There's a saying in data work that's been around long enough to become a cliché: garbage in, garbage out. It's a cliché because it's true, and it's especially true for AI.
A machine learning model learns from examples. Everything it knows about the world — every pattern it can recognize, every prediction it can make — comes from the data it was trained on. Which means the data isn't just an input. It's the source of the model's entire understanding of the problem. If that data is incomplete, mislabeled, biased, or just not representative of the situations the model will actually encounter, the model will be wrong in exactly those ways, confidently and at scale.
This is why the most experienced practitioners in AI spend a disproportionate amount of their time on data rather than on models. The model architecture matters. The training process matters. But neither of them can compensate for training data that doesn't reflect reality. A more sophisticated algorithm trained on bad data will almost always lose to a simpler algorithm trained on good data.
The two articles that follow are about the two main levers for improving data quality before training begins. Data preprocessing is the work of cleaning and preparing raw data — handling missing values, removing noise, normalizing formats, making the data consistent enough to learn from. Synthetic data generation is a different approach: when you don't have enough real data, or when the real data has gaps you can't fill, you generate artificial data that fills those gaps. Both are responses to the same underlying reality: the data you have is rarely the data you wish you had, and the gap between them is where a lot of AI projects succeed or fail.
The rest of this section gets into what happens after the data is ready. But data quality is the right place to start, because it's the constraint that everything else runs into first.
