LLM-Augmented Discrete Choice

Where Artificial Choices Need Anchors

Large language models can generate many plausible travel choices, but plausible is not the same as representative of real travelers. In the same choice situation, a human sample and an artificial sample may favor different options. If those differences are ignored, adding more synthetic responses can make a demand model more confident about the wrong behavior.

A human-versus-LLM mismatch appears first, followed by a correction learned from a smaller set of observed human choices. Once the artificial choice pattern has been adjusted toward that human anchor, synthetic cases can help fill gaps where real observations are scarce. The correction is deliberately imperfect: the point is to reduce systematic bias, not to make the two curves identical. This approach matters when collecting human choice data is expensive, while generating hypothetical scenarios is easy. The underlying question is how to use artificial data for scale without allowing it to overwrite the behavioral signal that only people can provide.