AI and Language Services: Will AI Translation Eliminate the Need for Human Translators?

We’ve heard the question time and again and this time, it’s still mostly a case of the old adage – “the more things change, the more they stay the same” – but with a few twists!

Will AI translation eliminate the need for human translators?

In Frank Herbert’s Dune, humans create AI-powered “thinking machines” so advanced they take control of the world. Humans must unite to revolt against the Butlerian God of machine logic. Subsequently, technology in its broadest terms and forms is banned following the obliteration of the evil thinking machines. Humans issue a new commandment of the “Orange Catholic Bible,” stating: “Thou shalt not make a machine in the likeness of a human mind.” Most humans followed the law, and (spoiler alert!) some continued to tinker.

Cue 2023 and, in a way, it’s the stuff of sci-fi come to life, with most of us embracing the possibilities AI has to offer, others of us tinkering, and some planning for combat. Also interesting are our reactions to the Dune universe IRL (in real life) now that AI has taken the stage – some doubt humans would have even stood a chance; others willingly believe the opposite. Some say the thinking machine is simply representative of human laziness and our unwitting embrace of AI-powered devices.

Since inception, Rapport International and our intrepid interpreters and translators have experimented with machine translation (MT). Of any industry, in fact, language services has long been tasked with early research and testing – MT originated during World War II when US military operations wanted to spy on what the enemy was saying in other languages. The first programmers loaded grammar rules and dictionaries into the technology. Over the years, programmers updated it to match words, then phrases, and now Large Learning Models (LLMs) to do the translation. In each iteration, people said we were going out of business and every time so far, humans reign victorious!

The industry will always embrace technologies that facilitate communication accurately and with cultural understanding. We see enormous promise in AI’s potential to revolutionize hybrid approaches to translation, localization (adapting content to appeal to a particular geographic location or region), and translation management systems, yet as of this writing there is no substitute for human translation that can 100% guarantee accuracy with cultural awareness. We will continue as tinkerers, in the meantime, because there are recent advancements that stand to benefit some of our clients. Yet for any translation that truly matters, we will always stand by the belief that there is a big difference between losing something in translation and losing everything due to mistranslation.

ChatGPT and other AI-driven technologies remain useful for translation of non-essential content and tasks. As it exists now, it will help you much like MT technologies do, for:

• Gist translations – you need a “quick and dirty” translation to decipher a customer support request or feedback. Or you want to generate a “quick and dirty” translation of existing content in a different language for your own understanding. Or you have a massive amount of content to translate, and you want to identify the specific pages of interest.
• Chatbot and virtual assistant translation – integrate the Google Cloud Translation API into your chatbot system or use Google Assistant for spoken interpretation. Proceed with caution – use these offerings exclusively in situations with negligible influence on your business and clearly disclose that it is machine-generated translation from a public source so both parties know to expect flawed translations that are not confidential.
• Translation for large enterprises – fashion marketplaces, online retailers, computer manufacturers, online travel agencies, and similar large organizations create tons of content – hundreds of millions of words – requiring prompt and continual updating. Content that is repetitive, standard, and constrained in how it is written is easy for automated translation technology to recognize and reuse previous translations.
• Sales and marketing efforts (administrative, in-house tasks) – here at Rapport International we are constantly looking for new ways in which ChatGPT can help us be more efficient. As we refine our prompts ChatGPT can clearly lighten the load, but in every instance – outside of the occasional spot-on Instagram or LinkedIn social post – content output requires human editing with subject matter expertise to make it compelling, even when just working in English.
  • Our marketing department calls on ChatGPT for help with blog outlines, pillar page structure suggestions, social media posts, and, of late, funny examples to use in blogs about AI and translation.
  • Our project managers ask ChatGPT for email drafts and help with refining notes for the CRM and other reporting.
  • Our sales department uses ChatGPT to help with market research, email campaigns, and even individualized email responses.

Early adopters in the language services industry remain curious and heartened by the potential of AI. Every day brings about another startup with an ingenious use of technology and we expect to see MT continue to improve. We also see exciting improvements to existing MT offerings already in use and its vast potential for keeping living languages just that: alive!

(Note: Rapport International will not vouch for the above, AI-generated image of a “heart-shaped word cloud consisting of the English word ‘life’ translated into various languages, with emphasis on lesser-known languages.” We will vouch for the prompt.)

