The aromatic amines are widely used industrial chemicals and can be found in tobacco smoke as well as in products generated during cooking. In a previous study, we established quantitative structure–activity relationship (QSAR) models linking the carcinogenic potency of non-heterocyclic carcinogenic aromatic amines to a series of molecular determinants. We also found that QSAR models for carcinogenic potency were inadequate in describing the difference between carcinogenic and non-carcinogenic amines [Benigni,R., Giuliani,A., Franke,R. and Gruska,A. (2000) Chem. Rev. , 100, 3697–3714]. In this paper, we derived specific QSAR models for separating active from inactive amines. It appeared that hydrophobicity (as measured by the octanol/water partition coefficient, log P ) played a major role in modulating the potency of the carcinogens, whereas mainly electronic (reactivity) and steric characteristics separated the carcinogens from the non-carcinogens. Interestingly, a similar pattern was previously demonstrated by us regarding their mutagenic activity [Benigni,R., Passerini,L., Gallo,G., Giorgi,F. and Cotta-Ramusino,M. (1998) Environ. Mol. Mutagen. , 32, 75–83]. Based on the QSAR models found, the molecular determinants of the mechanisms of action of aromatic amines are discussed in detail. The QSAR models obtained can be used directly for estimating the carcinogenicity of other non-heterocyclic aromatic amines for which experimental data are not available. With the QSARs in Benigni et al . (2000) and the present results, a two-step prediction of carcinogenicity of aromatic amines is possible: (i) step 1, yes/no activity from the discriminant functions; and (ii) step 2, if the answer from step 1 is yes then prediction of the degree of potency from the equations in Benigni et al . (2000). Thus, QSAR models can contribute to the following: the direct synthesis of safer chemicals; the estimation of the risk posed by amines present in the environment; setting priorities for further experimentation, thus also reducing the use of experimental animals. Whereas the quality of in vivo experimental data is often questioned, the robustness and intepretability of the present results strongly support the reliability of the rodent carcinogenicity assay.