2 weeks ago
This April, nan US Department of Justice joined Elon Musk’s xAI in suing nan authorities of Colorado to termination its AI anti-discrimination law.
When nan national authorities sides pinch a billionaire against a authorities trying to protect its residents from AI discrimination, that’s not only a Colorado story. That’s everyone’s story.
The justness department’s suit is portion of a coordinated national effort to reframe AI user protections arsenic ideological overreach. In July 2025, Donald Trump signed an executive bid connected “preventing woke AI”, equating bias mitigation measures to a leftist “woke” schedule that suppresses free reside and truth. The national National Policy Framework launched successful March included a push to pre-empt authorities laws connected AI, pinch Colorado’s rule targeted. The justness department’s involution successful Colorado marks nan first clip nan national authorities has sought to intervene successful a suit challenging a authorities AI law.
The AI measure successful reference, Senate Bill 205, was astir protecting group from favoritism successful high-risk AI systems making consequential decisions: who gets hired, who gets housing, who receives healthcare.
The original measure required bias audits, effect assessments and disclosure, requirements that drew pushback from nan business community. Colorado heard that and responded. Before xAI ever revenge its lawsuit, nan authorities revised those requirements pinch a reduced transparency model successful mid-March.
This benignant of measure matters. A 2019 study published successful Science recovered that a wide utilized healthcare algorithm deployed crossed US hospitals assigned Black patients half nan attraction of arsenic sick achromatic patients because it utilized healthcare costs arsenic a proxy for wellness needs, embedding existing inequities into objective decisions. Removing healthcare costs eliminated group bias successful nan exemplary and helped hospitals present much effective care.
But xAI based on nan Colorado bill would unit nan institution to beforehand nan state’s “ideological views connected various matters, group justness successful particular”, successful its Grok chatbot. Brianna Titone, a Colorado authorities typical and 1 of nan bill’s lead sponsors, disagrees. She captured nan misleading quality of xAI’s lawsuit: “SB 205 is astir consequential decisions … We’re not restricting speech. Our measure does not opportunity that Grok still can’t beryllium a dick.”
The justness department, meanwhile, called nan measure “state-mandated discrimination” that “obligates AI developers and deployers to discriminate”, an bonzer characterization of a rule that requires companies to cheque whether their systems are producing unlawfully discriminatory outcomes and hole them if so.
The justness department’s’s statement rested connected nan mendacious presumption that AI systems usage “neutral criteria”. Under this logic, altering models that usage neutral criteria messes pinch their objectivity, constituting discrimination. But arsenic nan infirmary study shows, criteria that look neutral tin nutrient worse outcomes for immoderate populations. This inclination of mediocre proxies harming marginalized groups has been shown successful various studies, specified arsenic for welfare allocation and college admissions. Bias besides enters erstwhile models person little training information connected immoderate groups, causing worse capacity for those populations – arsenic documented successful facial nickname research and my ain investigation connected large connection models. When companies improved practice successful their training data, capacity discrepancies astir disappeared.
Then there’s nan “winning nan AI race” framing, reflected successful nan justness section lawsuit. The title technologists talk astir and that nan US authorities refers to is nan pursuit of artificial wide intelligence and beating China – a frontier dominated by a fistful of well-capitalized firm labs. Rules barring AI from discriminating successful hiring and healthcare are abstracted from that race. Conflating nan 2 is misleading.
The only mainstream nationalist outlet to screen this communicative frankincense far, the Wall Street Journal, focused connected effects nan rule could person connected businesses and “killing nan entrepreneurial spirit”, a sentiment echoed successful nan justness section lawsuit. That statement doesn’t clasp up, either. No companies are cited arsenic leaving Colorado because of this regulation. The state’s governor, Jared Polis, states that “far more” firms are moving to Colorado than leaving. The 1 actual illustration offered successful nan Journal: Palantir cited nan measure arsenic a imaginable load successful a Securities and Exchange Commission (SEC) filing. But nan institution did not say this was nan logic it near Colorado for Florida. On nan other extremity of nan spectrum, companies including Microsoft have flagged bias and favoritism successful AI arsenic a worldly risk.
Small businesses person morganatic concerns, to which nan authorities responded. Still, utilizing AI to make consequential decisions astir who gets hired aliases receives healthcare is simply a choice. If a business wants to create aliases usage AI successful hiring aliases healthcare but doesn’t person nan bandwidth to cheque whether it discriminates, past they don’t person nan bandwidth to usage it responsibly. It’s worthy noting: AI devices that don’t activity for everyone present worse outcomes, create liability and erode trust. The healthcare algorithm that was fixed didn’t sacrifice efficiency; it sewage amended astatine its job.
But nan unit run worked. On 14 May, Polis signed into law a caller bill (SB 189), which repeals and replaces SB 205. Gone are nan requirements for companies to proactively measure their high-risk AI systems for bias, behaviour yearly reviews, study discovered favoritism to nan state, and usage reasonable attraction to reside known and foreseeable harms. What remains is simply a request for AI developers to stock method archiving (on intended uses, limitations, training data) pinch deployers but not pinch nan public. Consumers get a announcement that AI was progressive successful a determination and nan correct to petition quality review, which astir consumers will ne'er cognize to invoke. It’s amended than nothing, but falls short of ensuring meaningful, proactive accountability for high-risk systems.
If nan justness section tin subordinate a billionaire’s suit to termination a authorities rule meant to protect consumers and support transparency successful high-risk areas, what does that mean for different states trying to shield residents from AI harms successful nan absence of national regulation? The connection is clear: don’t try.
That’s not pro-America. AI is an incredibly powerful technology, and its imaginable is top erstwhile it useful for each of us. We merit nan transparency and protection to get there.
-
Dr Genevieve Smith is simply a postdoctoral investigation chap astatine Stanford University, laminitis of nan Responsible AI Initiative astatine nan UC Berkeley AI Research Lab and a personnel of master module astatine nan UC Berkeley Haas School of Business
