Copy-paste from Brussels: The EU’s error-riddled new sanctions package was generated using an LLM

The EU's most recent sanctions package against Russia is marred by numerous errors, from mixing Latin and Cyrillic characters to listing names without birthdates and using actors' stage names instead of real surnames. The errors are striking and inconsistent, indicating a lack of thoroughness in preparing the document. One glaring example is the address of Neva-Ferrit LLC, which was rendered in a haphazard fashion, mixing translation and transliteration in a nonsensical manner.

This suggests that large language models (LLMs) may have been used to generate parts of the document, as well as a Python transliteration library. The use of LLMs and copy-pasting instead of meticulous bureaucracy results in boilerplate language being applied to justify sanctions against 1,569 individuals and organizations. Inconsistencies are also evident in how Russian names are translated, transliterated, or both.

For instance, the address of Neva-Ferrit LLC appears to have been a product of such automated systems, with a mix of translation and transliteration that makes little sense. Inconsistencies in the translation and transliteration of Russian names and places further point to the use of AI-generated text. Another example is the English-language address for CJSC Tehkrim, which appears to be generated by a large language model.

The errors are consistent with those that occur when neural networks handle multiple languages at once. Overall, the EU's sanctions package is a prime example of the potential pitfalls of relying on AI for tasks that require accuracy and attention to detail.