The algorithm didn't know it was racist. It just knew the data Easy to understand, harder to ignore..
That's the line you hear over and over in tech circles — a kind of secular prayer. Day to day, *The model is neutral. So naturally, the math doesn't lie. We just fed it the world as it is.
But here's what that prayer leaves out: the world as it is was built on purpose. Redlining wasn't an accident. Still, predictive policing didn't emerge from a vacuum. The credit scores, the hiring filters, the risk assessment tools that decide who gets bail and who stays locked up — they all inherit the architecture of a system that sorted human worth by race long before anyone wrote a line of code.
Ruha Benjamin's Race After Technology: Abolitionist Tools for the New Jim Code isn't just a critique. It's a field guide for seeing the machinery underneath the marketing. And once you see it, you can't unsee it Worth knowing..
What Is the New Jim Code
Benjamin coins the term deliberately. It echoes Michelle Alexander's The New Jim Crow — but the target isn't mass incarceration alone. It's the coded layer now sitting on top of it. And beside it. And increasingly, inside it Practical, not theoretical..
The New Jim Code isn't one system. A hiring algorithm that downgrades resumes from historically Black colleges. So a healthcare triage tool that assigns lower risk scores to Black patients with the same symptoms as white patients. It's a pattern: automated decision-making that reproduces racial hierarchy while claiming objectivity. A facial recognition system that misidentifies darker faces at rates ten times higher than lighter ones.
This is where a lot of people lose the thread The details matter here..
These aren't bugs. They're features of the training data — which is to say, features of the history that produced the data It's one of those things that adds up. Nothing fancy..
The Myth of Neutrality
Tech loves the word "neutral." It's a shield. If the output is biased, the defense is always the same: *we didn't program bias in. The data spoke.
But data doesn't speak. Now, it's collected, curated, labeled, cleaned, weighted, and deployed by people inside institutions with budgets and incentives and blind spots. In real terms, a dataset of "successful employees" at a company that's been 90% white for decades? Which means that's not a neutral sample of talent. That's a record of who got hired, promoted, and retained under the old rules.
The moment you train a model on that record and ask it to predict "future success," you're not predicting potential. You're predicting continuity.
Benjamin calls this "coded inequity" — not because the code is hateful, but because it encodes the past as prophecy.
Why It Matters Now
Five years ago, this conversation lived in academic journals and activist newsletters. Today, it's in city council meetings, shareholder resolutions, and the terms of service you didn't read That's the whole idea..
Because the New Jim Code isn't theoretical anymore. It decides:
- Whether your mortgage application gets flagged for "manual review"
- Whether your child's school gets labeled "failing" and slated for closure
- Whether your neighborhood gets more police patrols or more pothole repairs
- Whether your resume reaches a human or gets auto-rejected at 2 a.m.
And the reach is expanding. Child welfare algorithms. Practically speaking, medicaid fraud detection. Tenant screening scores. Which means gig worker deactivation systems. Each one promises efficiency. Each one delivers disparity.
The Stakes Are Material
This isn't about "representation" in the abstract. It's about material harm.
A 2019 study found that a widely used hospital algorithm systematically prioritized healthier white patients over sicker Black patients for extra care resources. The algorithm didn't know that. But Black patients generate less spending at the same level of illness because of structural access barriers. So naturally, the variable? Healthcare spending — which the algorithm treated as a proxy for need. It just optimized for the number it had Worth keeping that in mind. That's the whole idea..
People died because of that proxy.
In Los Angeles, a predictive policing tool called PredPol directed officers to neighborhoods already over-policed. So a feedback loop dressed in math. More patrols meant more arrests meant more data meant more patrols. The department eventually dropped it — but not before years of compounding harm.
These systems don't just reflect inequality. They automate it. They scale it. They make it faster, cheaper, and harder to challenge — because "the computer said so" sounds final in a way "the officer decided" never did.
How It Works: The Architecture of Automated Inequality
If you want to spot the New Jim Code in the wild, look for these patterns. They show up again and again, across domains, across vendors, across good intentions.
1. Proxy Variables That Carry History
Race is rarely an explicit input. That would be illegal. But zip code? Credit history? Education? Employment gaps? Prior arrests? These are proxies — and in a segregated society, they carry racial signal with near-perfect fidelity Practical, not theoretical..
A hiring filter that screens for "continuous employment" penalizes caregivers, formerly incarcerated people, anyone who navigated a labor market that never welcomed them equally. In practice, the filter doesn't know race. It doesn't need to That alone is useful..
2. Feedback Loops Disguised as Learning
Predictive policing is the textbook case. But the same logic lives in child welfare "risk scores," in fraud detection for public benefits, in school "early warning" systems.
The system predicts risk → resources target the predicted area → more incidents get recorded (because more eyes are watching) → the model retrains on the new data → risk prediction goes up.
The loop tightens. Even so, the disparity deepens. And every iteration looks like "validation" — the model was right, look at the numbers.
3. Optimization for the Wrong Thing
Efficiency. In practice, cost savings. So naturally, throughput. Accuracy — defined narrowly, usually as "agreement with past decisions.
But past decisions were the problem That's the whole idea..
When a pretrial risk tool optimizes for "predicting failure to appear," it's optimizing against a history where Black defendants were held on higher bail, had less access to transportation, faced more rigid work schedules. The tool learns that Black defendants are "higher risk" — and then that prediction justifies holding them on higher bail.
Honestly, this part trips people up more than it should.
The circle closes. The tool didn't predict the future. It enforced the past.
4. Opacity as Policy
Proprietary algorithms. "Explainability" dashboards that show feature importance without context. Trade secret claims. Vendors who sell to public agencies but refuse to let auditors see the model.
This isn't accidental. If you can't see the logic, you can't challenge the outcome. In practice, opacity protects power. You can't demand due process from a black box.
Benjamin calls this "algorithmic governance by contract" — public authority outsourced to private code, shielded by NDAs And that's really what it comes down to. But it adds up..
Common Mistakes: What Most People Get Wrong
"Just Remove the Bias from the Data"
You can't. The data is the bias — or at least, the sediment of it. There's no "clean" dataset waiting underneath. The only way to get unbiased training data for a just society is to build a just society first.
Debiasing techniques help at the margins. They don't fix the structural logic.
"Diverse Teams Will Fix It"
Necessary. Not sufficient.
A diverse team inside the same incentive structure — ship fast, optimize for engagement, maximize shareholder value — will still build harmful systems. Representation without power redistribution changes the faces in the room, not the architecture of the product.
"Transparency Solves Everything"
Knowing that a system is biased doesn't automatically give you put to work to change it. Worth adding: public housing authorities know their tenant screening algorithms disproportionately reject Black applicants. They keep using them — because the vendor is cheaper than manual review, because the contract is signed, because no one with budget authority feels the pain.
Transparency is a tool. Not a solution.
"AI Is the Problem"
So, the New Jim Code predates deep learning. It predates "AI" as a buzzword. Credit scoring.
