Blogs

5 Critical AI Language Translation Gaps to Watch in 2025
Since ChatGPT dropped in 2022, large language models (LLMs) have been all the rage. ChatGPT and similar tools like Claude and Gemini have been heralded as high-powered productivity tools that can streamline your workflow and make many of our day-to-day work tasks significantly easier. And while it may be true that they can make our jobs a little bit easier, that doesn’t mean they can make our jobs easy. In the field of translation, LLMs have been widely applied alongside machine translation (MT) tools to produce fluent and accurate translations from one language into the next. But organizations looking to translate their content should be careful when employing tools like ChatGPT to translate texts — though they may yield accurate literal translations, these translations may not be fully adequate for your needs as a business. In this blog post, we’ll take a look at five key areas in which AI tools like ChatGPT are still lacking when it comes to translations. From their struggle to understand context to their tendency to editorialize, these are some of the most common issues you’ll want to look out for if you plan on using any of these LLMs to translate your content. Although LLMs can produce pretty accurate literal translations, they’re still not great for situations where you need to translate culturally nuanced language like idioms and other common expressions. According to a 2024 study, AI still struggles to make sense of (and thus, translate accurately) culturally nuanced phrases and ambiguities that human linguists are better prepared to parse out. So while these tools may excel with simple texts like routine forms and paperwork, you’ll still need a human in the loop when translating more complex texts like marketing brochures and web pages. If you’ve ever tried to correct ChatGPT after it produces inaccurate output, you know just how hard it is to get LLMs to correct themselves. But making corrections and editing our work is a key part of a translator’s work — language service professionals catch errors in the moment and adjust accordingly all the time. In a 2023 study, researchers found that LLMs have trouble “self-correcting” — that is, refining future output based, either on their “inherent capabilities” or on feedback to previous responses. According to the study, LLMs struggle to “self-correct their responses without external feedback, and at times, their performance even degrades after self-correction.” That means that these tools may produce errors and then repeat those errors throughout a given text. If you’re using these tools for translation, you’ll need to make sure somebody knowledgeable about both the target and source language is carefully reviewing the output and editing it to minimize these errors. LLMs have raised a wide range of data privacy concerns, and organizations working in highly regulated industries like healthcare and law should be leery of using them for translation tasks. Different industries and locations have different standards of data privacy, so it’s important to be aware of what is and isn’t acceptable for your circumstances. For example, LLMs are generally not HIPAA-compliant — healthcare organizations using this tool to translate texts into another language must make sure to mask any patient health information (PHI) such as name, date of birth, social security number, etc. before putting it into the tool. That means healthcare organizations must be careful to either manually or through automations (AvantShield) scan the original text for any such information and remove it entirely from the text, to avoid issues with HIPAA compliance. In addition to data privacy, another common concern that experts have raised about AI is its implicit biases. All sorts of AI models have drawn criticism for their biases — take, for example, Amazon’s recruiting tool that showed a bias against women applicants in the search and hire process. Such biases may also come up in LLMs performing translation tasks. LLMs are particularly notorious for tone-shifting and editorializing in their translations. They may shift the overall tone of a text to more closely align with its own standards of tone — for example, rephrasing a common yet pejorative buzzword like “woke” as something more neutral, like “aware of social inequality.” The translation of this phrase would in turn lose the connotation understood by the original terminology, harming the overall quality of the translation. Although LLMs can produce highly accurate and fluent text in languages like English, Spanish and French, the same isn’t true for all languages. These tools are trained on massive amounts of text in various languages, primarily taken from the internet — there’s more available input for languages like English and Spanish than there is for Pashto, for example. Languages with more training data will be easier to translate between; on the other hand, LLMs will struggle to produce accurate translations in languages that don’t have a large digital presence — like indigenous languages of the Americas, for example. LLMs may be useful tools, but they’re still far from adequate replacements for human linguists. Instead, human linguists should take a smart, balanced approach to incorporating AI tools into their workflow. By correctly identifying the proper use scenarios of AI in the translation process, we can leverage AI tools in an effective way that speeds things up without sacrificing the overall quality. Here’s why:
LLMs struggle to understand cultural or historical context that may be relevant to certain texts, making transcreation and localization tasks nearly impossible with AI alone.
These tools aren’t great at correcting themselves, even when they receive external feedback.
Organizations working in highly regulated industries like healthcare need to be careful not to violate data privacy laws.
AI is known for holding implicit biases which may affect the quality and content of the final translation.
The overall quality of translation will vary by language — languages that do not have a significant digital presence have less training data to draw from and produce an accurate translation.
At Avantpage, we’re well aware of the limits of AI tools in the translation process. We have a technologically driven translation workflow, with human linguists involved for quality control. If you need translation or localization services, contact us today at [email protected] or (530) 750-2040.

AI Interpreting: Is it ready for the healthcare sector?
