RISK SURVEY

The AI Safety community does not typically include translation errors in automatic translation when making an inventory of risks associated with AI. The primary purpose of this survey is to contribute to a collection of examples of high-risk errors, thereby strengthening the argument that, in certain use cases, translation errors can indeed cause harm and emphasizing the need for professional translators.

We are looking for documented examples of automatic translation, including neural machine translation (NMT), generative AI (GenAI) translation, etc., that will help us illustrate errors that could have caused harm if not corrected by a professional translator before the translation reaches the end users, be it an e-book, an article, an advertisement, or any other type of publication. These near-miss errors (the translation almost hit the target of being a quality translation) could appear in any of the following processes:

• producing a translation from a source text, using a tool that proposes suggestions from a termbase, a translation memory, an automatic translation system, or any combination of these three sources
• revising the work of another translator
• post-editing an entire document was produced using raw automatic translation

The form below was designed to collect examples that YOU found in your work reviewing material produced by automatic translation. We would like to know the subject field, the specific example with information related to where it would have appeared (newspaper, magazine, website name/URL, etc.), the proposed correction, and the possible harm the error could have caused.