Suppose I have extremely good statistical evidence that a random member of the student body at some ordinary college (I am not talking here of some underworld Thieves' University) has committed a crime of type C, and that the statistical evidence is probabilistically as strong as paradigmatic cases of "evidence that excludes a reasomable doubt". (Imagine, for instance, that the prevalence of criminal copyright infringement among college students was a couple of orders of magnitude higher than during the peak of the Napster era.[note 1]) Would it be permissible to convict a random member of the student solely on those grounds?
The idea would, I think, make most people uncomfortable. But why? It can't just be "because he might be innocent". For the same is true of someone convicted on paradigmatic evidence that excludes a reasonable doubt, and the chance of error in the latter case is no less than in the purely statistical case by my stipulations.
My suggestion for what is wrong here is that we want punishments to be dependent—say, counterfactually—on the crime. Even though we can be quite certain that Sally who is a student at the college has committed a crime of the relevant sort, we want it to be the case that had she not committed it, she would not be convicted.
Now, the above story might seem far-fetched. Here is a version of the story that isn't quite so far-fetched. The likelihood that a contract killer reports all his income to the tax authorities is vanishingly small. But we should not tack on a conviction for tax fraud unless we actually check the tax records.
Here is an interesting consequence of the counterfactual dependence thesis. Suppose that responsibility is compatible with determinism, and two days ago we were able to tell with certainty from Jones' brain scan and other data that a day later Jones would torture his uncle, leaving no physical marks, and then would erase all marks of the torture from the uncle's mind. Can we convict Jones of the torture? The uncle denies having been tortured, but that is exactly what we would expect. We have excellent evidence of Jones' guilt. But our evidence does not counterfactually depend on Jones' choice, and hence if we convict him, his punishment will not counterfactually depend on his choice. And hence, even though we have evidence of his guilt that excludes reasonable doubt, we are not permitted to convict Jones if I am right.