One sees this in math research. The model reports it has proved X. In fact it has given an erroneous numerical check of Y in a few atypical cases.

What makes math approachable is that the context is so well delimited (semantically) that one can guide the model with adequate correction.