A frequently discussed weakness in AI translation is the tendency for models to perform poorly with nuanced language like idioms and cultural references. Though progress has been made, AI translation still often produces overly literal outputs that require significant post-editing, especially when working in low-resource languages.
For example, the CBC reported in September 2025 on an initiative working to develop guidelines regarding English to Punjabi medical translations after the developers spent two years recording examples that appeared to be “word-for-word translations using AI, missing cultural and gender nuances.”
Similarly, when Slator performed an informal translation test of GPT-5 at the end of 2025, we observed generally accurate translations across five European languages, but with “recurrent issues in naturalness and idiomaticity, reducing their suitability for direct publication without human editing.”
These examples seem to be consistent with a February 2026 study from the data-for-AI company Appen, which tested translation performance from seven major AI models across 20 languages. The researchers found that while general cultural content was handled “relatively well,” the models performed poorly with idiomatic language and puns, frequently leaving idioms totally untranslated.
While rich and nuanced AI translations are the eventual goal, a more immediate concern may be the persistence of AI hallucination, where models incorporate inaccurate or fabricated information into their outputs.
In an October 2025 study, researchers from Alibaba found that AI hallucination is still a significant problem across major AI platforms. Evaluating 17 major LLMs in 11 “English-to-x” language pairs, the researchers reported an alarming translation hallucination rate of 33% to nearly 60% depending on the specific model and language pair.
Notably, the researchers also found that different models do not necessarily fail in the same way. For example, while some models were more likely to add fabricated information, others were more likely to translate into the wrong language, a phenomenon sometimes called “language confusion.”
Recent research has also identified persistent issues such as terminology inaccuracies in specialized domains and stylistic inconsistencies in translation outputs.
Where Do These Issues Come From?
Many of the limitations described above stem from structural characteristics of how AI translation systems are trained, evaluated, and deployed.
AI models — and the benchmarks used to evaluate them — tend to be heavily based on data from high-resource languages, and English in particular, leading to less robust capabilities in low-resource languages.
Also, prompt engineering cannot reliably compensate for gaps in a model’s underlying capabilities. An August 2025 study suggests that if a model has not effectively learned a language pair or stylistic nuance during training, prompting strategies may offer only limited improvements in translation quality.
Another area holding the industry back may be the inadequate application of automatic metrics when it comes to evaluating AI translations. Notably, it’s been observed that many metrics can be easily “hill-climbed,” meaning that models can be trained specifically to perform well on a given automatic metric.
Essentially, this means a model can be developed so that it performs well on paper without meaningfully improving its performance in real-world translation scenarios.
2025 Slator Pro Guide: Translation AI
The 2025 Slator Pro Guide Translation AI presents 15 impactful ways that AI can be used to enhance translation workflows.
Working Towards Solutions
In many cases, automation adds value by producing output faster than a person could and reducing certain types of “human error.” However, in the case of AI translation, the human touch is still necessary in many cases to address the issues we’ve discussed here.
This starts at the very beginning of the development pipeline when AI models are trained on large corpora of data.
While the use of synthetic data has gained much attention, some enterprises have put significant resources into prioritizing AI data quality through human intervention. Examples include major Language Solutions Integrators (LSIs) like RWS and Acolad, who have developed specialized data teams focused on expert-in-the-loop (EITL) data annotation and validation to augment their AI offerings.
Beyond enterprise initiatives, several research projects are working to improve the availability of high-quality multilingual training data. For example, Hugging Face recently released a large-scale translation dataset designed to improve model performance across multiple languages.
In other cases, developers have been addressing the English-dominance in training data by producing curated datasets using low-resource languages. For example, in 2026, Microsoft announced 11 winners to their “LINGUA” open call, which was aimed at offering funding and technical support to projects producing datasets for European low-resource languages.
Similarly, France’s INRIA research institute has launched projects aimed at improving AI translation quality for underrepresented languages by developing curated datasets and evaluation benchmarks.
At the other end of the pipeline, machine translation post-editing (MTPE) remains a common strategy where human linguists edit AI outputs to improve the accuracy and fluency of the final translation. In fact, 84% of respondents in the Slator 2025 LSI survey reported that clients had specifically asked for human editing services to improve AI translation outputs in the previous year.
Alongside these human-centered efforts, researchers are also exploring more advanced AI architectures to improve translation reliability. Experimental multi-agent AI frameworks assign specialized agents to tasks such as translation, cultural adaptation, domain adaptation, and quality evaluation, allowing multiple AI systems to collaborate and refine outputs throughout the translation process.
