Innovating at the Speed of AI

Artificial intelligence (AI) and large language models (LLMs) are reshaping the speed of innovation.

The opportunities these technologies hold lie in their unparalleled ability to accelerate innovation. Welocalize recently had conversations with leaders from over 40 of the largest global companies across various industries, including life sciences, software, and travel, to gauge how they view the changing landscape of business brought about by LLMs.

The Changing Landscape

There is more to generative AI (GenAI) and LLMs than Open AI’s ChatGPT and Google’s Bard. Some of the largest companies in the world, including Meta, IBM, Microsoft, Baidu, and Huawei, are developing and refining hundreds of LLMs.

Broad Adoption

“We expect wide-scale change, with rapid adoption of generative content within 12 to 18 months,” noted Darin Goble, Head of Solutions at Welocalize. However, while most companies are still in the early stages, they still need to be ready from a production predictability and tech stack perspective.

And yet, content managers, localization leaders, technology innovators, and many others are under pressure to deliver LLM implementation. An executive from the software industry shared this tension, with the push from their leadership to go live with LLM output versus the push from legal to wait.

Disruption to NMT

The leading language service providers (LSPs) have successfully used neural machine translation (NMT) in their localization projects for years. Combined with human post-editing, this has become a proven operating model. However, now, LLMs are about to disrupt, and possibly dethrone, NMT.

During one discussion, Goble clarified, “NMT still currently outperforms LLMs for localization, but we expect this to change due to the pace of innovation.”

Welocalize has run localization experiments that showed mixed performance from LLMs. However, it’s only a matter of time before this shifts. A localization manager from a multinational internet brand and services provider shared that they produced positive results for a proof of concept with AI as a machine translation (MT) tool.

Co-Pilot Authoring

Many software companies, from Microsoft and Google to Salesforce and GitHub, have already integrated GenAI co-pilots into their apps and content creation platforms, allowing companies to generate content and translate it into multiple target languages.

“This is the paradigmatic shift of this new wave of technology, and we see it as an eventual inevitability for some language pairs and content types,” Goble shared. Companies should expect larger volumes of content as LLMs are making content generation and translation easier and faster.

Revised Workflows

Professionals working with global content, LSPs, and content providers must rethink their workflows to validate content generated and presented in the target language. From source to target, where NMT and human translators localize content for subsequent review, Welocalize is testing a new workflow that becomes “prompt to target.”

The customer generates multilingual content using prompts, which then uses an AI quality estimate to determine the fit for purpose. If it does not pass, the LLM is tweaked, and the prompts are modified.

However, there were several questions and concerns, including:

• The cost difference between MT with post-editing and generative AI with post-editing
• The estimated time saved when applying the new workflow
• Combining and integrating everything from a tech stack perspective
• How generated content can pass quality evaluation without human evaluation
• If humans are no longer needed for translations
• How to customize prompts per target language
• Who will check the brand’s tone and voice for translated content
• When this new workflow will move from pilot to production

“It’s not only workflows, but job functions will need to evolve rapidly,” Goble pointed out. “Localization leaders will need to move upstream. And localization management will evolve into global content management.”

Risks and Mitigations

Despite LLMs’ vast opportunities, risks are involved, from copyright issues and hallucinations to data protection and regulation. The biggest concerns are the accuracy of generated and translated content and data security.

Trusting AI-generated content remains a significant issue. Generative AI has been known to produce unpredictable and problematic output, including biases, falsehoods, and hallucinations, which all stem from problems with underlying training data. A translator from a multinational pharmaceutical and biotechnology company asked about placing safeguards to prevent the AI from inventing false statements. A quality management leader from the same company expressed concerns about using LLMs outside of their intended purpose. There was also a question from the senior localization manager of a global multi-billion-dollar fashion enterprise about trusting AI’s quality estimation.

A primary concern is security, which includes data privacy and protection. Goble addressed this by pointing out, “Security concerns will be overcome by major tech players, with regulations likely to differ by region. We are working closely with Microsoft and other suppliers. We are keen to work with our clients to map out new workflows pertaining to GDPR (General Data Protection Regulation) and other infosec considerations in their industries.”

Here are some practical ways to ensure enterprise security and data privacy:

• Use AI/LLM services hosted by major cloud providers that incorporate enterprise security and data protection measures for their AI offerings.
• Ensure these providers use data encryption and isolate the foundational LLMs from client data.
• Educate employees on the proper and approved use of AI platforms and the risks of non-approved platforms.
• Form an internal AI committee to actively evaluate new technical, legal, and business impacts of LLMs.

Learn More

Welocalize is at the forefront of innovation, with over two decades of experience in AI and machine learning (ML). We have a dedicated team of ML engineers, data scientists, and computational linguists.

Our best-of-breed technology-agnostic strategy allows us to position our partners for success in this new frontier of LLMs. Our mission is to harness the transformative power of generative AI and pioneer its integration into localization workflows, content supply chains, and the broader evolving business landscape.

Hear more about innovating at the speed of AI during Welocalize’s session at SlatorCon.