What other creative professions can learn from translation about the effects of artificial intelligence

On Nov. 30, OpenAI publicly launched the ChatGPT chatbot built on its GPT-3 large language model (LLM). The new chatbot quickly caught the world’s attention, garnering over 1 million users in the first five days after its release. Its ability to provide plausible, well-written answers in many different languages and writing styles in response to natural language queries made in different languages astounded the general public and set off alarm bells in education and many of the creative professions, like copywriting, advertising, and journalism. Several of the largest school districts in the United States quickly banned ChatGPT outright from their networks, citing concerns over academic honesty. Even a leading artificial intelligence conference banned the use of ChatGPT to help write papers for submission. 

People in creative professions like advertising, technical writing, journalism, and illustration began imagining doomsday scenarios of being replaced by technology, with generative AI tools like chatbots taking over their work. To translators, this all sounds remarkably familiar.

The translation profession has not been known traditionally for being ahead of the technological curve, but in this case, it is by about six or seven years. The concerns currently running through academe and the creative professions are strikingly similar to those expressed by translators beginning in 2016 with the release of Google’s neural machine translation (NMT) service. These concerns only intensified when another NMT service, DeepL, was released in 2017. As a result, translation has had a head start in adapting and managing the technological change brought about by state-of-the-art artificial intelligence, specifically NMT. 

NMT technologies like deep neural networks and LLMs and the machine translation tools and chatbots built on them treat data in similar ways. “Translation is the key to everything else,” said German-English translator and translation technology expert Jost Zetzsche. “Once translation is done well by machines, everything else falls into place. This is how developers saw it, and it is why they focused on translation first.”

This same technology that enabled NMT to improve machine translation is now finding its way into seemingly unrelated products (think ChatGPT or Dall·E). The translation “problem” has not been solved with the latest advancements in NMT, but there are lessons to be learned in other professions from how the professional translation community has been affected by and adapted to the widespread use of NMT both by translators and by end users.

Five Lessons Learned from Translation Since the Launch of NMT

When a potentially disruptive technology is released, marketers differentiate the product from any competition and focus on wowing potential clients and the general public. Media attention depends greatly on that wow factor to create a big marketing splash and boost awareness. As a result, press releases and initial media coverage tend to overflow with superlatives and focus on any example that suggests the day of artificial general intelligence has arrived. Just think about all those Star Trek Universal Translator and Babelfish headlines you’ve seen over the years. However, the real impact of a new technology on actual workloads and workflows only becomes apparent over time. 

Here are five lessons that emerged from interviews conducted with practicing translators about how NMT has changed the way they work over the past six years. 

  • Translators Are Still Very Much in Demand

That’s the headline. Translators are still in high demand. But that doesn’t mean that they haven’t had to adapt. NMT tools like DeepL, Google Translate, and Microsoft Translate, are just that: tools. And like other tools, a hammer for instance, they can be used by anyone from skilled professionals to do-it-yourselfers with varying levels of skill and success. According to Dion Wiggins, CTO and co-founder of Omniscien Technologies, consumer NMT services have become quite mature with consumer-level use cases now well established. Google alone translates more than 100 billion words a day. 

That said, it is important to note that these use cases generally cover translation tasks that were never originally handled by professional translators in the first place. With this broad use of the technology, end users have become more aware of NMT’s abilities and limitations. In the end, NMT has made cross-language communication more available than ever, and although there are inevitable misapplications of the technology, consumers are increasingly savvy about the limitations to what NMT can do and when they will need a professional translator.

As literary translator Tim Gutteridge put it, “Translation is not a cake of a finite size. [Technology] will allow people to translate things. But it’s not a zero-sum game.” However, just where the line dividing when NMT is good enough and when a human translator is preferrable is dynamic and will continue to change and encroach upon certain translation market segments.

