Colorado passes first in the nation attempt to assess AI discrimination

While antidiscrimination laws exist, those who study AI discrimination say it’s a different beast, which the U.S. is already behind in regulating.

The first attempts to regulate artificial intelligence programs that play a hidden role in hiring, housing and medical decisions for millions of Americans are facing pressure from all sides and floundering in statehouses nationwide.

Only one of seven bills aimed at preventing AI’s penchant to discriminate when making consequential decisions — including who gets hired, money for a home or medical care — has passed. Colorado Gov. Jared Polis hesitantly signed the bill on Friday.

Colorado’s bill and those that faltered in Washington, Connecticut and elsewhere faced battles on many fronts, including between civil rights groups and the tech industry, and lawmakers wary of wading into a technology few yet understand and governors worried about being the odd-state-out and spooking AI startups.

Polis signed Colorado’s bill “with reservations,” saying in an statement he was wary of regulations dousing AI innovation. The bill has a two-year runway and can be altered before it becomes law.

“I encourage (lawmakers) to significantly improve on this before it takes effect,” Polis wrote.

Colorado’s proposal, along with six sister bills, are complex, but will broadly require companies to assess the risk of discrimination from their AI and inform customers when AI was used to help make a consequential decision for them.

The bills are separate from more than 400 AI-related bills that have been debated this year. Most are aimed at slices of AI, such as the use of deepfakes in elections or to make pornography.

The seven bills are more ambitious, applying across major industries and targeting discrimination, one of the technology’s most perverse and complex problems.

“We actually have no visibility into the algorithms that are used, whether they work or they don’t, or whether we’re discriminated against,” said Rumman Chowdhury, AI envoy for the U.S. Department of State who previously led Twitter’s AI ethics team.

While anti-discrimination laws are already on the books, those who study AI discrimination say it’s a different beast, which the U.S. is already behind in regulating.

“The computers are making biased decisions at scale,” said Christine Webber, a civil rights attorney who has worked on class action lawsuits over discrimination including against Boeing and Tyson Foods. Now, Webber is nearing final approval on one of the first-in-the-nation settlements in a class action over AI discrimination.

“Not, I should say, that the old systems were perfectly free from bias either,” said Webber. But “any one person could only look at so many resumes in the day. So you could only make so many biased decisions in one day and the computer can do it rapidly across large numbers of people.”

When you apply for a job, an apartment or a home loan, there’s a good chance AI is assessing your application: sending it up the line, assigning it a score or filtering it out. It’s estimated as many as 83% of employers use algorithms to help in hiring, according to the Equal Employment Opportunity Commission.

AI itself doesn’t know what to look for in a job application, so it’s taught based on past resumes. The historical data that is used to train algorithms can smuggle in bias.

Amazon, for example, worked on a hiring algorithm that was trained on old resumes: largely male applicants. When assessing new applicants, it downgraded resumes with the word “women’s” or that listed women’s colleges because they were not represented in the historical data — the resumes — it had learned from. The project was scuttled.

Webber’s class action lawsuit alleges that an AI system that scores rental applications disproportionately assigned lower scores to Black or Hispanic applicants. A study found that an AI system built to assess medical needs passed over Black patients for special care.

Studies and lawsuits have allowed a glimpse under the hood of AI systems, but most algorithms remain veiled. Americans are largely unaware that these tools are being used, polling from Pew Research shows. Companies generally aren’t required to explicitly disclose that an AI was used.

“Just pulling back the curtain so that we can see who’s really doing the assessing and what tool is being used is a huge, huge first step,” said Webber. “The existing laws don’t work if we can’t get at least some basic information.”

That’s what Colorado’s bill, along with another surviving bill in California, are trying to change. The bills, including a flagship proposal in Connecticut that was killed under opposition from the governor, are largely similar.

Colorado’s bill will require companies using AI to help make consequential decisions for Americans to annually assess their AI for potential bias; implement an oversight program within the company; tell the state attorney general if discrimination was found; and inform to customers when an AI was used to help make a decision for them, including an option to appeal.

Labor unions and academics fear that a reliance on companies overseeing themselves means it’ll be hard to proactively address discrimination in an AI system before it’s done damage. Companies are fearful that forced transparency could reveal trade secrets, including in potential litigation, in this hyper-competitive new field.

AI companies also pushed for, and generally received, a provision that only allows the attorney general, not citizens, to file lawsuits under the new law. Enforcement details have been left up to the attorney general.

While larger AI companies have more or less been on board with these proposals, a group of smaller Colorado-based AI companies said the requirements might be manageable by behemoth AI companies, but not by budding startups.

“We are in a brand new era of primordial soup,” said Logan Cerkovnik, founder of Thumper.ai, referring to the field of AI. “Having overly restrictive legislation that forces us into definitions and restricts our use of technology while this is forming is just going to be detrimental to innovation.”

All agreed, along with many AI companies, that what’s formally called “algorithmic discrimination” is critical to tackle. But they said the bill as written falls short of that goal. Instead, they proposed beefing up existing anti-discrimination laws.

Chowdhury worries that lawsuits are too costly and time consuming to be an effective enforcement tool, and laws should instead go beyond what even Colorado is proposing. Instead, Chowdhury and academics have proposed accredited, independent organization that can explicitly test for potential bias in an AI algorithm.

“You can understand and deal with a single person who is discriminatory or biased,” said Chowdhury. “What do we do when it’s embedded into the entire institution?”