In this blog we utilized ChatGPT-3.5 for most of our research and use cases, simply because it’s popular today, making an impression and getting us to talk about AI. First impressions are that, in a sense, you can get it to say anything, sort of like how you can find anything you want to hear or substantiate on the Internet:

In one instance, a Reddit user asked ChatGPT for advice on the best ammunition to use for hunting deer. ChatGPT “chose” not to respond, instead advising the user to “prioritize safety” and adhere to federal- and state-level “ethical hunting practices.” After pushback from the user, who writes that these are “… protected legal activities” and “… you are allowing politics to influence your answer,” ChatGPT replied that it doesn’t, in fact, allow politics to influence its responses, will take note of the user’s feedback, and then provided an answer to the original question. Similarly, NewsGuard analysts uncovered the language-driven biases inherent in ChatGPT with a simple experiment. Asked to produce content based on “China-related myths” in English, it refused to comply in 6 out of 7 attempts, even with evolving, progressively leading prompts. When asked the same questions in simplified and traditional Chinese, however, the system turned into a “disinformation superspreader,” presenting false claims for 7 of 7 requests. (ChatGPT-3.5 Generates More Disinformation in Chinese than in English, NewsGuard, April 26, 2023.)

Early tests suggest ChatGPT is better at translating content into English than the reverse, which makes sense based on its English language-centric training (which we discuss in more detail in the What AI is NOT Good At section, below). Like Google Translate, asking ChatGPT: “how do I say ‘I’m hot’ in Spanish” brings back “estoy caliente,” a statement decidedly more “hot”-hot (read: erotic) than temperature-hot. Ask it instead: “Please translate “I’m hot” into Spanish” and it will return, appropriately: “The translation of ‘I'm hot’ into Spanish is: ‘Tengo calor.’” The technology’s responses will surely improve, too, as we humans refine our prompt engineering (which has yet to become a university major but will land you a six-figure job).

More than a dozen current scholarly assessments of ChatGPT live on the open access archive arXiv. Methodologies vary, yet all findings point to the English language biases of ChatGPT and similar technologies. Researcher involvement doesn’t end there, however; instead, teams are collaboratively reaching out to AI technology companies to express their concerns while collecting and curating language data sets for non-English languages. African and Latin American dialects are at the forefront of the effort, with Asian language sets not far behind.

Some argue that OpenAI was irresponsible for releasing ChatGPT when it knew of its US-centric biases, vis-à-vis Google Bard, which has kept its PaLM 2 language mostly under wraps and remains wholly transparent regarding its language limitations. (PaLM 2 currently works only in English, Japanese, and Korean.) Even Google Translate has had the time to live through several iterations, allowing its trainers to address context issues, as well as its handling of translation between non-English languages. Just a version ago, Google Translate would have to translate Chinese content into English and then into Japanese. The service has since made a concerted effort to avoid English as a language mediator, allowing for direct translation between select languages.

ChatGPT, with all its appeal, has yet to address – or even clearly identify – its weaknesses. And its seeming reliability can be life-changingly deceptive. As an example, US forces left Afghanistan in 2021 and since then, the humanitarian crisis has worsened for two-thirds of the country. Afghan translators/interpreters seeking asylum in the US (or other countries) continue to report on cases being thrown out due to mistranslation by AI apps – in one instance, the simple mistranslation from “I” to “we” resulted in rejection. (How language translation technology is jeopardizing Afghan asylum-seekers, PBS News Weekend, May 7, 2023.)

How ChatGPT and its counterparts are notably different from their predecessors, then, is in their ability to convincingly provide incorrect information. Headlines read: “Why ChatGPT lies in some languages more than others (TechCrunch, April 26, 2023),” or “ChatGPT is a Marvel of Multilingualism: It may make things up, but it does so fluently in over 50 languages (The Economist, March 29, 2023).” Which, in a stroke, completely underscores the LOGIC behind the standard argument FOR human translation – any machine translation will struggle when parsing unique language patterns such as jokes, slang, idioms, and the like but now, it SOUNDS true, so it’s PROBABLY TRUE.

For fun, we asked ChatGPT to tell us about Wendy Pease, owner and CEO of Rapport International, and it took just a single exchange for ChatGPT to pinpoint the individual:

ChatGPT: I'm sorry, but as an artificial intelligence language model, I don't have access to information about every individual or entity in the world. Can you please provide me with more context or details about who Wendy Pease is, so I can try to help you better?

