Resumen: 

 The literature initiated by Green and Laffont (1986) studies signalling games with hard evidence. Evidence is modeled by restricting the message set of the sender, depending on his type. Glazer and Rubinstein (2004,2006) and Sher (2011) show that, when the sender's utility function is type independent, there is no advantage for the receiver in having commitment power. In this paper, I argue that this way of modelling evidence implicitly assumes it to be 100% accurate. I develop a model with imperfect evidence and show that being able to commit has value for the receiver, unless the evidence is perfectly accurate.