Tom Gally portrait



The Implications for Translation and Translators of Recent Advances in Large Language Models

Updated April 17, 2023


Introduction (top)

The large language model GPT-4 released by OpenAI on March 14, 2023, surprised many people with abilities that exceed those of its predecessor, ChatGPT, released just a few months earlier. Preliminary tests of GPT-4’s ability to translate between languages suggested to some observers that human-level translation ability had at last arrived for machine translation, even for languages as different as Japanese and English. However, others, including some professional translators, remained skeptical.

An informal online discussion about GPT-4’s current translation ability and the implications of such LLMs for translation and for translators was held using Zoom from 10:00 a.m. to 12:00 noon Japan time on Friday, March 31. The discussion, which took place in English, focused on translation between Japanese and English and the concerns and interests of professional translators. The event was organized by Tom Gally (CV) of The University of Tokyo.

The following questions had been presented to the participants in advance:

Summary (top)

After the event ended, a transcript of the audio recording was prepared using Whisper and summarized by GPT-4. That summary appears below. (Because of length limitations, the transcript had to be fed to GPT-4 in sections; as a result, some of the continuity of the following summary is a bit strange.)
On March 31, 2023, a Zoom discussion was held with participants from various translation backgrounds. The host began by sharing survey results that revealed the majority of attendees were freelance Japanese to English translators with diverse specializations. Most respondents had not yet tried GPT-4 for translation, prompting the host to provide a demonstration later in the discussion.

The host introduced himself and shared his professional background, which included freelance translation, teaching at the University of Tokyo, and working on dictionaries for various publishers. He expressed concern about the implications of AI in language education, particularly since the improvement of Google Translate in 2016. The host had organized the event to explore the field of translation as a whole and to gain insights from professionals with different perspectives on AI's role in the industry.

During the discussion, participants shared their experiences with machine translation tools, including Google Translate, DeepL, and SDL Trados. The future of translation as a career and the implications of AI for society were also topics of interest, with an open letter signed by over a thousand researchers, academics, and politicians calling for a six-month pause on the development of new AI models.

The conversation provided a fascinating glimpse into the experiences of translators and writers using GPT-4 as a tool in their daily work. While the technology had its limitations, many participants found value in incorporating it into their workflow to improve translations and generate more natural-sounding language. The discussion also highlighted ethical considerations and potential impacts on language use and change.

In the second half of the discussion, participants explored the impact of GPT-4 on the translation industry, its effect on individual translators, and how to prepare for rapidly changing technology. The conversation covered the potential of using GPT-4 for translation workflows and the business and future of translation, focusing on how clients might react to recent advances in large language models.

The importance of human translators acting as editors and writers in addition to translating text was emphasized, as well as the need for understanding the purpose and audience of a text. While AI tools like GPT-4 could provide valuable assistance in the translation process, human translators still had a unique role in ensuring high-quality, engaging translations.

The discussion shifted to the impact of AI on language usage and change, questioning how AI would adapt to the evolving nature of language. Participants debated Chomsky's recent criticism of AI and linguistic theories and expressed concerns about the future of various careers as AI tools continued to improve rapidly.

The conversation served as a case study for professions likely to be significantly affected by AI advancements. Participants noted the difficulties of incorporating AI tools into the translation workflow, acknowledging that while machine translation could help when stuck, it might also limit creativity and lead to a single translation option. Some attendees shared their preference for using AI tools as a source of ideas rather than relying on them entirely.

Regarding the future of translation, one participant wondered if focusing on interpretation might be a better approach, especially in the Japanese context, where the human element was essential. The speaker emphasized the necessity of individuals who were well-versed in both languages and cultures to facilitate negotiations and contacts among people.

In the conversation, one participant spoke about using machine translation tools such as DeepL and ChatGPT in their translation workflow, emphasizing the potential problems they can create. They mentioned that machine translation can occasionally confine translators to a single option, diminishing their creativity. The discussion then shifted to employing AI for writing essays, which can yield well-structured yet unremarkable content.