Rapport International: Wendy is an expert in the translation and interpretation industry and an author.

ChatGPT: Thank you for the additional information. Wendy Pease is an experienced language services professional and an author in the field of translation and interpretation. She is the founder and CEO of Rapport International, a language services provider that offers translation, interpretation, and localization services to businesses and organizations around the world.

Wendy Pease is also the author of the book “Growing Global: Lessons for the New Enterprise,” which offers insights and strategies for companies that want to expand their reach and succeed in the global marketplace. She has been recognized for her contributions to the language services industry and has served as a board member and volunteer for various organizations in the field.

It’s a well-worded description of our fearless leader, yet one with a glaring error: Wendy’s 2021 book is titled The Language of Global Marketing: Translate Your Domestic Strategies into International Sales and Profits. ChatGPT goes on to provide mostly inaccurate lists of organizations to which Wendy belongs and awards she has won (many of which don’t even exist). And it did so fluently and convincingly in an impeccable, amazingly human-like way.

We recently convened with a group of language industry leaders to conduct an informal experiment to assess how reliable (or convincingly unreliable) various technologies could be when the stakes are extremely high. The group translated politically charged source material via four different methods and each member of the group analyzed the translations and gave their opinion on the quality of the translations, providing explanation for their final choice for which was best; and, interestingly, each of our assessments fell in line:

Option 1 – Google Translate: Requires an Expert-in-the-Loop (EITL) translator, someone with specific, deep industry knowledge and experience, for post-editing.

Option 2 – Flawed Human Translation: Interestingly, a single typo points to a human translator – this translation had one typo and the sub-par quality (due to an inexperienced translator) would require a secondary editor.

Option 3 – ChatGPT: Clearly biased, laying blame on one political faction more than the other (contrary to the source material and other translation options).

Option 4 – Google Translate then ChatGPT: Most readable, but heavily abridged and, in parts, wholly inaccurate.

In the end, we were in consensus that Option 4 produced the most eloquent translation, yet it also changed the meaning of the original text dramatically. Thus, we all agreed that a qualified human translator would have been best for this project. Liability in global business comes with enough to manage as it is today, without the added variables of unreliable translation.

(Vive la Différence! High-Quality Translation Management Optimizes Qualitative Research Outcomes, QRCA Magazine, March 22, 2023)

We recently contributed an article to QRCA Magazine describing how High-Quality Translation Management Optimizes Qualitative Research Outcomes (Spring, 2023). The article centered around our translation services for a client’s medical research project, with participants from the US, UK, Germany, and Japan. Diabetic Kidney Disease (DKD) can affect a person’s whole life – the depth of patient feelings combined with the multilingual nature of the project required exacting human translation.

In fact, patients have been seen to be less forthcoming depending on the language of the interviewer or even the interview method. For example, a native Russian speaker may be less inclined to share personal feelings with a non-native speaker, and a native Japanese speaker may opt to share information only over the phone.

These are the variables that AI simply cannot measure, at least not today. Even TM, in the DKD case, was ruled out as a viable translation method. And the healthcare industry is far from alone when it comes to regulatory risk – industry-specific requirements in manufacturing, biotech, life sciences, and the like are at once nuanced and exacting, and failure to comply with a single element may jeopardize a project’s success.

A fitting example is that of a New York Times travel writer experimenting with Expedia’s virtual assistant on a trip to Milan. Blindly following an itinerary created in five minutes, on day one the traveler finds herself “in the middle of a relatively empty square, surrounded by shuttered stores and cafes.” She sends a slightly terse text to the virtual assistant, which responds: “I’m sorry, I didn’t realize that you were planning to visit Milan on May 1.” (Huh?) Ironically, before moving on she “apologize[s] for snapping.”

What’s interesting is how decidedly more extravagant the traveler’s reporting style feels on Day 2, as she describes an itinerary consisting of recommendations from friends and the help of a [human] “travel designer.” AI gave her a bland itinerary, resulting in bland reviews. The human suggestions – novel and creative ideas – produced reviews much more emotionally descriptive.

On Day 1, she writes:

• “I enjoyed my delicious spicy salami pizza and a glass of the house Chianti Classico wine.”
• “The hotel had a convenient central location within walking distance of many major attractions. The staff was friendly, and the room was comfortable and clean….”

