I have never met anyone who thinks of themselves as unfair. It's one of those beliefs that is close to universal and almost never checked, which is usually a sign that something is off. We treat fairness as a default setting, something that comes pre-installed with being a decent person. Then a situation shows up where being fair means ruling against your own kid, or your own team, or your own bonus, and the default turns out to be something else entirely.
The stranger test proves nothing
The definition I keep coming back to is simple: fairness is applying the same standard regardless of who is affected, and especially when the person affected is someone you love, someone you can't stand, or you. Being fair to a stranger in a dispute you have no stake in is easy. Anyone manages that. The question is whether the standard survives contact with your loyalties. Can you give credit to the colleague who irritates you? Concede that the person who wronged you has a point? Split the thing so that you personally get less? Fairness that has never cost anything has never been tested.
Most unfairness isn't malice, either. It's a tilt toward our own that runs below awareness, which is what makes it so hard to fix by intention alone. The best evidence I know for this is what happened when American orchestras began auditioning musicians behind a screen. Nobody on those committees thought they were discriminating; they were professionals judging sound. But once the screen hid who was playing, women started advancing at noticeably higher rates. Goldin and Rouse put numbers on it in a famous 2000 study, and although the statistics have been argued over since, the underlying point has held up: people whose entire job was listening could not fully separate the playing from the player. If they couldn't feel their tilt, I doubt I can feel mine.
What it isn't
Fairness is not treating everyone identically. Giving the same to someone in need and someone in plenty is uniform, not fair, and confusing the two produces a lot of confident bad judgment. It isn't niceness either; keeping everyone happy is often precisely how the unfair outcome gets made.
And it isn't rule-following. 'Those are the rules' can be an honest answer, but it can also be a place to hide from the harder question of whether the rule produces a fair result in this particular case. Rules are tools for fairness. Sometimes they're its enemy, and telling those situations apart is most of the actual work.
Ordinary versions
The versions that matter are mostly not institutional. Think of a parent refereeing a fight between their own child and the neighbor's kid, feeling the pull to side with their own, and ruling on what actually happened instead. I think that's one of the more consequential things a parent ever does, precisely because it looks so small. The child walks away knowing something about the world, that truth can outrank team, and my hunch is that this particular lesson can't really be taught, only shown. The manager who promotes the person they don't warm to, because that person earned it, belongs to the same family, and everyone in the building can tell whether it happens. Then there's the hardest version, being fair to someone who was unfair to you, which I would like to claim I'm good at and am not.
The machine learned the tilt from us
We are steadily handing these calls to software: who gets the loan, the interview, the parole hearing, the apartment. Part of the appeal is the promise of neutrality. The machine has no nephew applying, no colleague it resents. The promise isn't entirely empty, and I'll get to that. But the record so far is sobering. In 2016, ProPublica published its 'Machine Bias' investigation of COMPAS, a risk score used in American courts, and reported that Black defendants who did not go on to reoffend had been flagged as high risk at roughly twice the rate of white defendants who didn't. The vendor disputed the analysis using a different definition of fairness, and researchers later showed that the two definitions cannot both be satisfied at once when underlying rates differ. Which is its own lesson: the algorithm didn't remove the value judgment, it forced a judgment that had always been there out into the open, where nobody could agree on it.
Or take Britain's exam fiasco of 2020. With exams cancelled by the pandemic, England's regulator used an algorithm to standardize teacher-assessed grades, and it downgraded nearly forty percent of A-level results, falling hardest on strong students from large state schools while small classes, disproportionately at private schools, kept their teachers' optimism. The government scrapped the whole scheme within days, after teenagers were in the streets protesting a formula. What strikes me about that episode is not that the formula was skewed. It's that the skew arrived wrapped in a sheen of objectivity, so it took public protest, rather than one official's private discomfort, to undo it.
Here is where I have to concede something. The obvious moral of those stories, keep a human in the loop, is weaker than it sounds. Decades of research comparing expert intuition with simple statistical rules find that the rules match or beat the experts depressingly often, and human overrides frequently reintroduce the exact tilt the formula lacked. So I don't think the answer is human judgment instead of the machine. I think it's a person who actually cares about fairness standing where the machine's output lands: someone able to look at a clean, confident score and notice that it is wrong about the particular human in front of them. That noticing can't be automated, because it isn't a computation. It's caring about the answer.
A rough self-test
When we write scenarios for Agonora, the drafts that fail are always the ones where fairness is free, where the right answer costs the character nothing. Those measure reading comprehension, not fairness. The ones that work put a thumb on the scale: your friend, your bonus, your side. Which suggests a self-test I offer without much comfort. Try to remember the last time being fair actually cost you something real, an advantage surrendered, a point conceded to an opponent, a ruling against your own. I found the exercise harder than I expected. And I'm fairly sure the uncomfortable reading of a blank there is not that the moments never came. It's that they came, and went, and didn't register.