With the proposed method, you can fit arbitrarily closely to the data, so you can get your likelihood as good as you like, still using a single parameter. So yo

With the proposed method, you can fit arbitrarily closely to the data, so you can get your likelihood as good as you like, still using a single parameter. So you get good k(=1) and good likelihood, thus good AIC. The likelihood does not need to dominate the other term, it just has to be as good as the likelihood of the model you are comparing to.