Classical confidence limits are compared to Bayesian error bounds by studying relevant examples. The performance of the two methods is investigated relative to the properties coherence, precision, bias, universality, simplicity. A proposal to define error limits in various cases is derived from the comparison. It is based on the likelihood function only and follows in most cases the general practice in high energy physics. Classical methods are discarded because they violate the likelihood principle, they can produce physically inconsistent results, suffer from a lack of precision and generality. Also the extreme Bayesian approach with arbitrary choice of the prior probability density or priors deduced from scaling laws is rejected.