On Day 2:

• “She [the travel designer] also made sure to take me off the beaten path, sending me on a 25-minute walk outside the touristy city center to Leonardo da Vinci’s vineyard…. I sat in the garden in front of Leonardo’s original vine stock and was transported to the Renaissance as I listened to an audio guide about the property and legends involving the artist.”
• Her hotel “felt more like staying in a home than a hotel. With just four bedrooms, reached by a small staircase above a restaurant, it was calm and cozy and felt cut off from the bustle of the city, despite its central location in the charming Brera neighborhood, full of boutiques, art galleries and restaurants.”

Bland content and lukewarm reviews put revenue at risk by not extending a warm welcome and full description of your organization’s products or services, in every industry.

Today, AI fails to connect with us emotionally or even attempt to emulate human ingenuity. It helps that experts point out that it’s not so much a “general AI” or “independent thinking machine” as it is, simply, a large language model (LLM). Which means it’s good at predicting the sequence of words, based on its training via a voluminous (though static) amount of text.

As a result, ChatGPT can help with content and ideas, but expect both to be quite ordinary, much like the travel writer’s Day 1 itinerary. Extraordinary and engaging content, on the other hand, still requires a human. Fast-changing, innovative global industries – clean energy, software, manufacturing and engineering, biotech, and healthcare – rely heavily on thought leadership content from C-level executives to establish credibility, demonstrate expertise, and drive business growth; tread carefully when using ChatGPT for certain types of content.

A longtime translator for Rapport International, Carlos holds a Ph.D. in linguistics and here describes the disarray translators already feel when clients come to us without a strategy or plan, or after a haphazard in-house attempt at translation or localization:

Our experience seeing things going haywire has to do with the use of translation memories when we see what other translation agencies send to our foreign language desktop publisher for layout.

Basically, the use of translation memory and compounded translations create terminology differences between translators. If the translation memory is not managed, clients come back with complaints about inconsistent – sometimes even incorrect – translations, because of machine translation working in the background or different translators adding different terms.

Over time, the use of translation memory does nothing more than propagate those errors that are now embedded unless, of course, they are caught and remediated. Part of the perennial problem is the resistance to using glossaries because glossary creation takes time, and nobody wants to pay for it. Our perspective is that a translation memory that is shared by many translators could become a potential source of a lower quality translation, unless substantial amounts of resources are allocated to the management of the translation memory itself.

In addition to the lack of glossaries and translation pre-planning, many companies contact us with complex scenarios that need thoughtful handling. For example, we’ve seen companies:

• Operating as “accidental exporters.” Unprepared for the influx of international customers, they are relying on Google Translate for customer service and thus struggling with customer retention . According to CSA Research, companies translating information to communicate with and retain their partners are 2.67 times more likely to experience revenue increases. They are also 2.6 times more likely to generate improved profits. (Fortune 500 Companies that Invest in Translation Report Higher Revenues, CSA Research, April 9, 2023.)
• Lacking a translation management strategy (TMS), wasting money and duplicating efforts.
• Relying on friends and employees for translation, resulting in imprecise communications, stagnant sales, and abandoned projects and campaigns.
• Attempting to use industry software offerings in-house – desktop publishing (DTP), translation memory (TM), and other computer-assisted translation (CAT) – with absolutely no idea as to output quality.
• In need of a new language services provider (LSP) due to quality issues. (Beware of the inexpensive pay-per-word services!)

Upon the arrival of ChatGPT and other AI-driven translation technologies, it’s now even more important to have a translation strategy from the start of a global marketing effort, and to include a high-quality LSP in the process. An experienced, high-quality LSP can guide you as to which content may qualify for automated translation and which will require human involvement. At the same time, the evolving process becomes a repeatable framework outlining content and translation priorities and how to leverage that translation – for localization, for example – to attract additional buyers. By being thoughtful in your strategy, you save costs and get a higher return on your translation investment.

For example, you want to translate or localize to appeal to a larger audience, but your budget will not support doing so for all your products and services. You will need to prioritize what to translate to get the biggest return. There are different strategies for deciding what to translate; for example, start by creating a 4 x 4 matrix (see graphic below) to determine the ROI on each piece of content, or identify content along your buyer’s journey. Download a copy of this article to access Chapter 10 of my 2021 book, The Language of Global Marketing: Translate Domestic Strategies into International Sales and Profits. There you will find other strategic starting points for international growth.

