For a stretch this year I did work that sounds mind-numbing and mostly was: reading newspaper articles, one after another, and coding each one into a dataset by hand — recording its themes, its tone, its framing, according to a careful set of rules, hundreds and hundreds of times over. It's the kind of task people assume you'd want to automate away as fast as possible. And yet those months taught me more about the real nature of data than any tool or tutorial has, precisely because I had to make every judgement myself, one article at a time. Here's what stuck.
Every dataset is made by someone, on purpose
The biggest shift was visceral rather than intellectual. When you build a dataset by hand — when you personally decide, for article after article, "this counts as that category, this one doesn't" — you can never again look at a dataset as something that simply exists, neutral and given. You feel, in your hands, that every row was made by a person making a decision. The dataset didn't fall from the sky. Someone built it, choice by choice.
That sounds obvious written down. It is not obvious until you've done it. Most people treat data as found, like a rock — objective, just there. Coding a thousand articles teaches you in your bones that data is manufactured, like a chair — and that means it carries the fingerprints, the assumptions, and the occasional bad day of whoever manufactured it.
The categories are decisions, not discoveries
The second lesson was about the categories themselves. Before you code anything, someone has to decide what the categories are — what themes to track, what counts as which tone. And those decisions are not neutral. They shape everything that follows. A theme you didn't create a category for is a theme your dataset will be permanently blind to, no matter how carefully you code. The questions the data can eventually answer were quietly decided at the very start, by whoever drew up the categories — long before a single article was read.
I found that genuinely unsettling in a useful way. It means a dataset doesn't just contain answers; it contains, baked in, a set of assumptions about which questions were worth asking in the first place. Read a finished dataset and you're seeing the world through categories someone else chose. Understand those choices and you understand the data. Ignore them and the data will mislead you while looking perfectly objective.
The boredom was doing something
Here's the part I didn't expect: the tedium itself was load-bearing. Reading every article slowly, by hand, meant I couldn't hide from the ambiguous ones — the article that half-fit two categories, the tone that was technically neutral but clearly wasn't. An automated process would have forced those into a box and moved on, silently. Doing it by hand, I had to sit with each hard case and make a real decision, and in doing so I learned exactly where my own categories were fuzzy, where the rules broke down, where the "clean" data was actually a pile of judgement calls I'd rather not think about.
That's knowledge you can only get by grinding through it. Automate too early and you never meet the edge cases — you just launder them into the dataset unexamined, and they come back to haunt whoever trusts the numbers later.
Data isn't found, it's made. And you only really understand a dataset once you've felt the weight of the thousand small human decisions that built it.
The article that broke my categories
One article sticks in my memory because it quietly demolished my confidence in my own scheme. On the surface it was straightforward reporting — neutral, factual, exactly what my "neutral tone" category was built for. But read closely, its framing did something the words never said outright: through which facts it led with, whose voice got the last word, and what it left unmentioned, it took a clear position while keeping a perfectly neutral surface.
My categories had no box for that. "Tone" said neutral; the framing said anything but. And I had to make a call, knowing that whichever box I ticked would flatten something real into something simpler — and that the "clean" data point I produced would carry none of my hesitation into the dataset.
Multiply that by a thousand articles and you understand something permanent about data: the tidy final value hides every hard call that went into it. The dataset looks certain. The person who built it remembers exactly how uncertain it was.
Why I'm oddly grateful for it
I wouldn't volunteer to do it again. But I'm genuinely glad I did it once, because it gave me an instinct I now can't switch off: whenever I meet a dataset, some part of me immediately asks who made this, what did they decide, and what did they leave out? That question — the refusal to treat data as neutral and given — feels like it might be one of the most important things a person can carry into any work with data.
The tools will keep getting better at handling scale. What they'll never do is make the human decisions underneath disappear. Coding a thousand articles by hand taught me those decisions are always there, whether or not anyone admits it — and that reading data well means reading the decisions, not just the numbers.