Artificial intelligence (AI) is transforming many industries, including language services, like interpreting, and the healthcare sector. Working at the intersection of these areas, we see our clients struggling to figure out how best to benefit from exciting developments in automation and AI-powered technologies without compromising interpreting and translation quality. At Avantpage, we’re supporting healthcare organizations by implementing AI carefully, safely, and transparently to optimize business processes and translation workflows. Tools like machine translation can greatly enhance the work of human translators and there are ways that human interpreters can leverage AI to support their job as well. But while Avantpage embraces AI in certain aspects of language access, we encourage our clients to carefully consider the implications of services that offer to entirely replace linguists with AI interpreting technology. These offers of a service that’s faster and cheaper are understandably appealing when you’re on a tight budget. But are there more risks than benefits in a highly regulated industry where patient experience and outcomes are the top priorities? The National Council on Interpreting in Healthcare¹ recently published their “Guidance for Healthcare Organizations Evaluating the Potential Use of AI-generated Interpreting,” where they highlight many areas for organizations to ask deeper questions. We consider some of these here, highlighting five core concerns for health plans and hospitals considering an exclusively AI-interpreting model. A crucial part of language access planning is understanding which languages you need to provide. The core languages spoken in your patient communities should be easy to identify. But how will you offer support to a patient who needs an interpreter in a less common dialect or a language of lesser diffusion? It is important to remember that AI models are only as good as the data used to train them. If that data is limited (eg. for indigenous languages or other less common languages and dialects), you cannot rely on the same level of accuracy as a commonly spoken language such as Spanish. If you are using an AI interpreting model, does your provider have a backup solution for languages that are not available through AI? And on the topic of Spanish, can the AI tool pivot to provide a Spanish speaker from Puerto Rico versus one from Mexico, to match the needs of an individual caller? These are important questions to consider when considering an AI interpreting provider. As mentioned above, it is important to consider whether an AI interpreting service has been trained using sufficient data to provide accurate output for your members or patients. But the quality assurance questions don’t end there. Human interpreters are required to meet certain qualifying criteria. These include proof of subject-specific training, nationally recognized credentials, field experience, continuing education, and adherence to the National Code of Ethics. Years of experience (Avantpage requires a minimum of 3 years) have taught interpreters to self-monitor and to constantly look out for subtle misunderstandings between the parties. If a human interpreter identifies a communication issue, such as a cultural misunderstanding or a moment of unintelligible speech, they can intervene before it becomes a larger issue. How does an exclusively AI-generated interpreter guarantee the same level of nuanced expertise and service quality? Has the AI been trained with terminology that is specific to the healthcare industry? Does the service offer an opportunity to switch to a human interpreter if it becomes apparent that a quality issue is developing on the call? Distinct from quality assurance, quality control is about monitoring performance and improving a service to maximize quality and outcomes for patients. When it comes to written translation, tools with AI features learn and improve through detailed feedback from human linguists. If you are considering an exclusively AI interpreting model, find out how the tool will learn from mistakes. Will there be an opportunity for your staff, patients, or human interpreters to give feedback, and will that be incorporated into the model’s improvement? How does a user submit a complaint, how will errors be tracked, and who is ultimately responsible should a serious error occur? Will the service be spot-checked and monitored from the quality perspective, to be sure that patients and providers are receiving the level of quality they deserve? If you use a service like this, are you complying with federal regulations surrounding quality and AI, such as those outlined in Section 1557 of the ACA? Anyone who works in healthcare in the United States understands the importance of the Health Information Portability and Accountability Act (HIPAA) and, in the European Union, the General Data Protection Regulation (GDPR). When a health organization contracts with a language services provider such as Avantpage, that LSP must show that they offer translation and interpreting services that align with these regulations. This includes HIPAA training for staff and interpreters, secure and encrypted phone lines and servers, and many other data security measures that are monitored through annual audits. AI interpreting models work by recording conversations, converting them to text, translating them, and then voicing them in the target language. If an AI interpreting service is recording patient data, where is it being stored and for how long? What measures are in place to prevent a breach? Are patients aware that they are being recorded, and is there an option to opt out of this and be routed to a human interpreter on a non-recorded line? Addressing the factors above is an important part of building patient trust, which is arguably one of the most crucial factors when it comes to providing quality care and improving health outcomes. You want your members and patients to feel that you care at every turn in their journey with your organization. As the National Center for Interpreting in Health Care stated in their guidance¹, “qualified human interpreters often play a crucial, but understudied role not only in ensuring effective communication but also in building empathy and trust in healthcare settings, aspects of care that AI will not be able to fully replicate.” Where can you carefully implement AI within language services without eliminating the human approach? Can you be sure that your patients are not affected by unchecked racial biases, which can be unwittingly introduced in an AI interpreting tool by virtue of its limited training? If you do decide to offer AI interpreting services, is there an option for your patients to opt out and reach a human interpreter without any delay or other impact? Avantpage is excited about the progress in this field and we have been incorporating AI-powered processes into our workflows for some time. By asking questions such as those suggested above, healthcare organizations can make informed decisions about the extent to which they are willing to replace human steps with AI. To learn more about how we provide meaningful language access by combining AI with human expertise, contact our team today.

