In December 2022, we first reported on ChatGPT, the GPT-3 version of which had been made available to the public free of charge a month earlier. Over one million users had already signed up in just five days, indicating that OpenAI had introduced a technology poised to revolutionize the world once again.
After a little over a year, we can better assess the strengths and weaknesses of ChatGPT and, most importantly, determine whether it can serve as a viable alternative to DeepL and other machine translation systems, a question currently on the minds of many in the translation industry.
Is ChatGPT suitable for translation?
ChatGPT is a language-based application with which people can exchange information on a dialog basis. The potential uses of ChatGPT seem limitless, and since its launch, users have explored a wide range of scenarios with the chatbot with various degrees of success. Well-crafted prompts are essential for receiving the best possible answers, which is why large communities have formed to compile and exchange the best prompts. There is now a new job going by the name of “prompt engineer,” which involves effectively communicating with the various AIs.
If you pose the question to ChatGPT, it states that it specializes in text generation, text comprehension, code generation, language translation, conversation, factual knowledge, image and data processing, and the capacity to learn. You have to keep in mind that “specialize” is an elastic concept, particularly in regard to “factual knowledge,” which is debatable given that it is now widely known that the chatbot simply “hallucinates” some facts. To at least partially circumvent this issue, you can use ChatGPT through Bing, which has been possible since the establishment of a partnership with Microsoft. ChatGPT makes the Bing search engine database available for this purpose. The advantage of searching via Bing Chat is that the sources are linked, allowing users to verify individual items of information.
It is interesting that ChatGPT itself mentions translation, as the program is a chatbot and not a machine translation system like DeepL. For translation purposes, the chatbot simply needs to be provided with a text and asked to translate it into a specific language (ChatGPT currently supports 85 languages).
AI expert Katharina Zweig told German-language business newspaper Handelsblatt that the AI can translate even without specific training and that it has acquired an understanding of the “essence” of languages. This was possible thanks to the vast amounts of text data provided to it in various languages.
Nevertheless, it should be noted that ChatGPT was primarily trained using English data, which is why English communication may appear more “natural” in certain instances. For example, if you ask ChatGPT to write an e-mail in German, the first sentence following the salutation is often: “Ich hoffe, diese E-Mail erreicht Sie wohlauf.” This is a literal translation of the phrase frequently found in English e-mails “I hope this e-mail finds you well,” which is not commonly used in German.
Since English and German share relatively similar language patterns, translations between them are expected to yield better results compared to languages with less in common, such as German and Chinese. Nevertheless, we are already familiar with this issue from machine translations using “conventional” systems such as DeepL.
Differences between ChatGPT and DeepL
Unlike ChatGPT, DeepL was explicitly designed and trained for translation. For this reason, its translations are currently better than those of the chatbot, particularly when it comes to more complex texts.
Furthermore, DeepL offers certain features that enhance and streamline machine translation, which ChatGPT currently lacks and probably will not develop in the near future, as translation is not the primary focus of AI and OpenAI.
DeepL’s special features include the ability to set the form of address, create a glossary, offer alternative translation suggestions, and integrate with translation management systems (TMS) such as Across Language Server, which can automate and accelerate the translation process. One advantage of this is that once sentences have been post-edited, they are stored in the translation memory of the TMS. This makes it possible for the translations to be reused in case of repetition without the need to edit the sentences again.
In order to use ChatGPT for translation, you need to manually copy the text from the original file and paste it into the chat, and then the translation must in turn be copied and pasted into the target file. This may not be much of an issue for occasional translations, but the process is too cumbersome and slow for professional use.
The data security for ChatGPT and DeepL is comparable: In the free version, user data is used for training, but this can be avoided by using the paid premium version. With ChatGPT, the Team version is required for this, which costs €23 per month, while the Starter version for €5.99 per month is sufficient for DeepL. Integration with translation management systems is available only in the Advanced version, which is priced at €24.99 per month.
We believe that ChatGPT is not (currently) a suitable replacement for conventional machine translation systems such as DeepL, Google Translate, and others, as these systems were designed specifically for translation and have specific features that enhance translation quality.
Another alternative is customizable machine translation systems. In contrast to generic systems such as DeepL, customizable machine translation systems are trained with customer-specific data, which allows them to take into account the unique corporate language and terminology and thus provide more accurate translations. The resulting raw translations are better and require less post-editing. One example of this is AcrossMT, our in-house MT system, which is trained with your specific data (both translation memory and terminology) to deliver better translation results.
How ChatGPT can help with translation
However, this does not mean that ChatGPT cannot be beneficial for the translation process or for the translation itself. With its extensive knowledge, the chatbot can offer a wide range of tips and tricks relating to translation processes and translation management. Why not ask it the following questions:
- How can my company set up the most efficient translation process possible?
- What steps can I take to enhance translation management within my company?
- How can I find a translation service provider that is a good fit for my company?
- How can my company effectively integrate machine translation while maintaining the quality of the texts?
Tip: To get more precise answers, specify the industry and/or other general information about your company. You can also ask follow-up questions so that ChatGPT can provide more specific and detailed information.
ChatGPT can also assist with the translation process. For example, you can ask it to clarify specific terms or concepts in the source text. You can then ask it about the translation of these particular terms within the specific context. This is particularly useful if you are unable to locate an appropriate translation in the various dictionaries and translation databases such as Linguee or Reverso.
In these use cases, it is of course important to critically scrutinize the answers and compare them with other sources. As previously noted, ChatGPT may occasionally “hallucinate” false information, although this is rarely the case when general questions are posed. ChatGPT is a great place to start your research if you’re not very familiar with a topic and want to get a general understanding before conducting more in-depth research.
The bottom line
According to a survey conducted by the Statista Research Department on January 2, 2024, around 9% of respondents use ChatGPT several times a day, while an additional 18% ask the AI bot questions several times a week. Only 29.7% of respondents never use it. In other words, ChatGPT and similar AIs have become an integral part of our daily lives, and we have to learn to deal with them. This includes learning to ask the right questions and conducting thorough fact-checking.
Currently, the chatbot is not well-suited for routine machine translation tasks, but this could change in the future, especially with the rollout of MT systems based on ChatGPT’s datasets. Until then, companies interested in using machine translation should instead depend on proven MT systems that are integrated through an API with a conventional translation management system such as the Across Language Server.