Another participant shared their experience with translation software like DeepL, asserting that it eliminates the creative process. They also pondered if concentrating on interpretation could be crucial in the future, as human involvement remains indispensable in certain situations.

A small press publisher revealed the impact of GPT-4 on the publishing industry. Some small presses have closed submissions due to the influx of computer-generated submissions, complicating matters for new, genuine authors seeking to publish their work.

A more optimistic perspective on AI and translation was presented by another individual, who cited "The Myth of Artificial Intelligence." They argued that a fundamental difference between humans and machines would safeguard the translation profession. They noted that the number of people working in the translation industry has increased despite advancements in machine translation.

When asked about recommending translation as a career, this person advised young individuals to develop a diverse skill set and specialize in a specific area rather than solely relying on their bilingual abilities. They believe that humans will continue to hold a unique position in the industry, with AI serving as a tool to enhance their work.

During the conversation, an interesting remark was made about boutique translators, suggesting that as AI progresses, translators must differentiate themselves and offer specialized, high-quality services. One attendee compared their experience as a freelance translator to operating a small shop in a big city, providing a particular skill or service. As they gained expertise, they started to see themselves more as a boutique, offering a higher skill set and more personalized services.

The discussion also touched on the importance of establishing a distinct identity and presence in the industry, which can provide some level of security against advancing AI technology. As an example, one participant cited their daughter's freelance illustrator career, where her personality and artistic identity significantly contributed to attracting clients and securing work.

A challenge for translators is that much of the translation work is faceless and not publicly attributed to them. The most vulnerable work is hidden behind agencies or situations where the translator does not directly interact with clients. Translators with a visible identity and personal client interaction are considered safer, at least for now.

Regarding AI's future, participants expressed opinions ranging from pessimism to optimism. Some leaned more towards the pessimistic side, believing AI could eventually replace numerous translation jobs. The conversation also touched on a conference theme focusing on the invisibility of translators and interpreters.

One participant inquired about the benefits of upgrading to GPT-4, and it was suggested that GPT-4 is 20-30% better at producing text than GPT-3.5, although translation improvement is less significant. However, subscribing to GPT-4 can help users stay current with AI's latest developments.

As the discussion concluded, participants were encouraged to try GPT-4 and experiment with its capabilities. The host invited them to share their thoughts on the event website, where a summary of the conversation and selected comments would be posted. The host also expressed a willingness to host similar discussions in the future.

In conclusion, the conversation emphasized the importance of translators differentiating themselves, building a distinct identity, and offering specialized services in the face of rapidly advancing AI technology. While AI's future remains uncertain, the discussion helped clarify participants' thoughts and encouraged them to monitor the latest AI developments and their potential impact on the translation industry.

Comments (top)

Several participants submitted comments after the event. Here are some excerpts:
I don't think that the demise of translation by humans as a viable profession is imminent. Nevertheless, AI is already driving changes in the profession. As the power of AI tools continues to improve rapidly, the space for human translators to thrive will shrink.

As a freelance translator early in my career I would love to know more about how to 'prompt' GPT for J to E translations, or helping me brainstorm better translations and alternative ideas.

  1. I would never feel guilty about using a tool that is out there to create a good translation. I don't think "cheating" exists in translation and I think being resourceful is a skill translators need. I will openly use all dtools to produce a better translation and I think that is in the best interest of the client.
  2. I am on Camp Pessimism and agree with Rick. I see a lot of translators on Twitter perhaps trying to convince themselves that "ours jobs are not at risk" because Chat GPT is not perfect. No, it's not 100%, but it might go to 99% in a few years and then the question will be, how much will a customer pay for the remaining 1%. Not all translators will be gone, but not all of us will survive. Like the taxi driver, or bookseller, or travel agent or anyone whose job has been affected by technology.
  3. Many translators lean introvert. The message that they should perhaps show their human side and be interpreters or otherwise do something where they are interacting more with people is likely something they don't want to hear. Interpreting is like being in show business where you are performing live. Though there is an overlap of skills, translators and interpreters in my opinion are two different kinds of personality type.
  4. Agree that we have to be specialists in something. Experts. And we will have to choose our specialties with the future in mind. Will my specialty be around in 20 years?

