With artificial intelligence quickly establishing itself in many fields, AI tools are also increasingly being used in translation. Despite its progress, AI still lags behind the skills of human translators. Nevertheless, translators are already actively experimenting with AI tools and developing creative approaches to effectively incorporate them into their work.
The results from a study funded by Pro Helvetia and conducted by the A*dS (Authors of Switzerland) provide an assessment, which institutions can use as a starting point for further developing cultural policy instruments/funding measures. In the following interview, Cornelia Mechler, Managing Director of the A*dS, presents the study’s key findings.
What questions does the study on AI in literary translation address, and what was the approach?
The 15th Swiss symposium for translators, held annually by the A*dS, focused primarily on the automation of the literary translator profession. A project group led by translator Anita Rochedy conducted a survey to understand the phenomenon in Switzerland. It drew on a survey conducted by the Association of translators (VdÜ) and the French Literary Translators’ Association (ATLF) in spring 2023 andadapted it to Swiss specifics.
The quantitative study surveyed translators living in Switzerland who work in publishing, regardless of whether they are members of the A*dS. For the qualitative study, a group of five test translators was given texts. They were given the task of editing a text translated by AI while following various instructions, including using DeepL as a dictionary.
Could you summarise the key findings?
Editing a text pre-translated by AI, at best, results in no time saved and, at worst, leads to a significant loss of time. The various experiments also demonstrated that translation is ultimately a skill of reflection and analysis; it requires a sensitivity that machines have not yet achieved.
Translators will undoubtedly be faced with new tools more frequently in the future, and their working conditions will inevitably change. The issue of copyright also emerged in the context of AI. Machine-generated text, whether ‘created’ by ChatGPT or DeepL, does not yet fall under copyright law. In addition, these machine-generated translations only exist because the machine exploits the work of other translators; in other words, it has been fed by human translations that should actually fall under copyright law. Therefore, machine-generated texts and translations remain in a legal grey area of ‘intellectual creation’.
What can AI tools achieve in translation, and what are their limitations?
According to the reports from our participants, although a rough translation is produced more quickly using a tool like DeepL, the subsequent steps become so laborious and time-consuming, making it difficult to speak of saving any time at all. Every decision is preceded by negotiation with the machine and to what extent you should or shouldn’t intervene in the translation. The individual tasked with post-editing, who worked solely with the machine-generated translation without access to the original text, paradoxically stated they apparently felt the least freedom, as they did not dare to deviate from the text for fear of losing sight of the original, to which they had access only through the AI’s filter.
What knowledge of copyright law do users who want to use AI tools need to have when using such tools?
It is fundamentally advisable that for any AI tool used for translating copyrighted texts, it should be clarified in advance whether that system offers the option to disable saving the input text for the purpose of training the AI. DeepL, for instance, offers a free and a paid DeepL Pro version. The free version stores any input and processes it further, which they refer to as ‘quality improvement’. When using the paid version, texts are not stored
What challenges will these new AI tools pose for translators? How does the translation scene look now, and are there any concrete demands?
Developments in the field of AI are very dynamic. At present, the main problem is that there are many legal ambiguities and inadequate statutory regulations. Literary translations (in contrast to specialised translations) are considered an artistic activity, in which copyright therefore plays a decisive role. Apparently, AI or the use thereof is not yet an issue for most German-language literary publishers. However, it is to be expected that translators will receive more post-editing requests. It’s therefore extremely important that literary translators familiarise themselves with the material at hand – you can only argue and negotiate if you know what you’re talking about, after all.
For the A*dS, it’s clear that AI-generated translations can never be an alternative to human-generated translations of literature. There’s a lot at stake here. The study has already revealed a clear impoverishment of language, including a potentially significant loss of emotional, empathetic, sarcastic and ironic undertones that can only be recognised by humans. The working conditions of translators are also extremely precarious and are sure to worsen due to AI being used without much thought being put behind it.
The A*dS will be closely monitoring developments in AI and legal regulation in Switzerland, but especially in the EU. They will also be offering further training courses for authors and translators regarding the AI sector in autumn 2024.
