[This version is an update that went live at 1:45 p.m. on 12/12.] One of the first laws in the United States, New York City’s Automated Employment Decision Tool (AEDT) law was meant to go into effect on January 1 with the goal of decreasing prejudice in AI-driven recruiting and employment choices.

In a statement released this morning, the Department of Consumer and Worker Protection (DCWP) said it will delay implementation until April15, 2023. After receiving “a substantial number of public comments,” the agency has decided to hold a second public hearing.

Artificial intelligence and algorithm-based technologies cannot be used in the hiring or promotion of New York City residents or workers, according the AEDT statute, unless a thorough bias audit has been performed. The final line is that businesses in New York City, not the developers of AI technologies, will be responsible for meeting regulatory requirements.

Debevoise & Plimpton lawyer and co-chair of the firm’s Cybersecurity, Privacy, and Artificial Intelligence Practice Group Avi Gesser said many issues remain unresolved regarding the laws.

Even though the DCWP published suggested guidelines for enforcing the legislation in September and asked for feedback, the final regulations detailing the format of the audits have not yet been published. That makes it difficult for businesses to know what to do next to stay within the law.

Before the delay was announced, Gesser told VentureBeat, “I think some firms are waiting to see what the rules are, while some assume that the regulations will be implemented as they were in draught and are behaving accordingly.” A large number of businesses “are not even clear if the regulation applies to them.”

As more and more businesses resort to AI to aid in hiring and other HR processes, the city enacted the AEDT ordinance to regulate their use. According to a poll conducted in February 2022 by the Society for Human Resource Management, about one-quarter of firms are already using automation or artificial intelligence (AI) to assist with hiring. Even more so (42%), this is true of companies with 5,000 or more workers. These businesses employ AI software to do things like review resumes, match them with open positions, answer candidate queries, and do evaluations.

However, authorities and lawmakers have raised concerns about the broad adoption of these technologies due to the possibility of discrimination and prejudice. In 2021, a research indicated that AI-enabled anti-Black prejudice in hiring, and rumours of a 2018 Amazon recruiting engine that was cancelled because it “did not like women” have persisted for years.

As a result, in November 2021, the New York City Council approved the Automated Employment Decision Tool ordinance by a vote of 38 to 4. For the purposes of this legislation, “any computational process derived from machine learning, statistical modelling, data analytics, or artificial intelligence” is defined as “any process that issues simplified output, including a score, classification, or recommendation, and substantially assists humans in making employment decisions.”

According to Gesser, some confusion was cleared up by the draught guidelines that came out in September. He said that this was because “they restricted the definition of what defines AI.” As the authors put it, “[The AI] must significantly aid or replace the discretionary decision-making. Not nearly enough weight should be given to it if it is only one of several sources considered. It must serve as the deciding factor.”

Furthermore, the restrictions restricted the law’s applicability to more complicated models. Therefore, “to the degree that it’s simply a simple algorithm that evaluates some things, until it transforms it into like a score or conducts like some complicated analysis, it doesn’t count,” he added.

Auditing for bias is difficult

The new rule mandates that businesses evaluate the potential discriminatory effects of automated hiring choice tools based on factors such as race, ethnicity, and gender. However, as Gesser pointed out, conducting a thorough audit of AI tools for bias is no simple operation, since it necessitates extensive analysis and access to a vast amount of data.

He also noted that it is not obvious if an employer may depend on a developer’s third-party audit because the employer could not have access to the technology that would allow them to execute the audit. One issue is that this information is typically supplied by applicants on their own will, therefore many firms lack a comprehensive collection.

He said that the data may give an inaccurate portrayal of the company’s racial, cultural, and gender diversity. For instance, if you can only choose between male and female for your gender, then those who don’t comply to those binary norms have no way to express themselves.

There will be further explanations in the future

Gesser, who accurately anticipated that there would be a delay in the enforcement period, stated, “I expect there is going to be additional advice.”

Some businesses will conduct the audit themselves if they have the resources to do so, while others will rely on the audit conducted by their suppliers. However, as Gesser pointed out, “it is not apparent to me what compliance is meant to look like and what is adequate.”

He noted that this sort of control of AI is not out of the ordinary. “It’s so novel, there isn’t much of a track record to study,” he added. Furthermore, unlike AI in lending, where there is a limited set of allowed criteria and a lengthy history of utilising models, AI regulation in the hiring process is “extremely problematic.”

When it comes to finding a job, no two are same. “I’ve learned that every candidate is different,” he remarked. “It’s just a lot more work to figure out what’s prejudiced,” she said.

You never want perfection to be the enemy of good, as Gesser put it. In other words, the goal of certain AI recruitment tools is to lessen prejudice, allowing them to assess a bigger pool of candidates than would be possible with human assessment alone.

However, “regulators believe there is a risk that these technologies might be used incorrectly,” he added, “either purposefully or unwittingly.” To that end, we must ensure that everyone is acting responsibly.

Implications for Comprehensive AI Regulation

A number of states in the United States have already established legislation pertaining to artificial intelligence, and the European Union is now working on more comprehensive AI regulation.

Gesser posits that the “risk-based regulatory regime” vs. “rights-based productivity regime” is a common point of contention in discussions about the future of AI regulation. This is because the New York statute is “basically a rights-based system — anybody who uses the tool is subject to the exact same audit obligation,” as he put it. However, the EU AI Act is working to establish a risk-based approach to deal with the most dire potential effects of AI.

Hence, “it’s about realising that there are going to be certain low-risk use cases that don’t require a big load of regulation,” he explained.

Gesser speculated that AI legislation will follow the same path as privacy law, with a sweeping European law taking effect and giving way to a patchwork of state and industry regulations. As a result, “U.S. corporations will complain that there is this hodgepodge of legislation and that it is too divided,” he said. “Congress will be under intense pressure to pass a comprehensive artificial intelligence bill.”

Gesser suggests launching an internal governance and compliance programme regardless of the type of upcoming AI regulation.

“Whether it’s the New York law or EU law or some other, AI regulation is coming and it’s going to be really messy,” he said. “Every company has to go through its own journey towards what works for them — to balance the upside of the value of AI against the regulatory and reputational risks that come with it.”