AI adoption is accelerating, investment is growing, and organizations are deploying increasingly sophisticated language models. But beneath the impressive statistics lies a critical question: Are AI systems genuinely delivering the same quality of experience across different languages and cultures?

In this solo episode of VistaTalks, Host Simon Hodgkins explores one of artificial intelligence's most overlooked challenges: multilingual data quality. Drawing on industry research, emerging European benchmarks, AI governance frameworks, and real-world examples, Simon examines why language coverage alone no longer adequately measures AI performance.

From English-dominated training datasets to the consequences of inaccurate translations in healthcare, legal compliance, and customer experiences, this episode argues for a more human-centered approach to multilingual AI development.

AI Adoption Is Booming, but Is Data Quality Keeping Pace?

The AI industry continues to grow rapidly. Organizational AI adoption has reportedly reached approximately 88%, while Slator estimates that the global data-for-AI market is worth around $9.3 billion, with projections suggesting it could reach $21.5 billion by 2031.

These figures illustrate the scale of the opportunity. However, Simon argues that they reveal only part of the story. Despite the growing demand for high-quality training and evaluation data, participation in the data-for-AI market remains uneven. According to a Slator reader poll conducted in early 2026, 46.7% of respondents reported that their companies were not participating in the data-for-AI space, while only 19% described their organizations as very active.

The complexity extends to AI-powered translation. Polling data suggests 28% of companies use two to four AI translation models in production, while 16% use ten or more. Another 12% don't know how many models their organizations use. For Simon, these findings raise an important concern: as organizations adopt more AI technologies, are they investing enough in understanding the quality and consistency of the data powering them?

The Multilingual AI Problem

One of the most significant weaknesses in modern AI development is the disproportionate representation of English-language data.

Large language models frequently perform exceptionally well in English because their training datasets contain substantial volumes of English-language content. However, that performance does not automatically translate into equally reliable results in French, Hungarian, or other languages.

In July 2026, the European Commission launched the EU Massive Multitask Language Understanding benchmark, known as EU MMLU, to examine how effectively large language models perform across the European Union's linguistic diversity. The findings highlighted performance discrepancies that conventional AI benchmarks can overlook. Traditional multilingual evaluations often focus on whether a model can translate or answer questions in a particular language. But producing an answer is not the same as producing an accurate, culturally appropriate, and useful response.

Simon argues that measuring multilingual AI success by the number of languages supported risks creating a misleading picture of real-world performance. The challenge is not simply whether AI can communicate in another language. It is whether AI can communicate effectively, accurately, and appropriately within that language's cultural and practical context.

Why Language Quality Goes Beyond Translation

Language is far more complex than words and grammatical structures. Meaning depends on context, tone, cultural expectations, social conventions, and the situation in which communication takes place. A seemingly minor wording change can fundamentally alter a message. For example, changing a verb might turn a recommendation into an instruction. An incorrectly placed adjective could weaken a safety warning's clarity. A direct translation of a term or icon might lose its intended meaning when introduced into another culture.

These distinctions matter because AI models learn from the linguistic patterns represented in their training data. That data may include written text, audio recordings, annotations, translations, and human evaluations. However, real-world communication rarely follows a single standardized structure. People speak differently at home, in professional environments, and in public institutions. Regional slang, borrowed expressions, and cultural references shape how people understand information. Government services, legal proceedings, education, and healthcare introduce additional layers of linguistic complexity, often requiring precise terminology and contextual understanding.

When an AI model's understanding of language doesn't match how people communicate, its outputs can become unreliable. And the consequences extend well beyond awkward translations.

The Real-World Cost of Poor Multilingual AI

When multilingual AI systems fail, the people relying on them often bear the consequences.

A customer might struggle to resolve a support issue because an AI assistant misunderstands their request.

A consumer could abandon a purchase because product information is confusing or poorly localized.

An individual accessing public services might encounter unnecessary barriers because automated instructions fail to reflect their language or cultural context.

For businesses, these failures can increase support costs, reduce customer satisfaction, and cut revenue.

In more sensitive environments, the consequences can be considerably more serious.

Legal compliance information, medical communications, and safety instructions demand clarity and accuracy. An AI-generated response that works reliably in English but inconsistently in another language could create significant risk.

