Many companies are rolling out artificial intelligence (AI)-first or AI-augmented services that directly or indirectly interact with users. These services rely on large language models (LLMs) that are trained on data that is overwhelmingly in English and a handful of other languages — and that therefore perform less than optimally in most languages. This presents some potentially thorny issues when it comes to localizing AI-based services.  

Secondary and low-resource languages pose especially difficult challenges due to their sparse training data. To make matters worse, a lot of the online content in these languages is machine-translated, which creates a “garbage in, garbage out” problem if that material was used to train the model. So while your product might perform like ChatGPT-4 in English, it could be a lot less reliable in Hindi.

In light of these shortcomings, the following five best practices can help companies improve their localized AI-based services and provide the best experience possible for their customers.

1. Leverage Native Speakers to Test User Experience

One of the reasons it is a good idea to hire bilingual staff when possible is they can evaluate the product and look for functional issues while running it in their language. Native speakers can spot when AI is generating language that sounds right but is actually incorrect — something  that may go unnoticed by other testers (and you really don’t want to turn a broken experience loose on your users).

This is also something a transcreation agency can help with. These agencies hire copywriters who are native speakers of the target language. They can assess the quality of translations, style, and accuracy of responses in their respective languages.

2. Release New Languages Cautiously

It is generally not a good idea to “shotgun release” new languages without doing thorough testing first. Even then, you should label the release as “beta” and also compare user metrics versus English. Users are generally pretty tolerant of minor translation errors in the user interface (UI), so those won’t hurt usage much; but if they are using your service to do research and getting nonsensical results in their language, they may leave and not come back.

3. Use English as a Bridge Language

Companies can use a translation AI like DeepL or Google Translate LLM to translate to and from English. In this approach, the prompt is translated from a secondary language to English, and the response is translated back to the input language. Some AI providers may already be doing this behind the scenes, so it is worth investigating that before you add your own translation layer, which could just get in the way. 

This approach comes with some risks, as there will typically be a loss of information in each direction. The best models deliver accurate translations 80-90% of the time, which can make the difference between a good prompt and one that will produce garbled results. Specialist translation AIs generally outperform generalist AIs like ChatGPT for translation. So if you choose to do this, it’s probably a good idea to make it clear to users, so they can use that instead if they speak a better-resourced language.

Secondary and low-resource languages are often underserved or not served at all by translation engines. Sometimes you can find specialist translation platforms that target specific languages. What you’ll typically see is that translation accuracy is not as good for secondary languages, which will cause information loss in both directions. The good news is many people understand one or more of the other top international languages where AI platforms perform well. For example, French is widely spoken in parts of Africa, so users there might find the platforms work best in French versus local languages and dialects.

Overall, this approach may be the least-bad option for low-resource languages, but it is not ideal for most.

4. Collate Results of Multiple Queries

Information is often siloed by language. Consider a situation in which a user asks an AI about Japanese baseball players. While information about top players may be available in English, there is likely much more complete information in Japanese. In this situation, the best approach may be to do multiple queries in different languages and then collate the results to be presented to the user. 

Savvy users may already be aware of this issue and know how to prompt systems in a way that surfaces siloed information. But the average user will probably not be aware of this, so collating results is something AI providers can do behind the scenes to improve responses. 

5. If Possible, Utilize a Native Multilingual LLM

One encouraging development is that countries are developing their own AI models that are trained in relevant languages to build natively multilingual services. For example, in Switzerland, leading universities are developing an open-source, multilingual LLM that others can build on. I expect to see more developments like this to create AI models to serve other regions. However, while this is the best long-term solution, it is also an expensive one due to the cost required to train LLMs.

Conclusion

It doesn’t matter how good your UI and website look in Thai if the underlying AI service is dysfunctional in that language. For now, the five best practices presented above are the best strategies for companies that want to create better AI services for their users around the world.

However, there is no easy shortcut to localize language models without training them in other languages. Native multilingual LLMs may be the only way for AI providers to attain native proficiency in other languages. That’s why I hope that open-source models will win out in the long run, because they will enable countries that might otherwise be underserved to build their own models.