An AI model is not programmed with rules. It is trained: shown enormous amounts of text, images and other material, until it learns the patterns in them. That material is the training data, and it shapes everything about how the model behaves, what it knows, what it is good at, and what blind spots it has.
Two practical consequences follow. First, a model only knows what existed in its training data, up to a certain date. That date is called the knowledge cutoff, and it is why an AI can be confidently out of date about recent events unless it can search the web. Second, patterns in the data, including skewed or unfair ones, come along for the ride (see bias).
Why you keep hearing it
Because training data is where many of AI's controversies live: copyright disputes about what was used, debates about consent, and questions about why models behave the way they do.
What it means for you
When an AI gives you outdated information, think knowledge cutoff, not stupidity. And when your brand's own material becomes part of what AI systems learn from and cite, clarity in what you publish starts to matter in a whole new way.