Human checkpoints in AI-supported multilingual content: a conversation starter

Everywhere you go, everything you read these days, there’s a mention of artificial intelligence, commonly called AI. There is an untapped potential for AI in the digital strategy space, especially for multilingual digital strategy.

We’ve written before about the importance of starting with user research and having a clear strategy before designing experiences for communities with limited English proficiency. We’re well aware of the challenges teams face when trying to serve multilingual communities with limited funding and staff. AI can support this work, and we cannot ignore its potential.

But if teams want to see a positive impact from using AI, human intervention is essential. For that reason, from the start of a project or initiative, there is an opportunity to define when and how humans are integrated into AI-powered translation and transcreation workflows.

Over the past few months, we have engaged in informal discussions about the role of humans when using AI to translate content in other languages. This post is our attempt to begin identifying and documenting a few key moments when we, humans, remain critical. 

Our goal is to spark a conversation, learn from each other, and share insights on these ideas. We also hope you will contribute by suggesting additional moments when human input is needed throughout this process.

1. Determining content risk level

Not all content that we develop and maintain carries the same level of risk. Organizations can benefit from prioritizing review and editorial processes based on content risk level. Defining these categories upfront can help you scale and manage translation and transcreation efforts more effectively.

Here are a few examples. 

Low-risk content - Internal documents (such as discovery or exploratory drafts or internal summary materials) might be considered low risk since they will not be shared publicly. For these internal documents, AI translation alone may be sufficient.

Medium-risk content - All translated content should be reviewed by a human before being shared outside an organization. That said, evergreen content (such as program descriptions or seasonal information that’s updated every year) rarely changes. AI translation combined with light human review may be sufficient, as an in-depth review was performed when the content was first produced. It might also require lower levels of review if only minor changes are made over time. 

High-risk content - Legal rights, healthcare guidance, immigration case status, benefit eligibility, emergency alerts, and compliance requirements are all high-risk topics that require human review. The content within these areas is highly sensitive and needs frequent updates due to policy updates and other changes. This type of content requires review by multilingual content strategists and, many times, also by subject matter experts before publication.

2. Identifying use cases 

Different content types require different approaches. While AI can process text quickly, it might lack the nuance to distinguish between a formal legal notice, a press release, and an outreach flyer. Humans are needed not only to categorize risk levels for different content, but also to determine the appropriate tone, voice, and overall translation treatment for each type and use case. This ensures that specific guidelines, including glossaries and style guides, are applied correctly, providing AI with the specific context it needs to produce a message that resonates with audiences.

3. Creating multilingual style guides and glossaries

AI is a tool, not a solution. Without systems in place, AI will not perform as organizations need it to.

Developing multilingual style guides and glossaries helps ensure that your organization’s content strategy relies on consistent terminology, embraces plain-language principles, promotes findability, and remains culturally relevant. This is an area that requires human intervention. Humans are needed to build these resources and maintain them over time as living documents.

4. Developing prompts with clear translation and transcreation instructions 

While style guides and glossaries help with consistency, AI still needs clear guidance on how to handle a translation. A simple prompt like “translate this into Spanish” might produce a technically accurate result, but it often lacks the strategic intent behind the original message. To get a high-quality output, human intervention is needed to design detailed instructions based on user research and good understanding of the needs of the audiences you’re trying to reach. 

Detailed instructions include things such as specifying the desired reading level, identifying terms that should remain in English, and guidance on how to handle cultural nuances of the language. By providing this level of detail, we move from a word-for-word translation to a more thoughtful, user-centered translation or adaptation of the key messages.

5. Curating "golden case" examples 

AI tools benefit significantly from seeing what "good" looks like. "Golden cases" are high-quality, human-approved translations that serve as a gold standard for future work. By supplying AI with these curated examples, we reinforce the organization’s specific voice, tone, and overall style. Human intervention is key here as someone must identify these successful examples and organize them into a library that teaches AI how to reflect the unique personality and values of the organization.

6. Reviewing AI-produced content 

The revision stage is where most people would say human intervention is needed when it comes to AI translations, and for good reason. It acts as a guardrail, helping ensure that translations are accurate and ready to share with the public. This step goes beyond simple proofreading. Content strategists must ensure that translated AI outputs align with the organization’s overall communication and content strategy goals, address the needs of the audiences, and reflect the appropriate tone, voice, and messaging intent. Humans can also make editorial decisions when linguistic and cultural tradeoffs exist.

7. Testing with real users

We need to test translated content with real users to help identify confusing phrasing, unexpected interpretations, literacy barriers, and cultural disconnects. This testing is another key workflow where human intervention is required. 

8. Measuring comprehension and not just translation accuracy

Accuracy does not always equal understanding. Many organizations evaluate translation quality based solely on linguistic correctness. We must also ask questions about multilingual content that capture qualitative and quantitative data. What is learned from these evaluations can help the team fine-tune artifacts, including the AI tool. Sample questions include: 

• After reading the information in their language, do users understand what they need to do next? 

• Can users complete their intended tasks?

• How long do users spend on the page?

AI has opened new opportunities for organizations to serve broader audiences, including community members who rely on multilingual content to find the information and services they need. To be most effective, we must design AI translation workflows that include clear human checkpoints, defined risk thresholds, and continuous user feedback. 

What other human intervention points should we add to the list?

This blog post was a collaboration between Leilani Martínez, founder of KulturaTech, and Karla Fernández Sánchez. Here’s more about Karla:

Karla Fernández Sánchez is a bilingual content strategist and learning experience designer focused on making digital experiences clear, accessible, and inclusive. Her work centers on multilingual UX and plain language.

Most recently, she led bilingual content and translation efforts for public benefits programs, where she developed multilingual glossaries, validated translations with native speakers, and designed AI-supported workflows that balance efficiency with quality and cultural relevance. She is especially interested in the role of human insight in shaping accurate and trustworthy AI translations.

She also brings a background in learning experience design, where she has created learning journeys and training courses across edtech and video telematics. This experience informs her approach to content development and user experience, with a focus on creating solutions that meet people where they are.