That would count as harm and be disallowed by the current policy.

But a military application of using GPT to identify and filter misinformation would not be harm, and would have been prevented by the previous policy prohibiting any military use, but would be allowed under the current policy.

Of course, it gets murkier if the military application of identifying misinformation later ends up with a drone strike on the misinformer. In theory they could submit a usage description of "identify misinformation" which appears to do no harm, but then take the identifications to cause harm.

Which is part of why a broad ban on military use may have been more prudent than a ban only on harmful military usage.