Data Quality
How trustworthy the data used for training actually is
- Data quality asks how trustworthy the data you've gathered actually is, a different question from how much of it there is.
- It splits into several angles: are the values correct, is anything missing, is it stale, is it consistent, is it duplicated, each checked separately.
- A flaw left in the data hardens into a habit of the model through training. Changing the training method afterward rarely undoes it.
- Whether data counts as good depends on what it's for. Data that's plenty for one use is unusable for another.
- Checking it isn't a one-time thing. As long as data keeps coming in, quality keeps shifting too.
Contents
1The analogy
Before a summer water park opens for the day, it tests the water first. A sample gets scooped into a container to check how cloudy it looks, whether anything's settled at the bottom, and whether the disinfectant level sits within range. However good the slides are, if the water fails that check, the park doesn't open that day.
The data gathered to build an AI is that water. Whatever sits underneath, cloudy, stays cloudy no matter what gets stacked on top of it. Data quality is scooping up a container of that water and taking a close look.
Just as a water check never stops at one item, checking data covers several angles too. Whether values were recorded properly, whether any fields are empty, when the water was drawn, and whether yesterday's reading matches today's, each angle gets checked on its own.
2In detail
The checks split into five angles
The first angle checks whether values are correct. A weight entered into a height field, or a date with the month and day swapped, gets caught here. Data typed in by hand tends to carry the most of this kind of flaw.
The second checks for anything missing. A blank here and there can be filled in, or that one row dropped, but a whole field left empty means giving up on using that field at all.
The third checks how old the data is. You can't judge today's water park by water drawn years ago, and if the world has moved on while the data hasn't, the model just keeps describing a world that no longer exists.
The fourth and fifth check whether things line up and whether anything's duplicated. Is the same thing recorded differently in two places? Has the same record been swept in several times over? Let the same record in three times, and its voice alone gets three times louder.
A flaw flows straight downstream
Pour in cloudy water, and even a great filter later can't pull out dirt that's already gone in. Data works the same way. A model trained on a wrongly attached answer learns that mistake as a rule, and from then on, answers by that rule.
What makes it worse is that the flaw barely shows. If only a small slice of the data is wrong, the score on the surface still looks fine. It only misfires once something resembling that small slice shows up for real. To whoever's using it, that looks like an error with no explainable cause.
That's why time spent looking the data over before training even starts is never wasted. Find the flaw after training finishes, and fixing the data means running the whole thing again from scratch, a far more expensive fix.
What counts as good data depends on the use
The water that fills a pool and the water people drink are held to different standards. Data is the same. Estimating a corner store's daily sales roughly is fine with a ledger missing a few entries, but automating inventory reorders can't tolerate a single missing one.
That's why deciding what this data is for has to come before checking its quality. Once the use is set, how strictly to check each angle follows from it. Try to hold every angle to the strictest standard at once, and the effort never ends, while the angle that actually mattered gets missed.
The same dataset often turns out usable in one place and unusable in another.
The checking has to keep running
Water changes condition in a single day. Data does too. The format of incoming data can quietly shift one day, or whoever's been recording values can switch units, and everything coming in from that point gets thrown off.
That's why anywhere data keeps arriving, the checking keeps running too. How much is blank, whether values stay within their usual range, whether the daily volume looks like usual, measure these every day, and set an alert for whenever something looks off. A person reviews it by eye only the first time; a machine takes over the scooping after that.
When a service that's been running fine suddenly starts giving strange answers, checking where the data comes in is the right first move, before taking the model apart.
3More precisely
Data quality usually gets measured across dimensions like accuracy, completeness, consistency, timeliness, uniqueness, and validity. Each dimension is measured differently. Some, like the share of blank fields, are just a matter of counting, while others, like whether a value matches reality, can only be checked against some outside reference. If the data carries an answer key, that answer key's own quality needs checking separately too, often gauged by how much several people's independently attached answers agree with each other.
The analogy breaks in one place. Filtered water turns clear again, but a value entered wrong in the data often can't be recovered at all. What's deleted stays deleted, and the work has to continue with what's left. Water testing also has one shared standard to pass, but data has no such common passing line. The standard has to be set for the job at hand, and setting that standard is itself half of quality management.
Quality and skew are separate problems too, worth noting. Data with not a single wrong value and no blanks can still lean heavily toward one group. Passing every quality check doesn't say who the data actually represents, that needs checking on its own.
4Try it yourself
5Common misconceptions
It's easy to think more data is always better, but actually adding to flawed data adds to the flaws too, and the result can end up worse than a smaller, clean dataset.
It's easy to assume good training papers over quality problems, but actually the model learns a wrongly entered value as a rule, and training it longer only sets that rule deeper.
It's easy to think cleaning it up once means it's done, but actually quality starts shifting again the moment new data arrives, so the checking has to keep running.
7One-line summary
In shortData quality means scooping up a sample before pouring data into training and checking it angle by angle, and whatever flaw slips through unfiltered stays on as a habit of the model.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02