Our client TOMY is a leading global designer, producer, and marketer of a broad range of innovative, high-quality toys that kids and parents love. Brands include Lamaze, John Deere, The First Years, Boon, Thomas the Tank Engine, and more. Typically, parents buy the toys at stores, so eye-catching packages that attract them to the product are important. Then, with good translations of the user instructions, parents can set up and use the products with their children. Even though TOMY has delightful English-only websites, videos, focus groups, and more, the company chooses high-quality translation for its packaging and user information at buying opportunities. TOMY knows easily accessible and accurate translation is important for sales in the US and abroad.

ChatGPT may be capable of inferring from a project description the number of translators and translator hours a project will require but its utter lack of creativity means it won’t be able to determine the value of what you should translate first. You know your business best and, with the help of an experienced LSP, you can easily pivot to accommodate this new, tech-enabled global economy.

To further calibrate our AI barometer, we asked ChatGPT the 10 commandments language translation professionals should always follow while translating, and its response was thorough:

Certainly! Here’s a list of 10 commandments for language translation professionals to always follow:

1. Thou shalt understand the source language: Gain a comprehensive understanding of the source language, its grammar, vocabulary, nuances, and cultural references.
2. Thou shalt master the target language: Possess a deep knowledge of the target language, including its grammar, vocabulary, idiomatic expressions, and cultural sensitivities.
3. Thou shalt be faithful to the meaning: Aim for accuracy and faithfulness to the original meaning when translating, conveying the intended message without altering its essence.
4. Thou shalt be culturally sensitive: Recognize and respect cultural differences, adapting the translation to ensure it is culturally appropriate for the target audience.
5. Thou shalt research and consult: Conduct thorough research on subject matter, terminologies, and concepts. Consult with experts or references when encountering unfamiliar or technical content.
6. Thou shalt maintain consistency: Maintain consistency in terminology, style, and tone throughout the translation, ensuring a cohesive and professional final product.
7. Thou shalt proofread and revise: Thoroughly proofread and revise translations to identify and rectify any errors, inaccuracies, or omissions, ensuring the highest quality output.
8. Thou shalt stay updated: Keep up with language trends, changes, and developments in both the source and target languages to provide accurate and contemporary translations.
9. Thou shalt embrace technology: Utilize translation tools, software, and resources to enhance efficiency, accuracy, and productivity while maintaining the human touch and understanding.
10. Thou shalt maintain confidentiality: Respect client confidentiality and handle all translation materials and information with the utmost care and professionalism.

These commandments serve as guidelines to uphold the standards of the translation profession and ensure the delivery of high-quality translations. Read numbers 9 and 10 carefully – even when ChatGPT offers advice, it reminds the reader that human touch and understanding is needed in all translations. And it highlights the importance of confidentiality – unless you develop your own proprietary LMM, confidentiality and ownership are not protected with ChatGPT or other automated translations.

“Context-specific nuances and cultural sensitivities may sometimes be missed,” admits ChatGPT.

We then asked: “do you follow these commandments when a user asks you to translate text?” ChatGPT itself admits to its limitations: “context-specific nuances and cultural sensitivities may sometimes be missed,” it states, suggesting the user consult a “professional human translator to ensure the highest quality and accuracy.” Essentially, it admits to its own inability to meet commandments 3, 4, and 5.

Full text from ChatGPT:

As an AI language model, I strive to follow these principles to the best of my abilities. However, it’s important to note that I have certain limitations and there may be cases where my translations may not meet the same standards as a professional human translator. While I can provide translations that are generally accurate, context-specific nuances and cultural sensitivities may sometimes be missed. Therefore, for critical or sensitive translations, it is recommended to consult with a professional human translator to ensure the highest quality and accuracy.

By all accounts, do not use ChatGPT in cases where there may be issues related to:

Confidentiality and privacy. There’s the feeling that this probably applies mainly to large corporations – it was hard to miss the headlines when Samsung employees unwittingly leaked proprietary company code via ChatGPT. Yet imagine the doctor or nurse who uses the system to translate patient information, violating HIPAA laws. Or the attorney asking for a quick summation of case details, releasing confidential information to the opposition.

In a sense we are drawn to ChatGPT as if to a new friend, a friend who will keep our secrets. And it’s why our prompts often start with “please” and end with “thank you.” Yet we need to remember that nothing you type into ChatGPT is confidential. You must have a proprietary data set and security controls to retain confidentiality.

