How AI Enables Automation and Reshapes Jobs: Insights from the Localization Industry
How AI Enables Automation and Reshapes Jobs: Insights from the Localization Industry
AI advancements in automation and translation are reshaping industries, impacting jobs, and evolving the role of human translators in AI-driven futures. Plus an interesting reading and video...
Ciao,
In February, Radical Curiosity welcomed 50 new members 🎉. In this edition, I share insights into my ongoing exploration of content localization and the significant influence of artificial intelligence on this sector. It's important to note that the opinions expressed here are my own and may not represent the views of Translated, the company I am currently collaborating with.
Table of content
How AI Enables Automation and Reshapes Jobs: Insights from the Localization Industry
How startups beat incumbents (A Smart Bear)
Aaron White Dives Deep on Navigating AI's Horizon: Future-Proofing Products in the Age of AI
How AI Enables Automation and Reshapes Jobs: Insights from the Localization Industry
In the near future, advances in artificial intelligence (AI), particularly in generative AI, will significantly enhance the automation of business processes. This evolution will speed up existing automation practices and introduce entirely new categories of automation. The increasing adoption of AI tools is bound to affect the labor market. Consequently, it raises critical questions: How many jobs will be eliminated, and at what pace? Furthermore, will new job opportunities emerge to replace those that are lost?
Historically, each new wave of automation has led to the disappearance of entire categories of jobs while simultaneously creating new ones. It’s relatively straightforward to observe the elimination of jobs. However, forecasting the nature of new jobs and where they will emerge around the globe is far more challenging.
To understand the impact of these changes, it’s essential to consider both the tasks that can be more easily and safely delegated to artificial intelligence and the previously unfeasible automation types.
For the past year and a half, as a product leader at Translated, a language service and technology provider that localizes content for major tech companies like Airbnb, I’ve engaged deeply with the industry’s evolving landscape. A recurring theme is the impact of machine translation: many translators are worried about the prospect of job loss in the near future and the shift towards their work becoming predominantly the revision of machine-generated translations.
I’ll use the translation industry as a case study to illustrate the dual impact of artificial intelligence—both its destructive and generative effects on an industry. Two brief overviews will be beneficial for those not well-versed in this field. The first will cover the current state of machine translation technology. The second will provide insight into why content localization is complex, highlighting the intricacies involved.
Current Advances in Machine Translation Technology
Currently, two technologies can be used to automate translation: Neural Machine Translation (NMT) and Large Language Models (LLMs). NMT is the technology behind popular services like Google Translate, DeepL, or ModernMT and is regarded as a mature technology. These systems produce sufficiently accurate translations for various applications, though the style often lacks elegance and can be flat. Moreover, NMT is notably cost-effective; for instance, ModernMT offers translation services at $15 per million characters.
Large Language Models (LLMs) such as ChatGPT, Google Gemini, Anthropic Claude, and Meta LLaMA have seen rapid development in recent years. Consequently, there’s been considerable interest in leveraging them for translation tasks. According to the studies I’ve encountered, these models perform on par with Neural Machine Translation (NMT) systems, albeit at a significantly higher cost. Therefore, they are not yet deemed cost-effective for translation purposes. However, it’s important to note that LLMs are still in the early stages of development. It’s pretty plausible to anticipate that, within a few years, they could surpass NMT in terms of translation quality while becoming competitively priced compared to human translation services.
How content localization works
Typically, a professional translation of high quality involves the collaboration of two individuals: one person translates the content, and a second person reviews the translation for accuracy. In addition to these translators, a team of professionals is tasked with various responsibilities, including:
Planning and managing a company’s localization efforts, determining which materials to translate and into which languages.
Conducting quality assurance to ensure that translations adhere to the company’s style guides and glossaries.
Overseeing a technology infrastructure that can be particularly intricate for large companies, mainly when translating software and mobile application interfaces.
