Recent genome-wide analyses have detected numerous cancer-specific alternative splicing (AS) events. Whether transcripts containing cancer-specific AS events are likely to be translated into functional proteins or simply reflect noisy splicing, thereby determining their clinical relevance, is not known. Here we show that consistent with a noisy-splicing model, cancer-specific AS events generally tend to be rare, containing more premature stop codons and have less identifiable functional domains in both the human and mouse. Interestingly, common cancer-derived AS transcripts from tumour suppressor and oncogenes show marked changes in premature stop-codon frequency; with tumour suppressor genes exhibiting increased levels of premature stop codons whereas oncogenes have the opposite pattern. We conclude that tumours tend to have faithful oncogene splicing and a higher incidence of premature stop codons among tumour suppressor and cancer-specific splice variants showing the importance of considering splicing noise when analysing cancer-specific splicing changes.