Abstract The new paradigm of neural machine translation is leading to profound changes in the translation industry. Surprisingly good results have led to high expectations; however, there are substantial risks that have not yet been sufficiently taken into account. Risks exist on three levels: first, what kind of damage can clients and end users incur in safety-critical domains if the NMT result contains errors; second, who is liable for damage caused by the use of NMT; third, what cyber risks can the use of NMT entail, especially when free online engines are used. When establishing sustainable measures to reduce such risks, we also need to consider general principles of human behaviour if we want to make sure that all agents comply with provisions and requirements.