Avian influenza remains a serious risk to human health via zoonotic transmission, as well as a feasible pandemic threat. Although limited zoonotic cases have resulted from the current epizootic outbreak, the wide diversity of influenza viruses in avian hosts means the emergence of new strains that could transmit to humans more readily cannot be ruled out. There is therefore a need to anticipate zoonotic potential before spillover occurs. Here, we develop a novel zoonotic prediction model for avian influenza viruses, building upon "host-predictor" machine learning methods that estimate host potential given only a viral genome sequence. We construct a machine learning framework combining individual sub-models of influenza genome segments, each trained on many genomic and proteomic traits (e.g., k-mer composition, codon biases, protein physicochemistry). To prevent over-fitting to heavily sampled lineages and ensure models generalise to phylogenetically distant viruses, we pre-process training data by considering clusters of shared sequence identity. Curated training sets cover ~4,000 representative, complete genome sequences of avian influenza from 120 subtypes including 9 containing known zoonotic viruses. We combine best-performing models into a single ensemble that can distinguish zoonotic capability of sequences held out from training with strong performance (AUROC = 0.95, F1 score = 0.90), including sequences of rarely-sampled subtypes, e.g., H10N8. Interrogating ensemble model decisions also allows us to identify influential genomic motifs most associated with human infection. These findings suggest specific genomic traits that are key to understanding and monitoring evolution of influenza viruses that circulate within bird populations. Our ensemble model can estimate zoonotic potential for new sequence inputs, offering a means to quickly risk-assess emerging avian influenza strains as soon as a sequence becomes available.