Lag Selection and AIC

[b.auer]

Hello experts,

I am currently trying to estimate an equation of the type yt = a + b1yt-1 + ... + bpyt-p + c1xt-1 + ... + cqxt-q + et for a Granger causality test. I would like to select the optimal lag lengths p and q with AIC. Comparing different lag combinations I find that AIC is becoming lower and lower with increasing lags. But that is obviously not the idea of the measure because i can add plenty of lags without penalty. So whats wrong with my procedure?

Regards
Benjamin