In a memorable presentation during SlatorCon Silicon Valley on September 5, 2024, Silvio Picinini of eBay challenged the translation industry to rethink its processes in light of the transformative capabilities of language AI.

The Language and Data Quality Manager asked attendees, “What if we only had language AI? What if we built this from scratch today, at scale?” His presentation examined the implications of language AI for translation management systems (TMSs), context, glossaries, and evaluation, ultimately questioning the role of humans in this evolving environment.

Picinini began by defining language AI and emphasizing the key differences between it and previous machine translation (MT) models: language AI operates in the realm of numbers, while traditional MT processes are based on words. This fundamental shift enables language AI to capture meaning and context in ways that were previously impossible.

One of the most striking points Picinini made was that language AI’s access to vast amounts of data makes it potentially superior to traditional TMSs as well. “Language AI knows all the translations of all the translators, all the reviewers, and all the SMEs,” he said.

Collective knowledge is what allows language AI to generate suggestions that may be more informed and nuanced than those of a single human translator or reviewer.

Picinini also highlighted the limitations of TMSs in providing context for translators. He noted that these systems often break the context by presenting sentences out of order, making it difficult for translators to understand the broader meaning. 

Language AI, on the other hand, can translate with full context, even incorporating images and other multimodal data.

“Language AI knows all the translations of all translators, all the reviewers, and all the SMEs, all the lawyers … all the engineers .. all of that is there, and from all companies.” — Silvio Picinini, eBay Language and Data Quality Manager

Language AI Is Not Deterministic

Glossaries, traditionally created for humans, also face challenges in the era of language AI, Picinini pointed out. MT struggles to follow glossaries consistently, often leading to errors. 

Language AI, however, excels at term extraction and can assist humans in creating more effective glossaries, added the executive.

On the subject of [translation] evaluation, Picinini acknowledged that language AI is not deterministic, meaning that it can produce different outputs based on the same input. However, he also argued that humans are themselves non-deterministic, and that AI evaluation, with multiple agents and one “judge,” could be more reliable than traditional human evaluation.

Picinini indeed envisions a future where language AI plays a central role in translation, with humans focusing on managing glossaries and providing feedback to improve models. He also predicts that the localization industry’s expertise in evaluation will be increasingly valuable as companies seek to assess the quality of generative AI output.

While Picinini acknowledged that language AI is still in its early stages, he expressed optimism about its potential to transform the language industry at large.

The audience responded enthusiastically to Picinini’s presentation, with many questions about the practical implementation of his ideas. 

In response to a question about the cost of AI evaluation, Picinini acknowledged that it can be expensive. He also noted that at the same time, the cost of LLMs is rapidly decreasing, and suggested that AI evaluation could ultimately reduce costs thanks to a lesser need for human post-editing.

When asked about experimentation with language AI becoming production-ready, Picinini reiterated that it is still in the early stages, but his prediction is that language AI will become faster and more capable over time.

Silvio Picinini’s presentation offered an interesting vision for the future of translation, ultimately arguing, as have many other experts, that language AI has the potential to make translation more efficient, accurate, and accessible.