AI or Human Translation: A Roadmap for Your Translation Project
In the year since OpenAI launched ChatGPT in November 2022, we humans have been scrambling trying to figure out just what to make of artificial intelligence (AI). No matter the industry you work in — government, medicine, or even language services — the use of AI as a tool has been a hot topic for all of us this year. In the language services industry, we’ve been dealing with AI and translation for even longer now — since the rise of neural machine translation in the 2010s, AI has been an extremely powerful translation tool. But the highly fluent nature of texts produced through modern machine translation tools raises the question: When is human translation better than AI translation, and vice versa? The answer is complicated — it all depends on the type of content you need translated and the specific requirements you have for that translation. That’s why we’ve devised the following roadmap for you to determine whether you should request AI-powered translation, human translation, or a hybrid model for your language services. Here, we’ll take a look at the questions you should ask yourself before determining which approach to use in your translation needs.
The first question you’ll want to ask yourself is about the type of content that you need translated — AI is especially good at translating more technical, repetitive texts, however it still lacks the human touch necessary for more creative types of content, like pun-heavy marketing materials or nuanced blog posts, just to name a couple of examples. While a human will generally review AI translations before you get the final product, you may find that the translation process goes more smoothly when you start with a human translator from the get-go for certain types of content. You’ll also want to consider some of the specific requirements of the project when you’re wondering which route to go. Here are just a few factors to look at: If you have a tight deadline for a given project and the content type is suitable for AI translation, AI could be the right way to go. AI tends to be much faster, where human translators need more time. That said, if the content type isn’t ideal for AI translation, you may find that the review process ends up taking longer than desired — in cases where your content needs to be translated quickly by a human, consider requesting a rush translation from a trusted language service provider. Human-produced translations are going to cost you more than an AI translation. Still, it’s important to consider the fact that human translations are typically higher quality, and mistakes resulting from an AI translation could be costly. Consider the type of content you need translated first before determining what your budget constraints allow for.
Quality expectations vary across projects — choosing between AI and human translation may also hinge on factors like the expected accuracy, cultural sensitivity requirements, and the project’s potential impact on organizational outcomes and performance. Generally speaking, human translators will be able to ensure more accurate results, as AI often makes errors in vocabulary and context. Additionally, critical content like legal or medical documents demands the precision that human translators provide. AI, while advanced, often lacks the contextual understanding required for such materials. Language service experts often note that AI translates words, not meaning, while humans translate meaning, not words. This is particularly important to keep in mind with texts that may have culturally sensitive information — certain words and phrases might come across bluntly or insensitively when directly translated into another language, and as such, a human translator is important when considering cultural sensitivity. Human translators excel in capturing cultural nuances, ensuring that the translated content aligns seamlessly with the target audience. For projects where translation quality directly impacts outcomes — think educational materials that will assist a patient in their health decisions, or a child’s individualized educational plan that will determine their educational path — investing in human expertise is generally the more strategic choice.
While AI can produce content that meets your voice requirements, human translators are generally much better at creating content that adheres to your style and voice guidelines. A common complaint about AI translation tools is that they can’t consistently account for stylistic requirements — for example, if your brand’s content strictly follows the Chicago Manual of Style, AI may not be able to adequately follow those style guidelines. While machine translation glossaries and adaptive machine translation can be useful for AI translations, human translators still have an edge here. Humans are more well-equipped to translate jargon-heavy or highly stylized texts, and it’s much easier to get them to review and follow your organization’s unique style and tone guidelines.
Because AI-powered translation tools tend to struggle with accuracy more than human translators do, it’s important to consider the risks associated with your project — that is, what could go wrong, and how can you prevent the likelihood of that happening? For highly sensitive content — medical documents, legal texts, and anything else that’s heavily regulated — human translators provide the necessary expertise to navigate complex terminologies accurately. Inaccurate translations of these kinds of texts can be costly, have legal implications, and could even be life-threatening — that means it’s important to work with a trusted human translator who can provide the most accurate translation possible. If you determine that the overall risk is low and opt for machine translation, you can mitigate risk even further by incorporating thorough review and quality assurance processes. Quality assurance tools can perform objective measures of the quality of a given translation, while human reviewers are an absolute must, as they can fill in the gaps in any AI translation. Additionally, user feedback can play a useful role in mitigating the risk associated with AI or human translation. Consider testing out a translation with a small focus group to hear their thoughts on a given translation — this will give you a sample of what end users will think when they encounter the final product, allowing you to tweak details as necessary.
Before you decide to opt for an AI translation tool over a human translator, ask yourself the following questions: Whether or not you use AI or human translation depends on a wide range of factors — and sometimes, the answer still isn’t clear. In those instances, you may find that a hybrid approach works best. At Avantpage, we recognize the value of both AI and human capabilities. Our approach integrates AI features, human expertise, and hybrid methods to ensure a customized solution that works best for your project’s unique demands. If you’re looking to learn more about whether AI translation, human translation, or a hybrid approach is best for your project, contact us today at [email protected] or (530) 750-2040. We’ll help you determine and execute the most effective strategy for your project.