---

Humans are always curious about how to make our lives easier. Two years ago, GenAI appeared on the scene. Given that the technology (i.e., LLMs) has “language” in its name, it’s no surprise that our industry was eager to adopt it. Today, the initial wave of astonishment has gradually given way to moderate acceptance.

2023 may go down as the Big Year of GenAI Research and Experimentation. Experimenting with GenAI opened a new realm of possibilities. Still, more importantly, it helped identify risks associated with its uncontrolled use and the limitations and challenges of rushing blindly toward implementation.

Despite our collective and often highly collaborative efforts, it is fair to say that we are still far from having figured out AI. For every solution we seemingly find, a new big question appears. It’s not all doom and gloom, though — far from it. In pursuing sustainable AI adoption, we must consider the role of humans.

What is humans’ place in the brave new AI world?

A glance at your LinkedIn feed or into the “Other” Outlook folder is enough to reveal that, along with news of yet another breakthrough in AI research, the impact of AI on human activity and roles is one of the most frequently discussed topics. Personally, I am very optimistic. For me, it’s no longer about “humans in the loop.” With AI, it’s all about “humans at the core.”

Now, and until the next major leap in technology development, AI will only be as good (or imperfect) as its creators. Humans drive technology development, not the other way around. AI is the result of human ingenuity. It can address industry challenges such as scale, speed, and resource limitations. Yet, AI still relies on us to make sense of what it produces. For the time being, to think is a uniquely human capability.

In the field of quality management, which is closest to my heart, AI presents an opportunity to elevate traditional quality management concepts within the localization workflow. This is a shift from analytic, atomized quality gatekeeping to end-to-end quality management. Instead of focusing primarily on counting errors in the translation output, we start by exploring the key business and production attributes that define product quality. Questions like “Who is the user, and what’s important to them?” and “How do we balance risk, budget, and timeline?” become central. From there, we work backward, breaking down the use-case-driven production workflow into detailed tasks, ensuring that humans are engaged only when necessary to address challenges that arise because AI is “just” a technology.

The key to success lies in how precisely we can define and orchestrate the most appropriate mix of quality management and control activities performed by AI and humans throughout the entire localization workflow to ensure the desired product quality. We see buyer-side organizations and service providers investing thousands of hours and dollars into experiments to optimize AI capabilities, often arriving at similar conclusions: A human touch is needed to produce contextually and emotionally appropriate content — whether during translation or as part of quality evaluation. Only humans can connect the dots that matter in a given situation, determine what needs editing or tweaking, prepare the right corpus, or formulate the most effective prompts. Even how production workflows are set up is an astonishing symbiosis of humans and technology.

Today, humans shape AI’s potential and determine its trajectory. Only with humans at the core can humans’ needs and expectations truly be met.

If your logic is to automate and streamline with the help of AI — and the business logic driving these decisions is ironclad — then the next question becomes:

When does the human come into the mix?

As a strong advocate of the humans-at-the-core principle, my answer would be: always. This is a bold statement, especially when everyone strives to deploy AI to achieve a no-touch content delivery process for creating a product that fully resonates with its users.

Although we’re more down to earth these days, we still encounter unrealistic and faulty expectations regarding AI-powered solutions. There remains a lack of understanding about what underlies AI technology, the prerequisites for integrating AI into the localization workflow, and so on.

Researchers and experts working with LLMs are familiar with concepts that we, as general stakeholders or linguists, tend to overlook: Technology doesn’t think; it generates output based on how it has been trained. Transfer learning allows for rapid and extensive information sharing, but the training inputs are always broken down into very specific pieces of information. When humans train others — by sharing experience and knowledge — we do so holistically. We rarely break down knowledge into isolated concepts; instead, we share experiences and rely on intuition and the human ability to decode information.

We should remember these points whenever we consider integrating AI and carefully consider the context of AI applications. Take, for example, AI-assisted quality evaluation and the contexts in which it is applied:

• Segment-level scoring: Even trained LLMs produce segment-level scores that are difficult for a professional linguist to interpret. This is because humans assess the correctness of a sentence or segment within the context of the entire document, a specific situation or need, emotional tone, relevant knowledge, and more. However, these automated scores are extremely useful for quickly understanding quality improvements at a system level when training or creating language corpora.
• Error annotation: Despite using very sophisticated prompts, the types of errors identified by LLMs and the explanations provided are often inappropriate or incorrect. The root cause is that LLMs apply the contexts they were trained on, missing, for instance, the emotional nuances that require a different tone of voice, synonym, pun, or other subtleties.

