Apart from the alerting function of alarm calls, selection may favor cues that help individuals to distinguish between reliable and unreliable callers. However, this mechanism for selective response to real and false alarms may act only if the individual characteristics of the call are stable at least for some time. Here we test this implicit assumption for the caller's reliability hypothesis, studying individuality of alarm calls in a colony of free-living, individually marked speckled ground squirrels (Spermophilus suslicus). We recorded each of 20 study animals 4 times during repeated captures when calling from a live trap toward a human, with spans of 1 day, 2 weeks, and 1 year from the 1st capture. Ten alarm call notes per animal per capture were analyzed. Individual alarm call notes showed high similarity within captures but differed strongly between captures. Both multivariate analysis of variance and discriminant function analysis showed that vocal individuality decreased rapidly with an increase of the time span between recordings. However, vocal individuality always remained higher than expected random value. Examination of our data suggests that alarm calls are unstable, which contradicts the caller reliability hypothesis, because its implicit assumption of stable individual identity is not fulfilled. However, short-term stability still may be sufficient to ensure short-term individual recognition between kin and neighbors. Also, even if the alarm calls change structurally, because group members meet up daily, they can update their knowledge of the call structure of individuals, and this would likely allow them to distinguish between reliable and unreliable individuals.