Linguists today utilize various tools, including Translation Management Systems (TMS) and Computer-Assisted Translation (CAT) tools. These tools increasingly integrate artificial intelligence, providing linguists with suggestions automatically generated by machine translation engines, such as Translated’s ModernMT. As a result, translators rarely start a translation from scratch. In most instances, they refine and edit a piece produced by machine translation or a sentence retrieved from a translation memory.
The quality of translations generated by artificial intelligence has seen significant improvements in recent years. Consequently, human interventions have become minimal, leading many companies to bypass the initial human translation phase and opt solely for Machine Translation Post-Editing (MTPE) services. This shift offers two key benefits:
It substantially lowers the costs associated with localization.
It enhances machine translation quality over time, as human corrections are utilized to refine the AI models.
It may seem paradoxical, but today, when linguists translate or correct content, they contribute to improving machine translation capabilities. This iterative enhancement process will persist until companies deem machines sufficiently reliable, rendering human oversight unnecessary.
In the future, will translators become obsolete? There will likely still be a need for translators, but the demand will decrease compared to today. More significantly, the nature of their work will evolve, predominantly involving the supervision and training of machines. A select group of multilingual linguists and writers may still undertake the translation of novels and similar works, but company-generated content is expected to become largely automated.
However, artificial intelligence also presents new opportunities. Notably, there are at least two major types of textual content that would be impractical to translate without machine assistance: user-generated content and real-time chat conversations.
Today, if you make a reservation on Airbnb and need to chat with your host, you can write in your language, and what you write will be translated into the host’s language in real-time, and vice versa. At the same time, it’s only through machine translation that Airbnb can translate all its listings and reviews into more than sixty languages.
Similarly, Glovo utilizes ModernMT technology to translate restaurant menus into multiple languages automatically: Taste the World: How Our New Machine Translation Feature Transforms Your Ordering Experience.
Artificial Intelligence and Localization
I’ve outlined my predictions on how artificial intelligence’s impact will unfold over the next five years, differentiated by content type. The terms I’ll use are as follows:
Human Translation (HT): This is the highest quality and most costly service, requiring the involvement of two translators.
Machine Translation Post Edit (MTPE): This involves content initially translated by a machine, which a linguist subsequently reviews.
Statistical Machine Translation Post Edit (S-MTPE): Similar to MTPE. The review process is applied only to a sample of the content translated by the machine. The machine performs this sampling, assigning a confidence score to each translated sentence and highlighting those needing review.
Machine Translation (MT): This process involves no human intervention.
Real-time chat
Machine translation could facilitate communication between individuals who do not share a common language, significantly impacting various digital platforms. This functionality is particularly relevant for marketplaces like Airbnb, collaboration tools, and any application featuring an integrated messaging system. It’s expected that multilingual chat features will grow in popularity.
MT is the only possible solution for these applications due to its cost-effectiveness and the rapid pace at which translations must be delivered. Moreover, as these interactions involve private conversations, any form of human oversight and post-edit is deemed inappropriate.
Customer support
Machine translation is set to enhance customer support capabilities by enabling assistance in the customers’ native languages. This advancement is primarily fueled by the deployment of Large Language Model (LLM)-based automated response systems, such as ChatGPT and Google Gemini. These systems already support multiple languages, although their proficiency in English surpasses that in other languages. It might be practical to employ machine translation for converting text to and from English, coupled with an LLM, to interpret questions and formulate responses. This strategy also offers the benefit of focusing the training of the Large Language Model (LLM) on a single language rather than dispersing efforts across multiple languages.
MT remains the go-to solution in this context. LLMs are designed to produce precise, unambiguous text, reducing or eliminating the necessity for human revision.Â
User-generated content
Machine translation is a vital tool for handling user-generated content, given the vast amounts of data involved that would be unfeasible to translate manually. Beyond the previously mentioned examples, numerous social media platforms have started offering real-time content localization.
For such content, opting for S-MTPE over MT is preferred to enhance translation quality within the specific context of a platform. This is crucial because all communities tend to develop their unique dialects and expressions, which may not have any significance outside of their context.
