Signal transduction pathways translate the cellular context into context-dependent expression of genes. In response to extracellular stimuli, proteins have to be up-regulated quickly and reliably. However, rapid and reliable control of target genes by a signalling pathway faces two major challenges. Firstly, swift changes in protein levels require short-lived proteins, which is not resource-optimal. Secondly, gene expression is an intrinsically noisy process, and fluctuations in the protein numbers are likely to reduce the functionality of the proteins. Mammalian signalling pathways frequently induce the transcription of their own inhibitors, resulting in negative feedback regulation. However, the functional role of these transcriptional feedbacks in mammalian signal transduction is unclear. Here, we analyse a mathematical model of a prototypical signalling pathway, the MAPK cascade, in order to investigate how transcriptional negative feedbacks may help to overcome the challenge of fast and reliable gene induction. It is shown that a transcriptional negative feedback helps to decouple protein stability and response times, thus allowing for swift up-regulation even of long-lived proteins. Furthermore, transcriptional negative feedbacks filter out the extrinsic component of gene expression noise, which dominates the uncertainty in gene expression in mammalian cells, thus making gene expression more reliable. Model analysis predicts that both goals can be achieved if (i) proteins and mRNAs of the feedback regulators and (ii) mRNAs of all targets are short-lived. These predictions are confirmed by large-scale measurements of mRNA and protein half-lives. Therefore, the design of the mammalian signal transduction network with its rapid feedback inhibition allows for swift and reliable target gene expression.