Odd results with Sickle FASTQ trimming

[id0]

I was experimenting with Sickle (windowed adaptive trimming tool for FASTQ files using quality). It seems like a useful tool. I used it on some 2x150 Illumina data in paired-end mode and then aligned to mouse genome (mm9) using BWA. Everything seemed to work fine. When I later did recalibration with GATK, I looked at the produced quality charts and noticed something odd. Without trimming, the two reads look very similar in terms of quality. With sickle trimming, there is a very noticeable difference between the two reads. This seems very odd to me. Why would this happen?

Here is an example for one sample: