Generative AI is the digital world’s shiny new toy — but what is its real impact on the localization industry? The language services and technology market has seen an AI-fuelled hype cycle a few years ago with neural Machine Translation (MT), and now with generative AI via Large Language Models (LLMs) at the forefront, what can you expect in terms of AI localization today?
There’s certainly overlap between MT and generative AI. One of the consistent emergent abilities of LLMs with enough training material (e.g. OpenAI’s ChatGPT and Google’s Bard) is language translation. These models are trained to generate, not translate. But because of their multilingual data and training methodology, they can translate between certain languages anyway. In the same vein, sufficiently large and complex LLMs can also perform transcreation.
Believe it or not, the renewed excitement for AI-powered localization capabilities doesn’t stem solely from this emergent translation or transcreation capability. Localization isn’t just translation, after all. It’s an entire process and not a simple code switching between two languages. Generative AI — along with other developments in the space — present promising opportunities all across localization as a pipeline.
As Language Inspired CEO Kristaps Lapins himself says:
“AI-driven translation tools are transforming the way businesses handle multilingual content. As AI continues to evolve, it will undoubtedly streamline the localization process for many content types.”
What are the Current AI Localization Capabilities Available in the Market?
At the risk of oversimplifying a complex topic, it would be helpful to break down AI localization capabilities into three broad siloes:
Go into ChatGPT or Bard and you can try it yourself. Prompt it to translate an English sentence into another widely used language (or vice versa) and the chances are it will return an acceptable output. This alone isn’t really that exciting from the perspective of practical application, because it’s neither entirely novel nor state-of-the-art. It’s true however, that generative AI can potentially offer a new translation process and arguably a new pricing structure that’s similar but ultimately different from current standards because of the tokenization of LLMs.
Research has shown that LLMs like ChatGPT can certainly be made competitive. For example, some tweaks to system values allows ChatGPT to deliver “comparable [translation] results to commercial systems for high-resource languages.”
Furthermore, generative AI diversifies the approach to the translation process through its transcreation abilities. Instead of translating an English call-to-action (CTA), for example, into its German counterpart and then adjusting tone and style to appropriate cultural norms, a single prompt to an LLM can generate two completions at once.
This prompt, for instance:
“Write a one-sentence CTA for a digital coupon email subscription service in English and German, taking into account cultural norms for the different target audiences.”
Results in this output when given to ChatGPT:
English: "Subscribe now for exclusive digital coupons and savings tailored just for you!"
German: "Jetzt anmelden für exklusive digitale Gutscheine und maßgeschneiderte Ersparnisse!"
Is it the best translation you’ve ever seen? Probably not. Does it demonstrate the potential for speeding up translation of simple material, shortening pipelines by processing several languages at once instead of waiting for the original and then translating it multiple times? It certainly does.
More startling is that generative AI isn’t really restricted to language. Everyone knows about Stable Diffusion and Midjourney, and the entire copyright controversy resulting from increasingly better AI image generators. There are even early versions of AI video generators available to experiment with today.
Companies like Google have already taken steps towards multi-modal generative AI models that generate and complete not just text, but also images. Bard currently has limited capabilities when it comes to taking images as input, and the paid version of ChatGPT (via the latest Dall-E 3) can generate images as output. Microsoft’s in-preview Designer is essentially an AI-driven Canva, generating entire graphic designs with a single prompt.
So when localizing app assets, for example, generative AI can span not just text but also images. Integrating with additional AI capabilities, it’s a simple task with the right technology to transcribe audio, translate the transcription, and then generate text-to-speech for a fully translated audio clip in your preferred target language.
Administrative, Collaborative, and Concurrent Automation Capabilities
In essence, from a high-level perspective, the current capabilities of AI can truly transform localization not just through a novel method of translation or transcreation, but via bolstering the administrative and collaborative processes involved. Just simply processing multiple languages concurrently shortens the translation process of several languages into literal seconds per page of material.
An entirely new dimension opened up by LLMs is how users interface with them: via natural language. You can literally talk to LLMs and tell them what to do. This makes localization platforms like Lokalise AI, for example, become much “smarter” with how they can approach automated localization with as little human input as possible.
Additionally, interfaces supported by natural language drastically lowers the technical requirements for its users. You no longer need to know what knobs to twist and what sliders to pull to what value. You can literally tell your automated system to make the translation shorter, longer, less formal, friendlier — practically whatever you require.
Naturally, there are still apparent limitations and pitfalls:
Low-resource languages (language pairs without much training data) remain an issue that cannot be resolved without specialized and qualified human translators and interpreters
You can scale much quicker, but without professional human quality checks, you also multiply and magnify your mistakes easier
The management of translation and localization projects may appear to be democratized, but the skills required to manage complex projects certainly are not
If it isn’t obvious by now: more than ever, you need expert human interactions within these automated localization processes.
“This shift also transforms the role of language specialists,” says Kristaps Lapins. “It marks a transition from routine tasks to more specialized responsibilities that demand expertise. Particularly when dealing with intricate or culturally sensitive content, the human touch remains indispensable.”
Generative AI Adoption Trends for Localization
So with all of the above taken into consideration, how far has generative AI penetrated the localization market?
Lapins says “The use of AI in localization workflows is in the starting phase. Clients are discovering what the options are and there are a lot of questions to be answered, still, on this topic.”
In the case of Language Inspired, we are just starting to offer what can be called AI post-editing. It’s essentially an AI-enabled workflow where an LLM such as ChatGPT is the engine for the translations and the human-in-the-loop comes afterward for post-editing. Other companies, meanwhile, lean more towards a combination of MT plus AI. So a subtle but key difference now is that there’s an emerging model where instead of MT and post-editing, there’s generative AI and post-editing.
Lapins explains that as clients discover the functionality, the interest and search for potential use cases grows. “There are requests coming from FinTech, eCommerce, SaaS, small mobile apps — we observe that those companies that have already implemented MT in their workflow are keen on exploring what new AI capabilities are possible,” he says. The primary point of curiosity is what AI does better. Compared to incumbent MT-enabled workflows, is it cheaper, faster, or better in quality?
These are the questions clients want answered. And given the nascent stage of generative AI-enabled localization, a more definitive picture remains to be seen, for the most part.
The Ambitious Potential of AI Localization… but with a Human Touch
So what is the current state of AI capabilities in localization?
AI capabilities in localization are showing promising performance and further possibilities in:
Translation and transcreation
Concurrent processing of what used to be push- and pull-based queues of tasks
AI capabilities in localization have come to a point where it can enable a boom in multilingual content. But hand in hand with that, the complexity in workflows and speed of production require key human-in-the-loop safeguards to ensure quality and prevent mistakes from propagating out of control.
“Finding the right balance between AI and human translation is key to delivering spotless results and ensuring quality assurance,” Lapins emphasized.
But beyond all this, AI capabilities in localization present an opportunity for genuine hyper-localization and hyper-personalization. Generative AI in particular is making it possible for every user to go through individualized, localized experiences instead of generic ones, regardless if they're using a website or an app, or if they’re shopping online or reading a blog.
Think of it this way: ChatGPT has turned web search on its head. Instead of people browsing through top results in Google, they can simply ask a generative AI chatbot for the answer and get a personalized result.
What will drive this promising future forward are technologically forward-thinking platforms and the right human expertise. Lapins says it best:
“I'm thrilled about the potential for platforms like Lokalise AI to enrich businesses in their localization endeavors and eagerly anticipate actively participating in the ongoing evolution, where we harmonize AI capabilities with human expertise within the industry.”