  • Translators Have Learned that NMT Can Increase Their Productivity

All translators interviewed for this article said that the use of NMT tools has had a positive effect on their productivity, directly or indirectly. Many noted significant increases in their productivity because of how they employ NMT tools. 

The main time saver mentioned was using NMT just to get the target text quickly into the computer and on screen without having to type tens of thousands of words on a keyboard. Once in electronic form, the draft target text can be adapted and corrected using the translator’s preferred tool. Granted, this approach works better with certain types of texts that lend themselves to machine translation, like news articles or instruction manuals.

“DeepL was a game changer for translating press articles. I have easily doubled my output and shortened turnaround times. It streamlines my work. I have to do some post editing, based on what country the article is from. But it’s a tool, not a crutch,” said Donatella Ungredda, a US-based Spanish-English translator and interpreter.  

However, translators had to learn to recognize the quirks of NMT and the texts it does not handle well. Thomas West, a US-based translator specializing exclusively in legal translation, noted, “I ran an agency for 25 years. I edited a lot of human translation. I’m comfortable editing. But it has taken me time to get used to the mistakes that DeepL makes because they are different. Consistency is very important in English legal writing.”

“If a writer uses a synonym, it must be for a specific reason,” West continued. “DeepL is a smorgasbord of switching synonyms, which can be a mess for legal translation. DeepL does not do well with Latin American legal documents either specifically because of the terminological variation in the source text.”

  • Some Types of Translation and Market Segments Are Being Dominated by NMT

For decades, the translation profession has operated on the premise that every translation must be 100-percent accurate and of impeccable linguistic quality. With the advent of deep neural networks, the easy-to-spot awkward syntax and grammar produced by previous translation models have largely been replaced with other more subtle problems like hallucinations, deletions, and mistranslations. Even so, it is now becoming clear that more and more end users of translation have concluded that NMT can be good enough for certain types of texts and use cases. Informational texts that have a limited useful life, like news reports and different types of legal documents, such as court decisions, are now often translated using NMT services and are seldom post edited. Some end users are prioritizing turnaround times and costs over fluency and even translation accuracy. 

The shift to post-edited machine translation (PEMT) by large agencies and some direct clients is affecting bottom lines. For translators who have seen their work move to all post editing, their income has taken a hit, according to Zetzsche, as the size of contracts for post-editing work tend to be significantly smaller than for traditional translation work. As this trend has grown, some translators have transitioned to other language services or left the profession altogether, while others still have chosen to focus on working with direct clients in the high-level or premium market, which so far has proven more skeptical of the NMT trend. 

Another market segment that is evolving with the advent of NMT is legal translation. West explains that “good-enough” translation is gaining a foothold in the legal sector. “But you must really narrow down and focus on what the translation will be used for,” he said. “Legal translation is largely for informational purposes. The translation must be accurate, but it may have never needed to be word smithed because it is for informational purposes only. These are not texts that are going into a glossy brochure.”

  • Translators Have Been Creative and Cautious in How They Have Used NMT Tools

Frequently, translators assume that NMT will lead to them doing nothing but post editing. The past six years have shown otherwise. “Some colleagues are going to all post editing. I’ve tended to go in the opposite direction,” says Tim Gutteridge, a UK-based Spanish-to-English translator. For those that have seen their work transition to all post editing, their income has taken a hit, particularly translators who work almost exclusively for large translation companies, according to Zetzsche. 

Today, translators use NMT tools, such as Google Translate and DeepL, as suggestion tools while translating. Zetzsche explained, “Most translators use NMT much as they use other tools like translation memories, term bases, or corpora.”

US-based English-and-Spanish-to-Italian translator Riccardo Schiaffino agrees. “I use the suggestions probably more than 50% of the time,” he said. “You can see if [the tool] will be useful at the beginning of a project. Interactive mode speeds up my work a little bit. But it doesn’t improve the quality of the work. It’s like having a really good translation memory.” 