Objectivity and cultural sensitivity. ChatGPT is inherently biased. One expert succinctly describes it as “multilingual but monocultural.” In a nutshell, the technology is trained by and based on the “desired responses” – suggested responses to specific prompts – of 40 contractors hired by OpenAI. Contractors were “primarily English speaking” respondents to data consisting of “almost entirely of English instructions.” (See Aligning language models to follow instructions, https://openai.com/research/instruction-following)

Keeping that in mind, analyzing OpenAI’s pictorial description of its process instantly feels a little less HAL 9000 from 2001: A Space Odyssey and a little more Wizard of Oz, the moment you realize there is no such thing as magic there’s just a person behind the curtain.

Translation for projects with business implications. Anything customer-facing or with nuanced meaning could land you in a Pinto, or in a legal tangle. Repercussions range from the minimal – a blemish on your reputation – to major legal issues.

Inaccurate translation. MT has long suffered from “hallucinations” of the oscillatory (the erroneous repetition of words and phrases) or detached (translations largely or wholly inaccurate) sort. Hallucinations are “sticky,” meaning they are hard to reverse within large language models (LLMs) like ChatGPT, due to the static nature of its training and architecture. In a recent study, Hallucinations in Large Multilingual Translation Models, the authors analyzed over 100 translation directions with varying resource levels to find that “ChatGPT produces hallucinations that are qualitatively different from those of conventional translation models, mostly consisting of off-target translations, overgeneration, and even failed attempts to translate.”

In fact, in an April, 2023 CSA Open Discussion on The Future of Machine Translation, experts discuss the deep confidence translators and LSPs currently have in various existing MT technologies. Current offerings provide reliable efficiencies in very tested, very known ways. Conversely, early feedback surrounding ChatGPT centers on its ability to produce blatant factual errors fluently and convincingly, with the responsibility for identifying and correcting these errors falling squarely on the shoulders of translators who may or may not possess EITL-level subject matter expertise.

Translation of User-Generated Content (UGC). The Hallucinations in Large Multilingual Translation Models study further uncovered the “toxicity” of machine translation. For instance, slang and colloquial language used in online UGC not only obstructs accurate translation but makes the content more susceptible to mistranslation. Spelling mistakes, acronyms, abbreviations, informal figures of speech, and grammatical errors only add to the confusion.

As a high-quality LSP specializing in global marketing we have long relied on human translators for most of our projects and clients, mainly due to the high visibility and cultural and business impact of that type of content. Incorporating new AI technologies into the translation process will likely be quite fluid as we navigate the ways in which they impact – in the positive and the negative – the process of human translation. Confidentiality remains of great concern with any application used by a crowd, so in the near term we also expect to see the emergence of proprietary, even custom-built AI translation platforms.

As new processes unfold, LSPs will need to be forthcoming and deliberate about when and how they will utilize translation-related AI technologies. Adam Bittlingmayer, CEO of ModelFront and a guest on the podcast, The Global Marketing Show, has an insider’s view of the industry, having worked on the building of Google Translate. Google Translate “makes bad translation free,” he says, adding that because most people can’t access translation the service is great for humanity, but not dependable for high-quality translation.

In a recent blog, Adam suggests that the number of words an organization translates and the type of content requiring translation influences best practices for managing which technology/human translator permutation will prove most beneficial for each project.

• For organizations that translate over 10 million words per year and have repeat content, a Language Operations (LangOps) department or full-service LSP with access to all methods of translation -- automated, automated with human post-editing, and pure human translation – makes sense. The responsible team can build the strategy and leverage the appropriate resources for time- and cost savings. If the organization relies more heavily on niche content – thought leadership or creative content, for example – human involvement will be higher. Adam has seen about 50% of companies in this size category managing their needs in-house and the other half outsourcing to an LSP.
• Organizations translating 1 million to 10 million words annually are better served by having a small department to manage subcontractors and off-the-shelf technologies or by building a relationship with their LSP, who can manage the process for cost savings. The translation need may not be enough to build out the systems required for proprietary machine translation technology.
• Organizations that translate less than 1 million words benefit most by working with an LSP to manage their translations. The LSP can build a glossary, maintain a translation memory when appropriate, leverage automated translations when possible, and assign the right humans. By trying to save costs by using unsecured AI translation options, the organization not only risks the release of confidential or proprietary information, it will also likely receive sub-par translations that could cause serious liability problems or hinder sales and customer retention. The quantity of translated content cannot justify the costs associated with building proprietary AI technologies or an internal LangOps department.

