If your model speaks English best, you have a buying problem, not a modeling problem

The SAHARA benchmark tested 517 African languages across 16 NLP tasks. The performance gap it confirmed is not a function of linguistic complexity. It is a function of who collects the data, where, and for whom.

The SAHARA benchmark, published at ACL 2025, evaluated large language models across 517 African languages and 16 NLP tasks. It is the most extensive benchmark for African NLP ever assembled. Its core finding is blunt: a pronounced performance gap persists between English and the vast majority of African languages, including widely spoken ones like Hausa, Wolof, Oromo, and Kinyarwanda. The researchers attribute these disparities not to linguistic complexity but to policy-driven data inequities, decades of underinvestment in digital infrastructure for non-English languages.

That framing matters. If the gap were about linguistic difficulty (tonal systems, agglutinative morphology, low orthographic standardization), the fix would be better algorithms. But the SAHARA authors are saying something different: the gap is about who invested in data collection and for which languages. That makes it a procurement problem. And procurement problems land on the desks of people running healthcare systems and banking operations.

The gap is not closing. It is being relabeled.

A TechPolicy.Press analysis published in April 2026 frames the situation precisely: multilingual AI coverage is expanding, but the expansion is performative. Models add languages to their supported list. Benchmark scores tick upward on narrow tasks. But the communities who speak those languages have little say in how data is gathered or used downstream. The gap between adding a language to a model and including the community that speaks it in governance is a political choice.

This is not abstract. A 2025 Stanford HAI white paper confirmed that most major LLMs underperform for non-English and especially low-resource languages, are not attuned to relevant cultural contexts, and are not accessible across much of the Global South. The causes are structural: scarcity of labeled data and poor-quality data that fails to represent the languages or their sociocultural contexts.

More concerning: the problem extends to safety. A 2026 benchmark testing harmful prompts across English and West African languages found that safeguards holding in English degraded sharply in other languages, with refusal rates dropping by more than half. Alignment mechanisms do not reliably transfer across languages, turning multilingual disparity from a quality issue into a systemic risk.

What this looks like inside a hospital

In January 2026, the Joint Commission reclassified language access from a quality initiative to a patient safety requirement under its National Patient Safety Goals. Language access now runs through nearly every safety domain a hospital is measured on: patient identification, medication safety, emergency management, workforce competency. Goal 7 mandates that patients receive information in a language and format they understand.

The clinical stakes are documented. A study published in the International Journal for Quality in Health Care found that 49.1% of adverse events involving patients with limited English proficiency resulted in physical harm, compared with 29.5% for English-speaking patients. Among those harmed, 46.8% of LEP patient events ranged from moderate temporary harm to death, versus 24.4% for English speakers. The adverse events for LEP patients were also more likely to result from communication errors, 52.4% versus 35.9%.

Now overlay that clinical reality with AI-driven triage tools, patient portals, symptom checkers, and discharge instructions generated by large language models. If the model's safety alignment degrades in Spanish, Haitian Creole, or Tagalog, the harm pathway is not theoretical. It is the same pathway that already produces disproportionate adverse events, except now it scales with software.

Banking is facing the same structural gap

Financial services face a parallel problem. The CLEF-2026 FinMMEval Lab introduced the first multilingual and multimodal evaluation framework for financial large language models, specifically because existing financial AI benchmarks remain almost entirely monolingual and text-only. Its existence is an admission: financial AI has been evaluated in English and deployed as if that evaluation transferred.

The CFPB has been explicit. Approximately 26 million people in the United States have limited English proficiency. A January 2025 CFPB review found that while financial institutions increasingly recognize LEP consumers as an underserved market, significant barriers remain, and some entities have used in-language marketing to target LEP populations while burying key costs in English-only documents.

The data problem that sits underneath all of it

The pattern across healthcare and banking is the same: AI systems are being deployed in multilingual contexts while being trained and evaluated in English. The models themselves are not the bottleneck. The training data is. And training data for non-English languages at the quality level required for regulated industries (clinically accurate, financially precise, legally defensible) does not come from web scraping or synthetic generation.

It comes from capture infrastructure operating in the countries where those languages are spoken, staffed by native speakers who understand the domain context. Prior authorization workflows in Tagalog. Claims adjudication terminology in Arabic. Sepsis screening questionnaires in Haitian Creole. Each of these requires annotators who are both native speakers and domain practitioners, not one or the other.

That capability does not exist inside most healthcare systems or banks. It cannot be assembled ad hoc when a regulator asks for multilingual performance documentation. It requires standing operations across 119 or more countries, native-speaker networks spanning 150 or more languages, and quality infrastructure ensuring the data meets evidentiary standards of regulated industries.

The multilingual AI gap is real, widening in relative terms even as absolute coverage expands, and it will increasingly determine which organizations pass regulatory scrutiny in healthcare accreditation and financial examinations. The question is not whether your model supports Yoruba or Vietnamese or Arabic. It is whether the data that trained it in those languages was produced by people who understand what a correct answer looks like in a clinical, financial, or legal context. That expertise is rarely inside the organization buying the model.