“I use DeepL to find other options. It helps you determine when a translation is off,” said Hedwig Spitzer, a French to Spanish translator and conference interpreter based in Peru. 

Spitzer also shared another novel application of NMT tools to assist with conference interpreting. “I use DeepL when I receive speeches at the last minute. Because speakers often read dense speeches at the speed of light, we usually do a quick machine translation of the speech if it is a public event. But it still must be reviewed.”

This is a growing practice in conference interpreting as more and more events are heavily scripted and recorded for on-demand viewing in multiple languages, further blurring the lines between translation and interpreting.  

Interestingly, even when using the paid versions of NMT tools, many translators are still skeptical about data security claims, and many refuse to employ the technology for certain clients with particularly sensitive topics or when the translated content is of a confidential nature. Some clients even include clauses in their contracts prohibiting the use of machine translation. This concern, however, may be overstated.

“The privacy problem is kind of solved for most engines,” Zetzsche said. “As long as you’re using the API, Google, Microsoft, and DeepL are guaranteeing not to use the data. The problem, in my opinion, is that translators have used that privacy argument against MT so much that they either can’t or don’t want to realize this.”

One thing is certain. Data privacy and the copyright of the data used to train LLMs are two significant controversies yet to be resolved when it comes to generative AI.  

  • The Job of Translators will Continue to Evolve

The use of NMT in business scenarios is still evolving, with applications of the technology still in development. Professional translators have had to adapt and will need to continue to adapt as NMT is applied in more business scenarios. 

Practicing translators have eventually adopted new technologies into workflows, while at the same time emphasizing the more human aspects of translation, like cultural knowledge, domain expertise, and having a feel for what sounds right, as essential to producing professional-quality translations. Ceding ground to technology in professional practice is still a sensitive subject.

Case in point, PEMT. The majority of translators interviewed for this article, for example, have begun offering PEMT as one of their services, albeit reluctantly. Schiaffino began using MT for his work when DeepL was introduced in 2017. As more of his customers move toward PEMT, he decided to offer the whole machine translation and post editing process as a service, as a way to improve quality control.

“I don’t like doing post editing very much but will do it for certain clients,” he said.

Other translators interviewed for this article noted that they use machine translation in combination with post editing on a regular basis but asked to remain anonymous. 

“The very top of the market is not going to be touched at all by this. The rest in one way or another will be touched. If we are smart, we are going to use engines like DeepL ourselves,” said Schiaffino.

One thing is sure, the role of the professional translator has evolved over the last six years and will continue to evolve at ever-increasing speed. 

What’s on the Horizon?

The accelerated development of generative AI tools, like LLMs and AI assistants, is poised to have a significant effect on the professional practice of translation and may radically change the notion of what professional translation is. It is this set of emerging technologies that may prove to be the dark horse in this race of innovation and disruption. And the language services industry would do well to keep a close eye on how and where it is running. 

During a recent online discussion on the future of machine translation, Alon Lavie, VP of language technologies at Unbabel and consulting professor at Carnegie Mellon University, predicted that in a few years’ time, much content will not need to be pre-generated. Clients will rely on AI agents to tell them what they want to know in real time. According to Lavie, there may not be a need to pre-generate content that would then need to be translated. In the future, content will be generated on the fly in multiple languages.

What will this mean for current translation workflows? How will this and other technological developments affect the definition of what it means to be a translator? 

Jay Marciano, director of MT outreach and strategy at Lengoo, puts it this way: “As errors become much more related to the meaning than to grammatical problems, subject area expertise almost becomes more important than language skills. If you have an MA in translation, you need to have more subject area expertise. Your work will change, but it will not go away.”

Barry Olsen
Barry Slaughter Olsen is the vice president of client success at KUDO, a conference interpreter, and professor of translation and interpretation (currently on a leave of absence) at the Middlebury Institute of International Studies.


Weekly Digest

Subscribe to stay updated

MultiLingual Media LLC