Thank you for the encouragement, and I’m glad you’ve found value in my commentary.
… it’s also important to track which errors seem like part of a pattern of motivated error, and which seem to be mere mistakes. The former class seems much more dangerous to me, since such errors are correlated.
I agree with this as an object-level policy / approach, but I think not quite for the same reason as yours.
It seems to me that the line between “motivated error” and “mere mistake” is thin, and hard to locate, and possibly not actually existent. We humans are very good at self-deception, after all. Operating on the assumption that something can be identified as clearly being a “mere mistake” (or, conversely, as clearly being a “motivated error”) is dangerous.
That said, I think that there is clearly a spectrum, and I do endorse tracking at least roughly in which region of the spectrum any given case lies, because doing so creates some good incentives (i.e., it avoids disincentivizing post-hoc honesty). On the other hand, it also creates some bad incentives, e.g. the incentive for the sort of self-deception described above. Truthfully, I don’t know what the optimal approach is, here. Constant vigilance against any failures in this whole class is, however, warranted in any case.
Thank you for the encouragement, and I’m glad you’ve found value in my commentary.
I agree with this as an object-level policy / approach, but I think not quite for the same reason as yours.
It seems to me that the line between “motivated error” and “mere mistake” is thin, and hard to locate, and possibly not actually existent. We humans are very good at self-deception, after all. Operating on the assumption that something can be identified as clearly being a “mere mistake” (or, conversely, as clearly being a “motivated error”) is dangerous.
That said, I think that there is clearly a spectrum, and I do endorse tracking at least roughly in which region of the spectrum any given case lies, because doing so creates some good incentives (i.e., it avoids disincentivizing post-hoc honesty). On the other hand, it also creates some bad incentives, e.g. the incentive for the sort of self-deception described above. Truthfully, I don’t know what the optimal approach is, here. Constant vigilance against any failures in this whole class is, however, warranted in any case.