The Record
Tracking the risks, rulings, and realities of human oversight in AI.
The Human Appeal Exemption Is Not a Way Around the Review
California's ADMT rules let a business avoid the consumer opt-out right only if the appeal is handled by someone who can actually revisit the decision. A template appeal process doesn't solve the review problem. It preserves the evidence that no one reviewed the person the first time or the second.
Automation Bias Is a Design Problem, Not a Training Problem
A reviewer can be trained, qualified, and authorized and still defer to the model on almost every decision, because automation bias is a measured cognitive default that training doesn't fix. California's human-involvement standard depends on how the review is designed and recorded, not on how carefully reviewers are told to look.
Cigna, PxDx, and the 1.2-Second Review
Cigna's PxDx let medical directors deny more than 300,000 claims in two months at an average of 1.2 seconds each. The figure wasn't lost in an unrecoverable log; it sat in an internal scorecard, and it became the core of the case. PxDx isn't even AI, which is what makes it the cleanest illustration of what automated-decision liability actually turns on. Not the technology, but whether the individualized review the plan promised ever happened.
AI Liability Doesn't Follow the Org Chart
When Workday's model screens out applicants and UnitedHealth's model denies post-acute care, the liability surface spans the vendor who built the model and the deployer who recorded the decision. Workday is being sued as an agent of employers it never employed. UnitedHealth is liable for determinations its plan documents attributed to physicians. Existing coverage follows entity lines. The exposure doesn't.
Model Accuracy Is Not a Compliance Defense
A credit model that correctly denies a loan and a credit model that incorrectly denies a loan produce the same compliance question under California's ADMT rules. The regulation doesn't ask whether the output was right. It asks whether the human reviewer was real. Model accuracy makes that question harder to answer, not easier.
UnitedHealth, nH Predict, and What a 90% Reversal Rate Proves
When over 90% of AI-driven prior authorization denials are reversed on appeal, the appeal record is not evidence that the system works. It's retroactive proof that the original review didn't happen. The Lokken class action doesn't turn on whether nH Predict was wrong. It turns on what UnitedHealth's own plan documents promised a human reviewer would do.
Rite Aid, Target, and the Employee Review Problem in Retail AI
The FTC's action against Rite Aid wasn't primarily about whether the facial recognition technology was accurate. It was about what happened when employees received match alerts with no verification protocol, no oversight, and no documentation standard.
Prenuvo and the Documentation Problem in AI-Assisted Screening
A whole-body MRI that reported normal findings, a catastrophic stroke eight months later, and a liability record consisting entirely of a one-page report. The Prenuvo case isn't about whether the AI was wrong. It's about what's missing from the trail.
An Employee Monitoring Tool Doesn't Have to Make Decisions to Trigger California's ADMT Rules
Automated profiling based on systematic observation of employees can require a risk assessment even when no employment action follows. The trigger is the inference from observation, not the action taken on it.
Write Isolation Doesn't Prove the Human Was the Decision-Maker
Preventing the AI from writing a credit determination into the loan origination system doesn't prevent it from making the decision in every way the regulation cares about. The question isn't who held the pen. It's who chose the outcome.
A Good Model and a Rubber-Stamp Produce the Same Override Rate
A highly accurate credit AI model and an underwriter who upholds every denial without reading the application produce identical override metrics. California's ADMT rules are built to catch the second pattern, but the number alone can't distinguish between the two.
Human-in-the-Loop Is Not Enough Under California ADMT Rules
A recruiter click isn't enough to avoid California's ADMT rules. The reviewer must understand the output, check other relevant information about the candidate, and have real authority to change the decision.