This brings me to the next big question that needs to be asked:

When can you trust the machine?

You’ll note that the question is not whether we can trust the machine. The real question is identifying, on the fly, with the context and knowledge of our situation at any given moment when we can trust the machine. AI can work wonders — creating reams of content, checking it, and generating detailed issue logs with just one click. In the AI world, the traditional “human’s role of producing (linguistic, quality) data” has shifted to “only a human can decide if this data is appropriate for the particular context” or, more generally, “whether and how much to trust this data.”

For AI technology, language is just data. For humans, language is a flexible tool that combines cultural concepts, modality, and context to communicate information in a way that is appropriate and effective for the situation. Tech delivers “output”; humans make it trustworthy and genuine. Humans — end-users — are at the end of the funnel, the ultimate beneficiaries of our expertise. As we’ve touched on elsewhere in this issue of Global Ambitions, there’s also the idea that humans still prefer human-created content. This is another reason to emphasize the human-at-the-core concept and the importance of elevating traditional quality management within the localization workflow.

As I’ve mentioned, quality is a complex concept composed of various perspectives and elements, all contributing to the final result — a memorable user experience. Today, when AI is still something of a “black box,” the only way to ensure quality, in my view, is by defining what quality means at every step of the process until the final result is achieved. This includes the translated word, the quality of the system you’ve built, the quality of the data set it uses, the quality of the people who interact with the system, and, last but not least, the quality of collaboration between those who design and use it. Quality is an inherent component of every step in the sequence, not just during the translation task. It’s the sum of the parts that makes or breaks the result.

Historically, buyers have been deciding about their localization operation (or any business operation!) based on a set of variables — cost, time, quality — and, usually, compromises are involved. This will continue to be how it’s done. The answer to the question of when to trust the machine hinges on understanding the place and role of the human first (back to the idea of human-machine synergy) and comes with a series of smaller questions you will need to ask:

• Which tasks must involve a human in the process?
• What profile is needed for the particular task?
• What are your requirements and expectations — from the system, AI tech, and results that will appear in your product?
• How do you know you’ve done enough to manage risk and make end-users happy?

What pathway is there for humans to evolve?

Evolution is constant. By now, it is abundantly clear that we must adapt to thrive. When reflecting on the role humans will continue playing in the localization operation of the future, it’s evident that our current skills and competencies won’t be enough to coexist with AI.

Take, for example, the role of a translator. Typically, you follow the source text, rely on your intuition, and use context and available reference materials to select the right word for the translation. You do not necessarily pause to consider the sociocultural influences or implications of the word you’ve chosen (if you’re good at your craft, it simply flows). This classic scenario occurs countless times in a linguist’s workday. Now, imagine you’re a linguist tasked with helping to create or edit linguistic data for machine training purposes. This could involve translating very specific and complex concepts or sifting through pages of profanity-laden segments so the machine knows what not to use. You could say this is a non-standard task, but now, with LLMs, it’s calling for a new set of creative or expert skills that linguists will need to develop.

Then there is the relatively new discipline of prompt engineering and the related questions of how and where it fits in the workflow and whether it’s a task for linguists. The traditional linguist, however, is not a prompt engineer (remember the earlier idea of humans and machines “thinking” differently). To effectively guide the system through prompts, you need to be a linguist and an engineer both. Prompting requires a skill set closer to that of a developer who understands the underlying architecture of the language model. As we’re heading towards embedding prompting in conventional CAT tools, it’s clear that our only pathway is upskilling.

And one last question for the road: What does the future hold?

There is a paradox here. With AI, the human is more visible and important than ever before. Technology enabled the democratization of access to content in the users’ native languages and is broadening our horizons. The humans are the centerpiece — we, localizers, do our work for them to help them accomplish their objectives and broaden their horizons. So, in the grand scheme of things, how risky and important are the minutiae of translation errors?

In the long run, AI may support our humanity and multiculturalism more than any individual’s or company’s effort. AI forces us to reexamine our place in the chain and forces us to work on ourselves and our traits — most of all those we didn’t think we had or didn’t think about developing before. This is our pathway to coexisting with AI.