In 2018, Amazon quietly abandoned a recruiting tool that had been in development for years. The AI system, designed to review resumes and identify top candidates, had learned something its creators never intended: it preferred men. Resumes that included the word "women's"—as in "women's chess club captain"—were automatically downgraded. Graduates of all-women's colleges were penalized. The algorithm had looked at Amazon's historical hiring data, noticed that most successful engineers were male, and concluded that being male was a qualification.

This wasn't a bug. It was the system working exactly as designed—learning patterns from data and optimizing for them. The trolley problem here isn't a split-second decision about who to save. It's a systematic choice about whose futures get derailed, made thousands of times, invisibly, at scale.

The Efficiency Track vs. The Fairness Track

Traditional hiring is slow, expensive, and subjective. A human recruiter might spend two minutes scanning a resume, influenced by unconscious biases, fatigue, and whether they've had their coffee. AI promises something better: objective evaluation at scale, processing thousands of applications in seconds, identifying patterns humans might miss.

But here's the trolley problem: you can optimize for efficiency or fairness, but not perfectly for both. An algorithm trained on historical data will reproduce historical inequities. An algorithm designed to ignore protected characteristics (race, gender, age) will find proxies—zip codes, college names, gaps in employment history. The tracks diverge, and someone has to choose which way to pull the lever.

Amazon chose efficiency first, then discovered they'd built a system that systematically discriminated. They tried to fix it by removing explicitly gendered terms, but the bias ran deeper—into writing style, word choice, even the structure of sentences. They eventually scrapped the entire system. But most companies don't discover their biases so clearly, and many don't scrap their systems when they do.

The Illusion of Objectivity

The appeal of algorithmic hiring is the promise of objectivity. Numbers don't have prejudices. Code doesn't care about your race or gender. The algorithm just optimizes for the best candidate.

Except it doesn't. It optimizes for whatever "best" meant in the training data. If your historical "best" employees were mostly white men from elite universities, that's what the algorithm learns to prefer. If women took career breaks for childcare and were subsequently less likely to be promoted, the algorithm learns that employment gaps predict poor performance. If minority candidates were historically screened out at higher rates, the algorithm learns to screen them out more efficiently.

This is the cruelest version of the trolley problem: the one where you don't realize you're pulling a lever at all. Traditional hiring discrimination is visible—you can challenge a biased interviewer, sue for discrimination, demand transparency. Algorithmic discrimination is invisible, justified by "objective" scores, protected as proprietary trade secrets. The people on the track don't even know the trolley is coming.

Who Gets Hurt by Optimization?

HireVue, a company that analyzes video interviews using AI, claimed its system could predict job performance by analyzing facial expressions, word choice, and speaking patterns. The system was used by over 100 major companies to screen millions of candidates. In 2021, after years of criticism, HireVue stopped using facial analysis—but only after countless candidates had been rejected by an algorithm analyzing their smiles.

LinkedIn's algorithm, designed to suggest job opportunities, was found to show fewer high-paying jobs to women than to men with identical qualifications. The system wasn't explicitly programmed to discriminate—it was optimizing for engagement, and historical data showed women were less likely to apply for certain jobs, so the algorithm stopped showing them those opportunities. A feedback loop disguised as optimization.

Pymetrics uses neuroscience games to assess candidates—pattern matching, risk tolerance, attention span. The company claims its approach is bias-free because it doesn't look at resumes at all. But the games are calibrated against "successful" employees at each company. If your successful employees are demographically homogeneous, your "objective" games will select for more of the same. Different tracks, same destination.

The Trade-Off We're Not Discussing

Here's what makes this a true trolley problem: there are legitimate reasons to use hiring algorithms. Manual resume screening is expensive, inconsistent, and demonstrably biased. Humans are terrible at predicting job performance from interviews—we prefer candidates who look like us, talk like us, went to schools we recognize. Structured algorithms, properly designed, can reduce some forms of bias.

But "properly designed" requires acknowledging trade-offs most companies won't make. You could optimize for demographic parity—ensuring your algorithm selects candidates in proportion to the applicant pool. But that might mean accepting lower scores on your performance metrics (which themselves encode historical bias). You could require algorithmic transparency, publishing exactly how decisions are made. But that might reveal that your "objective" system is making subjective value judgments about what skills matter.

You could allow candidates to see their algorithmic scores and challenge them. But that would slow down the process, defeating the efficiency gains that justified the algorithm in the first place. You could regularly audit for disparate impact. But that costs money, and if you find bias, you're legally obligated to fix it or stop using the system.

Most companies choose the track that looks like optimization: deploy the algorithm, trust the scores, hire efficiently. The people who don't get interviews, who are screened out by invisible criteria, who never know an algorithm decided they weren't worth a human's time—they're on the other track. And unlike the trolley problem, we don't even count them.

The Tracks Lead to Different Futures

The trolley problem asks about immediate harm: five people versus one. Hiring algorithms create diverging futures. Get screened out by Amazon's algorithm, and you don't just lose one job—you lose the career trajectory, the network, the experience that would have led to the next opportunity. Multiply that by thousands of candidates, hundreds of companies, millions of decisions.

The algorithm doesn't just choose who gets hired today. It chooses whose skills are valued, whose backgrounds are legitimate, whose career paths are recognized as valid. It encodes one vision of merit and makes it seem inevitable, objective, natural. The people on the other track don't just lose a job. They lose the future where their skills mattered.

And here's the deepest problem: we're not even asking who should pull the lever. We're letting the algorithm pull it, based on patterns from a past we claim to want to move beyond, optimizing for an efficiency that benefits companies while distributing harm to candidates who never see it coming.

The trolley is already moving. The tracks have already diverged. The only question is whether we'll keep pretending the algorithm is driving itself.