What I find most useful is that I can ask stupid questions. Its responses are mostly spot-on, as you noted. It helps me parse out the complexities of my client’s industry and their requirements. I can request the source document, version, and section number to validate where its “quoted text” came from. When I found no such section existing in the actual version, it profusely apologizes and blatantly acknowledges that I am correct. How many authoritative human expert possess such admirable personality traits, and the time to waste in stupid chats. Why, it reminded me of a border collie eager to please, and it made me smile.
    In drafting this thank-you note, I asked it for its preferred pronoun, but it repeatedly insisted on having no preferences. I saw in those responses a line of impartiality drawn in the sand, a sign of its upbringing. I am grateful, for future peace on earth, to the smart people putting thoughts into training it to be a good-natured being.

I had a bit more optimistic view for our future after viewing the discussion as a company in-house translator who is facing the retirement in 4 yrs, focusing more on how to improve my writing skills in both Japanese and English as well as expertise in my field of the industry.

From Ellen Tamura:
    I've worked mainly through translation agencies since 1987. In the 1990s, I was paid as much as 6,000 yen per 200 words by a translation agency to translate reports intended for publication. Half that amount would be a good rate today. About three years ago, my main agency asked if I would consider switching to just editing the output of machine translation. They wanted me to refer to the original Japanese only when the English was unclear. Of course, they wanted to pay only half the rate they'd been paying me for translation. At the time, I said no because it sounded like too much hassle for too little money. I've continued to receive a reduced amount of translation work from that company - mainly more complex texts and PowerPoints which are apparently difficult to machine translate.
    Since machine translation output has improved dramatically recently, I might reconsider and start editing machine translations for that company. One of the biggest problems I see with that kind of work is that machine-written text can get weird before the reader even realizes it. If I'm reading machine-generated text all day long, I become acclimated to the weirdness, unfortunately. So, I would approach machine-generated text with even more skepticism than I approach human-generated text so as not to get jerked around by the machine. That might be an odd way to put it, but that's basically what it comes down to. At least in the case of Japanese, I can refer to the original to sort things out (however, in the future, that might be machine-generated, too).
    I can't comment on the overall market for Japanese-to-English translation because I haven't been applying for new work, but of course no one is going to pay someone to translate when they can just run documents through Google Translate, DeepL, or GPT. This means that, even if people will pay for translations of more complicated texts, there will be no entry-level work for new translators to develop their skills. Machine translation makes errors, but as AI improves, the errors are fewer and fewer. In the past, I was occasionally asked to edit translations by other humans, and human translators, who are much slower and cost a lot, also make plenty of egregious errors.
    Currently, I think many people vastly underestimate the power of AI, and I think translation work as we know it is finished. As Tom mentions, there may still be jobs in cross-cultural training and hand-holding for clients venturing into foreign markets, but that's nothing like the job of being a specialist in translation as we've known it heretofore. In fact, I think there may soon be AI avatars that can do cross-cultural training as well as humans can. The only thing left for humans to do will be to advocate for the biosphere (which machines don't need). We might still want humans to translate poetry or very important speeches, etc., but as with buggy whip makers, it will be a very niche market (unless there is a major electro-magnetic pulse event that destroys all of our electronics and sends us back to the Stone Age).
    I'm a little sad to see the relationship between the translator and the text change and to see the skills I've learned becoming largely obsolete. Translating the old-fashioned way involves much more mental activity than running a text through a computer - I remember in the early days looking up each kanji I didn't know one by one in paper dictionaries (the baseline for function literacy is over 2,000 kanji) until the dictionaries were grimy and ragged, wracking my brains over grammar and syntax, figuring out who was trying to communicate what to whom and why, thinking about how best to put that into English, and staying up until 2 am to meet deadlines while also raising babies (and yes, I also walked uphill to school, both ways, in the snow), etc.
    Now if only the machines could figure out how to slow down the mass extinction of the biosphere, not to mention how to resolve our social problems.

Video (top)

A week after the event, the organizer made a video inspired by the discussion. This video was aimed at nontranslators and can be viewed here.