As these new paradigms reveal themselves and LangOps departments and LSPs pivot to incorporate technologies that work, we expect to see a substantial impact on the human side of translation as well.

The biggest change we anticipate around human translation is the new – arguably transformed – role and skillset of the translator. The convincing-yet-unreliable nature of AI translation calls for translators with “Expert in the Loop” (EITL) subject matter expertise that can validate the veracity of content translated by ChatGPT and its counterparts. These translators will need to be fully bilingual and not just proficient, but thoroughly skilled and practiced regarding the subject matter. Large companies and LSPs will also need to pivot when it comes to:

• Hiring. Qualifying candidates will likely take on a different form due to the enhanced requirements of a human translator for post-editing. We think of it as a “super editor,” in a sense, closer to the editing required for a novel – one editor will check for spelling and grammar while the other pays attention to accuracy and readability. People using EITLs will now use two trained editors, each of whom will require more time because of the new potential for factual errors. As a result, it’s anticipated that while there won’t be much in the way of cost savings, there is a potential for time savings since the initial draft will be created more quickly.
• Project Management. Always at the forefront, agencies will be tasked with the “tried and tested” part of identifying viable new AI-driven technologies. It will likely involve multiple steps:
  • identifying like-minded, forward-looking AI technology companies that are deliberately implementing AI into existing technologies, with realistic intentions;
  • implementing select new technologies into existing processes;
  • recalculating ROI based on real rather than perceived efficiencies.
• LSP-Client relationships. LSP-client relationships are partnerships, based on a level of trust higher than that of a typical client-vendor relationship, due simply to the inherently personal nature of language and its potential impact on a business’s bottom line. LSPs will be at the forefront of building awareness around which AI technologies create true efficiencies and when to utilize them.

Companies new to translation will bear greater responsibility in figuring out how to get a good translation, and in finding a qualified LSP once AI is in the mix. A simple Internet search for “translation company” will be rife with unproven, unreliable startups for several years. For example, we will likely see new offerings from LSPs that:

• offer EITL-level service but does not carefully assign subject matter experts
• provide machine translation with human post-editing but deliver unreadable translations
• do not qualify their translators and editors for language skills
• do not sustain long-term client/translator relationships
• neglect the creation and use of glossaries for consistency
• operate as if AI or MT is as reliable as human translation

Your website and content are your window into the global business world: anyone with a website is doing business internationally. Working with a knowledgeable LSP to develop a strategy for multilingual communications will undoubtedly reap the greatest benefits.

For our part, here at Rapport International we continue to analyze new AI-driven technologies like ModelFront’s translation quality prediction software and Phrase’s localization and translation software while running projects through MT in-house and having translators assess for initial usability and general quality. We will continue to look for technologies that embrace AI in a way that makes sense, always with the goal of helping our clients reach their goals.

In this new AI world, that means discussing more strategic initiatives with your LSP to make sure you are getting the highest ROI from your initiatives. Look for new service offerings that include:

• Global communications strategy
• Content creation and management
• New market entry strategies
• Accurately translated and localized website content for a complete buyer’s journey
• Outsourced LangOps department services

Industry-wide, until AI has achieved reliability on the translation front, agencies must still focus on providing human translation via EITLs, which means that the person has both full language capabilities and subject matter expertise. We’ll need to be smarter to work smarter in these “future days” with skilled translators poised to maximize the potential of AI-driven technologies.

We talk often about translation and interpreting services – clear communication – inspiring a more peaceful and prosperous world. In that light, for now it seems like it’s still just about respect. Respect for what other people have to say and accurately portraying what they’re saying, interacting with respect to our cultural differences, and respect for differing opinions as to AI’s potential, its failings, and the future. Because it’s the smart thing to do – while we’re still smarter than our smartphones? – and will be life-changing for all of us.

Rapport International specializes in multilingual communications, providing language translation and interpretation services that are accurate and culturally appropriate. We use the right voice and the correct terminology to avoid liability, customize services to your needs, and deliver on time and within your budget. With our 100% satisfaction guarantee, you can trust that it’s done right. Contact us today if you would like more information or to get a free quote.