AI models learn from what humans have written and photographed (see training data). All the imbalances in that material come along: which faces appear in stock photos, which names get associated with which jobs, which accents and cultures are treated as default. The model absorbs those patterns and reproduces them, at scale, without anyone deciding it should.
Bias shows up in small, cumulative ways. Ask an image model for "a CEO" and watch who it draws. Ask for "a beautiful home" and watch which continent's architecture appears. None of it is malicious. All of it compounds.
Why you keep hearing it
Because it is one of the central ethical debates in AI, and because regulators, researchers and journalists keep finding examples. AI companies actively work to reduce it, with real but partial success.
What it means for you
For marketers this is unusually concrete, because we manufacture imagery and language all day. If AI generates your visuals and copy, its defaults become your brand's defaults unless someone is paying attention. Who appears in your generated images? Whose names show up in your examples? The tools will not watch this for you. Someone on the team has to, and it is a genuinely good use of a marketer's judgment.