If the durations of the blocks/events are varying, then the AM2 pair of regressors will not be orthogonal.

Whether they "should" be orthogonal is an interesting question. Under the hypothesis of white noise -- as assumed in the statistics generated by 3dDeconvolve -- orthogonal regressors would have independent statistical tests, making the calculation of the significance of a conjunction test (t1 significant AND t2 significant) relatively straightforward.

However, under the more correct assumption of non-white noise -- as in 3dREMLfit -- then two separate t tests are not necessarily independent even if the regressors are orthogonal -- which complicates the computation of the significance of a conjunction test. And if the regressors are not orthogonal, then even with the white noise assumption the t tests are not independent.

Beyond this, I'm not really ready to say. If you want to perform such a conjunction test, accurate significance testing will require more thought. If you aren't interested in a conjunction test, then I'm not sure why you asked the question.