Animals' relationship to humans, i.e., their perception of humans, is an important parameter to be included into on-farm welfare assessment schemes. The objectives of this study were to evaluate interobserver reliability and between-experimenter repeatability of three previously used and two new tests feasible for on-farm assessment of animals' relationship to humans in loose-housed dairy cows. Additionally, inter-test relationships were analysed to obtain first information on the potential validity of the two new tests. Two experimenters performed five tests (avoidance distance test at the feeding place [AvoidF] and in the barn [AvoidB], approach test to a stationary human in the barn [Appr], passing and touching lying animals [LyingPT], walking through the herd and trying to touch animals [WalkT]) on 16 commercial dairy farms in a balanced order. The person actually not testing the cows observed and recorded the test results from a distance (observer). Inter-observer reliability was investigated by relating results of observer and experimenter. Between-experimenter repeatability, comprising inter-observer reliability as well as consistency of reactions of animals to humans, was evaluated by correlating the test results of both experimenters. Observer or experimenter biases were evaluated by testing for significant differences. High inter-observer reliability was found for all tests under consideration: correlation coefficients exceeded 0.9, percentage agreement for LyingPT was 100%, and negligible observer biases were found. Between-experimenter repeatability of the two avoidance distance tests (AvoidF and AvoidB) was high to very high (r = 0.7 to >0.9), both at an individual as well as at herd level. Evaluated at herd level, LyingPT was also highly repeatable (r = 0.7), whereas WalkT and Appr showed only moderate (0.4 to <0.7) between-experimenter repeatability. With regard to inter-test relationships (used for a preliminary investigation of convergent validity of the new tests), the avoidance distance test in the barn (AvoidB) highly correlated with AvoidF and WalkT (r = 0.7-0.9) and moderately (0.5) with LyingPT. The approach test (Appr) did not correlate significantly with any of the other tests (r = 0.0-0.4). Principal component analysis results showed that all but the measure of Appr loaded highly vertical bar >= 0.7 vertical bar on one component. In sum, regarding reliability and convergent validity, both avoidance distance tests can be recommended as tests for assessing the cows' relationship to humans. In addition, the high interrelationship between WalkT and the avoidance distance test in the barn (AvoidB) makes WalkT a promising test for further development despite the absence of sufficient between-experimenter repeatability in the present study. (C) 2008 Elsevier B.V. All rights reserved.