could become an important part of the national AI debate.
