The principal goal of this study was to determine whether the results from a set of selected currently available alternative methods as used by cosmetics companies are valid for predicting the eye irritation potential of cosmetics formulations and ingredients and, as a consequence, could be valid replacements for the Draize eye irritation test. For the first time in a validation study, prediction models (PMs) that convert the in vitro data from an assay to a prediction of eye irritation were developed for each alternative method before the study began. The PM is an unequivocal description of the relationship between the in vitro and the in vivo data and allows an objective assessment of the reliability and relevance of the alternative methods. In this study, 10 alternative methods were evaluated using 55 test substances selected as representative of substances commonly used in the cosmetics industry (23 ingredients and 32 formulations). Twenty of the single ingredients were common to the European Commission/British Home Office (EC/HO) eye irritation validation study (Balls et al., 1995b). The test substances were coded and supplied to the participating laboratories. The results were collected centrally and analysed independently, using statistical methods that had been agreed before the testing phase began. Each alternative method was then evaluated for reliability and relevance in assessing eye irritation potential. Using the criteria of both reliability and relevance as defined in the study, the preliminary results indicate that none of the alternative methods evaluated could be confirmed as a valid replacement for the Draize eye irritation test across the full irritation scale. However, three alternative methods-the fluorescein leakage test, the red blood cell assay (classification model) and the tissue equivalent assay-each satisfied one criterion of reliability or relevance. Further investigation of the decoded data from this study to explore more fully the relationship between the in vitro data and the in vivo data is recommended. Such a review may allow the development of new prediction models to be tested in a subsequent validation study.