Regarding the actual automotive safety norms the use of artificial intelligence (AI) in safety critical environments like autonomous driving is not possible. This paper introduces a new conceptual safety modelling approach and a safety argumentation to certify AI algorithms in a safety related context. Therefore, a model of an AI-system is presented first. Afterwards, methods and safety argumentation are applied to the model, whereas it is limited to a specific subset of AI-systems, i.e. off-board learning deterministic neural networks in this case. Other cases are left over for future research. The result is a consistent safety analysis approach that applies state of the art safety argumentations from other domains to the automotive domain. This will enforce the adaptation of the functional safety norm ISO26262 to enable general AI methods in safety critical systems in future.