Say your regression result for a variable in levels and a version in logs get opposite-sign, statistically significant results. What would you check second?
(I assume you would check first what I did: whether there are zeroes causing mischief. There are not.)