Knowledge base and manuals
HT is predominantly used to translate this content, ensuring high accuracy and linguistic quality. However, for some companies needing to localize vast amounts of material into multiple languages, MT becomes a viable, cost-effective alternative. This approach not only helps reduce expenses but also broadens the range of languages that can be supported.
For handling such content, S-MTPE and MTPE emerge as ideal solutions. Budgetary considerations can influence the choice between the two. For instance, a company might prefer to review all translations thoroughly (MTPE) for widely spoken languages. Conversely, they might opt for selective spot-checking (S-MTPE) for less common languages to manage costs effectively.
E-learning
E-learning represents a sector where artificial intelligence (AI) serves a dual purpose. First, similarly to knowledge bases, machine translation can be employed to decrease localization expenses and/or increase the number of languages a piece of content is available in. Second, generative AI introduces new possibilities. For instance, synthetic voices can be used to dub audio and video content at a significantly lower cost than professional dubbing services, yet delivering more satisfactory quality.
Marketing and communication
For marketing and communication content, the use of HT or, at a minimum, MTPE remains necessary. In this field, it’s not just about accurately translating content; it’s crucial that the text is engaging to read and that linguistic nuances are either preserved or appropriately adapted when transitioning from one language to another.
Specialized contentÂ
Specialized content typically still necessitates the expertise of human translators (HT). It’s challenging to envision relying exclusively on machines to translate complex documents such as contracts or scientific papers. However, there are certain areas where machines could excel, particularly in scenarios involving highly structured language and specialized vocabularies designed to minimize ambiguities, such as in medical diagnoses.
Conclusions
If my analysis holds, artificial intelligence is poised to effectuate the following impacts on the localization industry in the forthcoming years:
Companies will increasingly opt for machine translation over human translators to cut costs and broaden the languages covered. This shift will likely result in diminished job prospects for translators.
Machine translation will be strategically employed for content where a compromise on language quality is deemed acceptable to broaden reach. The industry will swiftly devise strategies to determine the minimal level of human proofreading (S-MTPE) necessary for each type of content.
Services offering multilingual chat capabilities will rise.
Social platforms enabling real-time translation of user-generated content will become more common.
The role of translators will evolve towards more oversight of machine-generated work.
Synthetic voices will pave the way for cost-effective audio and video content dubbing, thus birthing a new market for synthetic dubbing that doesn’t exist.
Generative AI will facilitate the creation of personalized, on-demand content. For instance, leveraging a service’s knowledge base to produce customized tutorials for utilizing specific features will become straightforward. This content will be automatically translated, and the oversight of the generated content and its translations will be easily conceivable.
How startups beat incumbents
A startup can beat a large, successful incumbent if it does things the incumbent can not or will not do. Here are those things:
Taking risks that cannot be quantified
Addressing a profitable niche
Doing delightful, valuable things that don’t scale
Unsurpassed customer service
Leveraging new technology
Having an opinionated personality
Doing things that aren’t zero-sum
Being worse-but-acceptable in most dimensions
Being low-cost against a profit center
Read the full article on A Smart Bear: it is very interesting!
Aaron White Dives Deep on Navigating AI's Horizon: Future-Proofing Products in the Age of AI
Aaron White and hosts Fareed Mosavat and Brian Balfour dive into how AI is changing the game in product development. They discuss the big picture of AI’s role across industries, stressing the need for builders to always look ahead and aim for the big wins, not just small efficiency gains.
Aaron breaks down AI in product design into three main roles: the co-pilot, the human as the guide, and the AI that takes the lead. He points out that while many companies are good at using AI to make us more productive, the real breakthroughs come from profoundly understanding and creatively solving specific challenges. The talk also covers how AI transforms jobs that rely on knowledge, especially in how it deals with routine tasks versus exceptions, suggesting that humans should step in mainly for those exceptions.
The interview was published in two episodes on Reforge:
If you were forwarded this email or if you come from a social media, you can sign up to receive an article like this every Sunday.Thanks for taking the time to read this episode of my newsletter. I hope I’ve been helpful. If you think my sketchbook might interest someone else, I’d be glad if you shared it on social media and forwarded it to your friends and colleagues.