Remote Work's Hidden Perk: Bias Disappears When You Can't See People

A study published in Economics of Education Review found that the well-documented 'beauty premium' in student grading vanished when university classes moved online. In face-to-face settings, more attractive students received measurably higher grades for the same quality of work. Remove the face, and the effect dropped to zero.

The study leveraged the natural experiment of COVID-era remote learning to compare grading patterns for the same instructors teaching the same courses, before and after the shift to online delivery. The attractiveness premium — previously estimated at roughly 0.2 standard deviations in grade outcomes — simply evaporated.

This should matter to anyone who's ever sat on the other side of a hiring panel, a code review, or a promotion committee.

The implications for tech are direct. The industry spent 2020-2022 discovering that remote work didn't tank productivity, then spent 2023-2025 dragging people back to offices anyway. The usual arguments center on collaboration, culture, and serendipitous hallway conversations. What rarely gets discussed is what remote work *removes*: the full catalog of appearance-based cognitive biases that humans apply unconsciously in every in-person interaction.

Think about what this means for code review. A pull request reviewed asynchronously on GitHub is evaluated on the code. The same change presented in a meeting room carries the invisible weight of how the presenter looks, sounds, and carries themselves. There's a reason 'blind auditions' increased female musicians in top orchestras by 25-46% — removing visual information forces evaluation on merit.

This isn't an argument that remote work is universally better. It's an argument that in-person work has costs we don't account for because they're invisible to the people who benefit from them. Every RTO mandate is, among other things, a decision to re-enable a set of biases that disproportionately affect people who don't match the pattern of what a 'senior engineer' or 'engineering leader' is supposed to look like.

The study also raises an uncomfortable question about AI-assisted hiring and evaluation. If we're training models on historical performance data collected in face-to-face environments, we're baking in beauty premium effects. The bias doesn't need to be in the algorithm — it's already in the training labels.

For engineering managers: this is one more data point suggesting that asynchronous, text-first communication isn't just a productivity preference. It's a debiasing mechanism. The next time someone argues that 'you lose something' when you move interactions online, it's worth asking exactly what's being lost — and who benefits from keeping it.