Machine learning systems are trained on the record of what humans have written and done. Whatever else they are, they are a reflection — and reflections are useful precisely because they are unflattering.
Two temptations should be declined at the outset. The first is to treat these systems as oracles, ascribing to statistical fluency a wisdom it does not possess. The second is to treat them as demons, which flatters us by implying the trouble arrived from outside. Neither is accurate and both prevent thinking.
What is accurate is more interesting. A model trained on human output inherits human patterns — including the biases, the evasions and the characteristic ways we deceive ourselves. When it produces something ugly, the finding is not chiefly about the machine. And when it produces something clarifying, that too came from somewhere.
The distinction I find most useful is between capability and direction. These systems are extraordinary at capability: pattern-finding, translation, synthesis, prediction, the compression of vast material into usable form. They have no direction of their own. Direction is supplied by whoever specifies the objective — and the objective is a moral choice wearing technical clothing.