Remember when Amazon had to scrap its AI recruiting tool because it hated women?
Humans are biased, sure. But human bias is usually messy, inconsistent, and limited to the room the interviewer is sitting in. AI bias is systemic and scalable. According to MIT Tech Review, AI is more likely than humans to form these biases during the hiring process. The problem is that these models aren’t discovering “merit”—they are just echoing the historical preferences of the people who wrote the original training data.
If your previous ten successful hires all went to the same three universities and played lacrosse, the model doesn’t see “talent.” It sees a pattern. It then aggressively filters out anyone who doesn’t fit that specific mold (probably because it’s easier than talking to people). It is basically a high-speed mirror of our own worst tendencies, just wrapped in a veneer of mathematical objectivity.
The industry loves to talk about “alignment” and “fairness constraints,” but that is mostly theater. You cannot simply tell a model to “ignore gender” if gender is correlated with a thousand other variables in the dataset, like gaps in employment or specific extracurriculars. Trying to fix a biased model after the fact is like a chef using a pre-made sauce and pretending it’s a reduction; you can add a bit of salt at the end, but the base is already ruined.
Do we really want a black box deciding who gets a paycheck? The reality is that these tools don’t find the best candidate; they find the candidate who is best at gaming the prompt. It is a race to the bottom where the most “optimizable” resume wins, not the most competent engineer.
It’s a disaster waiting to happen.
Then there is the weird side of the house: weather data sabotage. When the data used for climate modeling gets messed with, the results aren’t just “off”—they are dangerous. Whether it is intentional manipulation or just systemic failure, the reliability of the inputs is the only thing that actually matters. If the training set is poisoned, the output is a hallucination with a forecast icon attached to it.
It is a reminder that the “data is truth” mantra is a lie. Data is just a record of what happened, and records can be faked, deleted, or skewed. If we cannot trust the sensors providing the raw numbers, the most sophisticated transformer architecture in the world is just a very expensive random number generator. Or maybe the hardware is the problem—see below.
The current state of AI hiring has created a bizarre kind of friction. We have these “efficient” systems that still result in a candidate waiting three weeks for a bot to send a generic rejection email. It is the worst of both worlds: the coldness of automation and the latency of a broken bureaucracy.
We have reached the point where the “standard” resume is a liability because it is designed for a machine that doesn’t understand nuance. If you are applying for a dev role, you are not fighting other devs anymore; you are fighting a weights-and-biases configuration that might hate the word “passionate” or love the word “synergy” for reasons no one can explain.
By Q4, we’ll see a wave of “anti-AI” resume formats designed specifically to break these scrapers or trick them into passing candidates through to a human. We will see a return to the portfolio and the referral because the middleman—the AI screener—has become a bottleneck of incompetence.
