Written by Ransom Boynton, Senior Associate
I recently rewatched 2001: A Space Odyssey, directed by Stanley Kubrick, based on an AI-generated recommendation from a streaming platform. The irony is not lost on me, given what follows.
“Good morning, Dave.”
For most employers, the voice of artificial intelligence is not quite that ominous. But we have seen this plot before. Tools that start out helpful have a way of becoming something else by the end of the movie. That got me thinking: as AI systems start playing a bigger role in hiring, scheduling, and promotion decisions, can AI cause potential liability for unlawful discrimination?
Artificial intelligence is already part of everyday employment decisions. Not just in Silicon Valley, but on factory floors and in service operations across California. Resume screeners, scheduling tools, and promotion algorithms are becoming standard tools for businesses trying to run more efficiently.
But as these tools spread, the question gets more practical: what happens if the system produces a discriminatory result, even if no one intended it? Can an employer be on the hook for what the algorithm does? In other words, who is responsible for HAL’s decisions?
The real issue is not whether AI can create disparities. It can. Your streaming recommendations are proof enough. The harder question, and one courts are only starting to grapple with, is whether those disparities turn into liability under the Fair Employment and Housing Act (FEHA).
Two realistic scenarios might show where this is heading.
Scenario One: A Hiring Algorithm Filters Out “Unreliable” Workers
A mid-size manufacturer rolls out a third-party hiring tool to screen applicants for warehouse and assembly roles. The software scores candidates based on things like job stability, employment gaps, and commute distance. HR only looks at applicants who meet a certain score.
In other words, HAL is not running the company, but it is helping decide who gets through the door.
After a year, a pattern starts to show up: older applicants and people with prior medical leave histories are being screened out at higher rates. A rejected applicant files suit, alleging age and disability discrimination under FEHA.
At first glance, this looks like a straightforward discrimination case. But things get more complicated when the screening decision comes from an algorithm instead of a person.
California courts recognize that even neutral policies can create liability if they disproportionately affect a protected group. (Guz v. Bechtel National, Inc. (2000) 24 Cal.4th 317.) An algorithm fits comfortably into that category, and statistical disparities may be enough to get a plaintiff past the starting line.
But that is only step one.
An employer can still win by showing the criteria are job-related and tied to legitimate business needs. Courts have long recognized that businesses can rely on practical, real-world factors, even if those factors have uneven effects. (Harris v. City of Santa Monica (2013) 56 Cal.4th 203.)
Where these cases really get difficult is proving causation. The algorithm did not actually hire or reject anyone. It produced a score. People set the thresholds, reviewed the candidates, and made the final call.
HAL, it turns out, is still taking instructions.
At the end of the day, FEHA asks a simple question: was a protected characteristic a real reason for the decision? (Guz v. Bechtel National, Inc., supra, 24 Cal.4th 317.) When both human judgment and algorithmic input are involved, answering that question gets a lot messier.
Scenario Two: A Scheduling Tool Quietly Shapes Who Gets Ahead
Now consider a service-based company using software to assign shifts and flag employees for promotion. The system favors employees with broader availability, more overtime acceptance, and higher productivity.
No red eyes. No calm voice. Just a dashboard.
Over time, employees with caregiving obligations, disproportionately women, end up with fewer premium shifts and fewer promotion opportunities. One employee files suit for sex discrimination and failure to promote.
Again, the theory is likely disparate impact: a neutral system produces a gender-based disparity. That is enough to shift the burden.
But that is not the end of the story.
Availability for work is not a suspicious or illegitimate criterion under FEHA. In many industries, it is critical. Employers still have to cover shifts, meet client demands, and keep operations running smoothly. Courts recognize that business realities matter, even when they lead to uneven outcomes. (Guz v. Bechtel National, Inc., supra, 24 Cal.4th 317.)
The real question is what actually caused the disparity. Was it the system, or the employees’ own scheduling choices? If someone limits their availability for legitimate reasons and the system responds to that, proving causation becomes much harder.
