|
|
| |
 |
By Aaron Mak |
Presented by Alliance for a Better Future |
With help from Hassan Ali Kanu
|
Supreme Court Justice Ruth Bader Ginsburg inspired a project to use AI for identifying discriminatory laws. | Charles Dharapak/AP Photo |
Researchers at the Stanford Institute for Human-Centered AI and the university’s RegLab published a study last week demonstrating how large language models can help identify discriminatory local laws. The AI system helped human reviewers identify thousands of local statutes that could violate anti-discrimination laws by treating people differently based on race, gender and citizenship. The jurisdictions with those statutes contain millions of people. DFD spoke with HAI fellow Daniel E. Ho, who helped conduct the study. This interview has been edited for length and clarity. Why did AI seem suited for this task? The reason why an AI system really helps to solve a problem like this is the sheer scale of local laws. In 1950, civil rights icon Pauli Murray published a 700-page volume documenting segregation laws at the state level. Thurgood Marshall called it the “bible” of Brown v. Board of Education. A couple of decades later, Ruth Bader Ginsburg, when she was a law professor, hired a small army of Columbia law students to sift through the U.S. Code to write a report called “Sex Bias in the U.S. Code.” Trying to do something similar at the local code level is a really daunting task. In San Francisco alone the municipal code and its associated resolutions run some 16 million words. That's why we felt it would be really useful to develop an [AI] system that really tackled this. In what ways did the system align with human analysis of the laws, and diverge? We validated these kinds of assessments in two ways. We curated a validation set [using] an amicus brief by historians compiling discriminatory statutory provisions, the Pauli Murray compilation of Jim Crow laws and a compilation having to do with gay marriage litigation. We asked how often the system is able to track and identify those provisions. The accuracy there was quite high: 98%. We also had human reviewers, alongside with the AI system, go through and try to prioritize amongst the shortlisted set of provisions how strong the claim is that this violates anti-discrimination law. There's some room for improvement. Once we've already zoomed in on [certain] provisions, the AI system can in some instances be overly eager. There are instances where the law countenances classification based on gender — for instance, [if you] have separate facilities for a county jail, there's typically a rationale there that courts will credit. The AI system was more eager than some of the human reviewers in prioritizing that as an anti-discrimination law violation.
|
| |
POLITICO's Decoded Summit The future of AI policy is being defined in power centers from coast to coast. Join POLITICO's Decoded Summit – Washington, D.C. on Sept. 16 for conversations on AI regulation, global competition, enterprise adoption, and more.
Register for the livestream. |
| |
| |
The study developed rating systems that the AI could use to evaluate whether a law might be discriminatory. How did you think about translating the legal analyses that courts apply into a system that AI could use? Courts, of course, don't use a rating system. What you would see is much more the doctrinal analysis that identifies: Is the law classifying based on a protected attribute? Depending on the protected attribute, what level of scrutiny should be attached to this? And then we're weighing the state's justification against these other interests and seeing whether it passes a kind of legal bar. That was the structure of the doctrinal inquiry that informed the design of this AI system. This was not a simple thing where we just prompt a model with respect to a provision. The AI system really was trying to track that doctrinal inquiry. Why did you only focus on laws that explicitly treat certain protected classes differently, rather than ones that are written neutrally but have discriminatory impacts? Do you think AI could also identify discriminatory laws that seem neutral on their face? Disparate treatment is a little bit of an easier question that is more within reach of the current capabilities of AI systems, because there's a lot that you can do just based on the surface of the statute to track whether or not the statute is classifying by a protected attribute. Disparate impact is context-based inquiry. There's a robust causality requirement. So if you have a facially neutral law that might have an adverse effect on the group, you really have to show that this policy is causing these disparities.
|
| |
A message from Alliance for a Better Future: There is a better future. A future where AI strengthens families, empowers workers, and reflects the values that make America great. AI is not just another technological advancement. It’s a civilization shift. American greatness has grown from goodness. By ensuring AI meets our highest ideals, America can lead the world with AI that is not just powerful, but trusted. Learn More. |
| |
|
|
| SpaceXAI sues man accused of using Grok to ‘nudify’ minors |
|
Elon Musk’s SpaceXAI has sued another man who’s been criminally charged with sex crimes after allegedly using the company's Grok chatbot to “nudify” pictures of minors — marking at least the third such civil lawsuit by the artificial intelligence company. SpaceXAI last week sued Harry Tiffany IV, alleging the Bucks County, Pennsylvania man illegally breached its usage terms when he used the application to generate sexualized images and videos of children. Tiffany is facing up to 100 years in prison on federal criminal charges for producing and possessing sexual abuse material depicting four girls. Authorities began investigating him after receiving legally mandated tips from SpaceXAI. The company’s suit was filed before Judge Reed O’Connor, a George W. Bush appointee in the Northern District of Texas known to reliably favor conservative and Republican-backed policies. The lawsuits appear to be part of a strategy to eventually transfer into O’Connor’s courtroom a growing wave of suits against the company that are being filed by alleged victims of the three men. SpaceXAI did not respond to requests for comment. An attorney representing Tiffany also did not immediately respond to a request for comment. SpaceXAI is currently facing at least six lawsuits involving more than 15 identified plaintiffs over the use of Grok to digitally undress adults and minors, including proposed class actions. The company has sought to transfer a number of those cases to the Northern District of Texas. SpaceXAI’s civil suits seek to hold the accused child sex abusers responsible for “all reasonable expenses” xAI incurs defending itself against suits their victims file against the company, including damages for “reputational harm.”
|
| |
A message from Alliance for a Better Future:  
|
| |
|
|
| OpenAI backs biodefense bills |
|
OpenAI is supporting a slate of bills the company says will help prevent artificial intelligence models from supercharging the threat of biological weapons and synthetic viruses. The bills aim to improve access to data needed to build AI defenses against synthetic and naturally occurring biothreats, and establish standards for the safe development of those tools. The company exclusively told POLITICO it is issuing new endorsements for three bipartisan bills surrounding the issue: the Web of Biological Data Act, the AI-Ready Bio-Data Standards Act and the SCALE Biology Act. The first two proposals were introduced in both the House and Senate, the latter in the House. Richard Johnson, OpenAI’s national security risk mitigation lead, told POLITICO that data and standards addressed in the bills are essential for developing better methods of evaluating how increasingly powerful models may impact biosecurity. “We need to constantly be able to measure capabilities and risk, and then be able to compare where we are in that stage to everything else that's happening in the industry,” he said.
|
| |
A message from Alliance for a Better Future: Alliance for a Better Future – AI the American Way means AI by the people, of the people, and for the people. Families in Charge means AI that serves families, rather than replacing parents; parents should be in the driver’s seat. Opportunity for Everyone means technology that lifts every American; shared prosperity is a dream worth fighting for. The future is not yet written, but the window is narrow. The stakes are lasting. Learn More. |
| |
|
|
|
|
|
Screenshot/X |
|
| |
Spencer Cox on America’s political divide In the latest episode of On the Road, Jonathan Martin sits down with Utah Gov. Spencer Cox to discuss his new book, Off Ramp, the future of the Republican Party, immigration reform and what he says it will take to rebuild trust and community in American life.
Watch now. |
| |
| |
|
|
|
|
Stay in touch with the whole team: Gabby Miller (gmiller@politico.com); Aaron Mak (amak@politico.com); Bob King (bking@politico.com); Nate Robson (nrobson@politico.com); John Hewitt Jones (jhewittjones@politico.com). |
| |
Follow us on X
|
| |
Follow us
|
| |
Posts
0 Comments