Simon highlights a fundamental problem: organizations may believe they have successfully deployed multilingual AI because their systems technically support multiple languages, even when the quality of those experiences varies substantially.

The issue is not just technological performance. It is also about accessibility, inclusion, and trust.

AI Governance Has a Linguistic Blind Spot

As AI adoption expands, governments and international organizations are developing frameworks intended to improve transparency, accountability, and safety.

The 2026 International AI Safety Report, developed with contributions from more than 100 experts and supported by over 30 countries, examines risks including malicious use, security vulnerabilities, unreliable information, and threats to human autonomy.

However, Simon observes that disparities in linguistic performance are not explicitly identified as a distinct risk category.

Similar questions arise when examining major AI governance initiatives.

The European Union's AI Act establishes risk-based requirements for AI systems. South Korea's AI Basic Act addresses transparency and safety, while the OECD AI Principles promote inclusive growth and opportunities for underrepresented populations.

Although these frameworks establish important principles, Simon argues that they do not adequately address the specific challenge of ensuring consistent AI performance across languages in high-risk applications.

An AI system might satisfy existing evaluation requirements while delivering substantially different experiences depending on the language used.

This raises a question for policymakers, developers, and organizations deploying AI:

Should linguistic reliability become an explicit part of AI governance and risk assessment?

For Simon, the answer lies in recognizing multilingual quality as an essential part of responsible AI development, not an optional enhancement.

Why Human-Centered Multilingual Data Matters

Improving multilingual AI performance requires more than simply adding more translated content to training datasets.

The quality, relevance, and cultural accuracy of that data are equally important.

Simon points to the development of the EU MMLU benchmark as an example of how human expertise can improve AI evaluation.

Rather than relying exclusively on automated translations of existing assessment questions, the initiative involved nearly 240 human translators from 21 universities. Their work focused on adapting more than 1,000 questions to preserve meaning, cultural relevance, difficulty, and assessment value across different languages. This approach recognizes that an effective multilingual benchmark must measure more than literal translation accuracy. It must establish whether the underlying knowledge and reasoning tasks remain comparable across languages.

Human linguistic expertise plays a critical role in identifying the subtle distinctions that automated processes can overlook. By combining cultural intelligence, localization expertise, and rigorous data evaluation, organizations can develop AI systems that better reflect how people actually communicate. The objective is not simply to make AI available in more languages. The goal is to ensure people can use AI confidently and effectively in their preferred language.

Building Better Multilingual AI Foundations

During the episode, Simon also discusses how Vistatec approaches these challenges through VistatecData.

The service brings together multilingual expertise, cultural intelligence, localization-quality processes, and broader AI data capabilities to help organizations build, evaluate, and improve AI systems.

The underlying philosophy is straightforward: reliable multilingual AI depends on data that accurately represents how people communicate, behave, and interact across languages and markets.

This includes understanding regional variations, industry-specific terminology, cultural expectations, and the contextual factors that shape meaning.

Rather than treating multilingual support as a feature added after AI development, the approach emphasizes incorporating linguistic quality from the beginning. This matters as organizations move beyond experimentation and introduce AI into customer-facing services, business operations, and high-stakes environments.

By improving the quality of multilingual training and evaluation data, organizations can develop systems that are more capable, consistent, and trustworthy.

The Future of Multilingual AI Is About Quality, Not Just Coverage

The next stage of multilingual AI development will require a shift in how we measure success.

Supporting dozens of languages may demonstrate technological reach, but it doesn't necessarily show equitable performance. Organizations must increasingly ask whether their systems provide comparable accuracy, relevance, and usability across linguistic communities. That means investing in culturally representative datasets, human-led evaluation, language-specific testing, and continuous quality improvement.

It also means recognizing that linguistic performance should be considered alongside other measures of AI safety, reliability, and effectiveness.

The central message from this VistaTalks episode is clear: AI's multilingual future will not be defined simply by how many languages systems can process, but by how reliably they serve the people who speak them.

As artificial intelligence becomes embedded in everyday life, treating language and cultural context as foundational elements of data quality will be essential.

Because truly effective AI must do more than understand words. It must understand people.