Wrong and More Confident: A Field Experiment on Language Models Taking a Graduate Economics Exam

A red herring, an irrelevant passage added to a problem, makes a language model reason incorrectly and answer incorrectly far more often.

Yet the model still writes out a full explanation, and the answer it gives remains consistent with the steps it shows.

The red herring corrupts the reasoning, while leaving the explanation intact and coherent.

I show this on the Graduate Economic Reasoning Benchmark (GERB), sixty graduate-level microeconomics problems, each a detailed setup with a verified answer and a step-by-step reference solution, presented with and without the red herring and answered by thirty-eight language models in a within-task $2\times2$ design.

The red herring lowers the probability of a correct answer by 12.3 percentage points, about a quarter of what these models answer correctly, and thirty-seven of the thirty-eight are less accurate under it.

Reasoning capability confers no protection.

The effect does not differ detectably between models that reason by default and models with no reasoning mode, nor between open-weight and closed-weight models, though open-weight models reach comparable accuracy at a substantially lower cost per correct answer.

The damage is largest on the problems the model rates as easy.

Worse, if anything the red herring leads a model to rate a problem as easier, not harder, than its clean version, even as it answers that problem wrong more often.

The clean problems are already hard, and on average the models answer fewer than three of every five correctly.

A wrong answer still comes with a full explanation, is coherent with its own reasoning, and the model stays confident, so the error cannot be caught without checking it against the verified answer.