A data lake stores large volumes of raw data in its original format, which makes it cheap and flexible to load. Without metadata, clear owners, and rules about what enters and how it is checked, the lake fills with files that nobody can explain. Duplicate copies, unknown sources, outdated extracts, and columns with unclear meaning pile up. At that point people call it a data swamp: the data is still there, but finding the right table and trusting it takes longer than starting over.
For example, an analyst looking for customer churn data may find five similar tables with different row counts and no note on which one finance uses. Teams recover with a data catalog, ownership for each dataset, and lineage. For AI work, the cost shows up in a specific way: a model or agent may read a table that looks correct but no longer means what the task assumes.