Courts have made clear that plaintiffs cannot just point to statistics and call it a day. (Duran v. U.S. Bank National Assn. (2014) 59 Cal.4th 1.) That matters here, where the outcome may reflect individual choices as much as anything the system is doing.
The Real Litigation Battlegrounds
These cases are not going to turn on whether AI is “biased.” It is. The real fights are likely to center on three questions:
1. Who actually made the decision?
Was the algorithm making the call, or just offering input? The more human involvement there is, the harder it is to pin liability on the tool.
2. What actually caused the outcome?
It is not enough to show a disparity. Plaintiffs have to show the system caused it, not employee choices, business needs, or other factors. Courts have rejected theories that rely on broad statistics disconnected from how decisions are really made. (Serri v. Santa Clara University (2014) 226 Cal.App.4th 830.)
3. Can the employer justify the system?
Employers can still rely on legitimate, nondiscriminatory business criteria, even if those criteria produce uneven results, so long as they are not a cover for discrimination. (Guz v. Bechtel National, Inc., supra, 24 Cal.4th 317.)
An Emerging Area—Without Court Guidance
So far, California appellate courts have not squarely addressed whether AI-driven employment decisions can create FEHA liability. That will not last long.
As these tools become more common, lawsuits will follow. When they do, courts are not starting from scratch. They will lean on existing law around statistical proof, causation, and business justification, especially cases that push back on overreliance on generalized data. (Serri v. Santa Clara University, supra, 226 Cal.App.4th 830; Duran v. U.S. Bank National Assn., supra, 59 Cal.4th 1.)
In short, the law is already there. The facts are just getting far more complicated.
Discovery Will Shape These Cases
If these cases move forward, discovery is where things will get interesting. But not just because of what gets produced. Discovery will force the parties to answer a more basic question: who actually made the decision?
Plaintiffs are going to focus on the choices the employer made before the system ever produced a result. What factors were selected? What information was used? What assumptions went into the tool? The argument will be that the employer chose to rely on criteria that may have looked neutral on their face, but in practice lead to uneven results. In other words, the bias did not come out of nowhere. It came from the way the system was set up.
Employers, on the other hand, will be tempted to say: “the algorithm did it!”
That argument is not likely to go very far on its own. These systems do not operate in a vacuum. Someone chose the tool, decided what information it would use, and determined how its output would be applied. The system may generate a result, but it does so within the framework the employer put in place.
At some point, someone will inevitably ask for “the algorithm.” There isn’t one. That’s like asking for the source code to The Matrix. There is no single document that explains everything. What exists instead is a mix of data, settings, and assumptions that shape how the system works. That makes it harder to draw a straight line from any one decision to any one source.
Employers will still have to respond to discovery, while protecting sensitive information where appropriate. At the same time, plaintiffs will likely rely heavily on statistical analyses to try to show patterns in the outcomes. Courts, however, are unlikely to accept those analyses at face value without a clear explanation of how they connect to the actual decisions being made.
Which brings us back to the central question: when a system produces a discriminatory result, who made the decision, the algorithm or the employer?
The Takeaway
AI does not change the basic elements of a FEHA claim. Plaintiffs still have to prove causation. Employers can still justify their decisions. Courts will still expect real evidence.
What AI changes is how complicated those questions become. For employers, the risk is not that these tools exist. It is using them without understanding how they work, what they measure, and whether their results can be explained.
For litigators, these cases will likely turn on who made the decision and why. The algorithm may generate the outcome, but someone arguably chose how it works. And when that question matters most, “the algorithm did it” will not be enough.
If you’re not sure whether your current process creates exposure, now is the time to find out. Contact Koegle Law Group today to learn more.
[Contact Us] | [Schedule a Strategy Call] | [Subscribe to Our Newsletter]
Disclaimer: This article is intended for general informational purposes only and does not constitute legal advice. The information contained herein may not reflect the most current legal developments and is not guaranteed to be correct, complete, or up to date. Nothing in this article should be construed as creating an attorney-client relationship. Employers and HR professionals should consult with competent employment counsel regarding their specific facts and circumstances before taking any action.