- The LAPD spent a decade using an AI that predicted crime with less than 1% accuracy.
Then they wondered why it kept sending cops to Black and Latino neighborhoods.
This is what happens when you build AI on "dirty data" and zero accountability.
🧵 - Geolitica (formerly PredPol) used earthquake aftershock algorithms to predict crime.
Read that again.
They took seismology models and pointed them at human beings. The training data? Historical arrest records already saturated with racial bias. - The result was a textbook feedback loop:
Algorithm flags neighborhood → more cops deployed → more arrests for minor offenses → new arrests fed back into system → neighborhood flagged again.
Bias didn't just persist. It compounded. - Chicago went further. Their "Strategic Subject List" flagged 400,000+ people as potential threats.
56% of Black men aged 20-29 in the city were on it.
57% of "priority targets" had zero violent arrests.
The algorithm criminalized being young, Black, and alive in Chicago. - Both systems shared the same fatal flaw: epistemic opacity.
Even the police departments using them couldn't explain how predictions were made. Proprietary black boxes. No audits. No accountability.
For nearly a decade.
Now here's what should terrify every enterprise leader: - Your company is probably making the same mistake.
Simple LLM wrappers — GPT with a UI on top — inherit identical structural risks. No domain reasoning. No bias checks. No governance. - We tested this directly. Naive LLM agents hit ~51% accuracy on security vulnerability triage. Basically a coin flip.
Deep AI systems with composable agents, specialized reasoning layers, and domain knowledge? ~89%.
Architecture matters more than the model. - The fix isn't abandoning AI. It's maturing it.
That means mathematical fairness metrics baked into development — demographic parity, equalized odds — not bolted on after launch.
Pre-processing. In-processing. Post-processing. Bias mitigation at every stage. - It means explainability isn't optional. If your AI can't show which features drove a decision — and prove those features are valid — you don't have a tool.
You have a liability. - 40+ U.S. cities have now banned or restricted predictive policing tech. The White House mandated impact assessments for rights-impacting AI in 2024.
The regulatory wave isn't coming. It's here.
And it's expanding far beyond law enforcement. - Here's what we keep asking ourselves:
If your company's AI makes a high-stakes decision tomorrow — hiring, lending, triage — could you explain exactly why it chose what it chose?
And would that explanation survive an audit?
#ResponsibleAI #AIGovernance - We wrote the full analysis — predictive policing failures, feedback loops, enterprise parallels, and our governance framework — here:
https://veriprajna.com/whitepapers/architectures-of-trust-beyond-superficial-ai-algorithmic-integrity