Predictable noise: the hidden flaw in machine judgement

If it is cloudy outside, a college admissions officer might care more about your grades; if it is sunny, your sports and clubs might suddenly look far more appealing. Human judgment fluctuates with mood and weather. Economists call this noise.

For years, management gurus suggested an easy out: replace fallible humans with algorithms. As the Harvard Business Review once argued, the key advantage of algorithms is that they are noise-free. Feed a formula the same input, and you always get the same output. But researcher Prakhar Ganesh argues that this premise is completely wrong. Machine judgment has not eliminated noise at all; it has merely baked random interference into what he calls predictable noise.

The illusion of cold, hard math

Why do people assume AI is noise-free? Because deployed machine learning models are essentially fixed mathematical formulas. If an automated system rejects your college application today, it will reject the exact same profile tomorrow. That deterministic behavior feels reassuringly consistent.

Yet the training process behind that formula is anything but clean. Ganesh uses a simple guessing game to prove the point: two players, Alice and Bob, are asked to guess a hidden number using identical rules and the exact same hints. Even though their algorithm is identical, they end up with completely different final guesses simply because of the random initial guess they started with.

AI systems work the exact same way. They rely on various forms of randomness during training, from weight initialization to optimization paths. Once training stops, the model becomes deterministic, but the noise from that random learning path is now frozen inside the formula for good.

This is fundamentally different from bias. Bias is systematic and usually explainable, like an interviewer having a personal preference for a certain school. The randomness frozen inside an AI, however, is completely meaningless.

What happens when algorithms pick students

Deploying these systems to run high-stakes selections like university admissions exposes sharp practical flaws.

AI might beat humans on average statistical metrics, but that tells us very little about its judgment on a single real-world human being. Grades and past records are incomplete indicators, and humans regularly break past trajectories for better or worse.

If colleges hand admissions over to AI, Ganesh points out two immediate problems. First, similar schools running similar models will all extend offers to the exact same cluster of candidates, leaving lecture halls full of empty seats when those students pick just one campus. Second, applicants will be judged entirely on past records without any allowance for unpredictable growth.

Why it matters

For builders and enterprise leaders, this argument pulls the rug out from under the lazy assumption that mathematical consistency equals fairness. If your organization is rushing to replace human screening with AI under the guise of removing human noise, you need to pause. You are not eliminating noise; you are just converting human mood swings into frozen, invisible mathematical artifacts.

Across the industry, this changes how we think about algorithmic decision-making. We think tech advocates have gotten away with equating repeatable outputs with objective judgment for far too long. A moody human reviewer is frustrating, but human variance occasionally offers a second chance to non-traditional applicants. An algorithm with predictable noise offers no grace, just a locked door disguised as statistical truth.