When AI Can’t Understand Your Language, Democracy Breaks Down

What if artificial intelligence looks at your language — and sees nothing?

For millions of people, this is not a metaphor. It is the daily reality of interacting with AI systems that simply do not understand the language people use to participate in public life.

Today’s most advanced AI models perform impressively in English and a handful of dominant languages, while hundreds of others remain almost entirely invisible. And when AI cannot understand a language, it cannot reliably inform, protect or empower the people who speak it — especially in civic and democratic spaces where accuracy matters the most.

This is not a niche fairness issue but rather a growing democratic fault line.

A new divide

For decades, the digital divide was defined by access to internet connectivity. But a new divide is now emerging across connected societies: the divide between the languages AI can “see” and those it can’t. And it’s widening fast.

The newly released SAHARA benchmark evaluated 517 African languages across a range of AI tasks. The results were stark: English performs near the top of every task, while many African languages — including those spoken by millions, such as Fulfulde, Wolof, Hausa, Oromo, and Kinyarwanda — consistently cluster among the lowest performers across reasoning, generation, and classification tasks.

The discrepancy appears to have little to do with linguistic complexity. The SAHARA benchmark attributes these gaps to uneven data availability — languages with sparse or incomplete digital datasets consistently score lower, even when they are linguistically straightforward. This reflects decades of underinvestment in datasets, documentation and digital infrastructure. When a model has no training data for a language, they do not degrade gracefully. Research on multilingual AI shows markedly higher hallucination rates in low-resource languages, along with well-documented tendencies to amplify stereotypes and misclassify dialect or culturally specific content.

And the issue is not limited to spoken languages. A recent report by the European Union of the Deaf documents how many AI tools for sign languages misinterpret grammar, flatten cultural nuance and misinterpret even basic expressions. This often happens because they were developed without meaningful Deaf community involvement, reinforcing the sense that accessibility is something bolted on after the fact rather than a democratic requirement.

Across both spoken and signed languages, the evidence points toward a consistent pattern: AI systems struggle most with languages that receive the least political and economic investment — even when those languages serve millions. And that has direct implications for democratic equity.

When AI doesn’t understand you, civic safety falls

Much of the conversation around AI safety focuses on principle alignment, model robustness and misuse prevention. But safeguards to that effect only function when a model can understand the input it receives.

Real-world cases may reveal some of the consequences. Ahead of Kenya’s 2022 elections, a Global Witness and Foxglove investigation found that Facebook approved ads containing explicit hate speech in both English and Swahili — content that should not have passed moderation. The report does not identify the reason for the lapse, but Meta has repeatedly emphasized that AI handles a substantial share of its moderation workload. The incident therefore likely reflects a broader pattern: safety mechanisms break down most sharply in languages where AI systems have limited underlying support. In Nigeria, recent research demonstrates that widely used hate-speech systems dramatically overestimate their performance when applied to political discourse involving code-switching, Pidgin and other local language patterns.

Multilingual safety research consistently shows that content moderation, toxicity detection and misinformation classification degrade sharply in low-resource languages. In practice, this means hate speech or election-related incitement goes undetected in regional or minority languages; benign civic content is misclassified as harmful, suppressing legitimate political speech; and safety filters that work reliably in English collapse in languages with sparse training data.

Evaluations of LLM-based moderation systems reach similar conclusions: when models are asked to assess toxicity across languages, agreement and reliability drop as one goes from high-to-low resource languages defined by how much training data exist. These breakdowns shape the information environment for millions of speakers.

The stakes go beyond platform moderation. The World Bank’s Generative AI Foundations report underscores how AI tools used in health, agriculture and education produce inconsistent — and sometimes dangerous — outputs when prompts are translated into low-resource languages. As public institutions integrate AI into service delivery, these failures become governance failures, not just technical glitches.

And in elections, the risks compound. Research shows that election-related disinformation campaigns already target communities whose languages receive the weakest moderation and safety protections. Public-facing chatbots and "multilingual" portals, often powered by brittle translation systems, have also been shown to routinely misinterpret queries in lower-resource languages, a batten that can lead to incorrect guidance on benefits, rights or political process. When a citizen receives misleading civic information, the effect is indistinguishable from disenfranchisement.

Linguistic inequity in AI is not a peripheral fairness concern. It determines who has access to accurate information, who can contest decisions and who can meaningfully participate in democratic life.

A governance blind spot with democratic consequences

Governments are beginning to sound the alarm. South Africa, during its G20 digital economy presidency, openly warned that linguistic inequity in AI risks excluding billions of people from the digital economy. UNESCO’s recommendation on the ethics of AI underscores the principles of inclusiveness and non-discrimination. Yet major global governance frameworks still treat multilingual capability as optional. For instance, the European Union AI Act — one of the world’s most comprehensive governance frameworks — does not require developers to report or guarantee model performance across the languages spoken by affected users, except in limited consumer-facing cases. Multilingual performance remains largely advisory rather than mandatory — a stretch-goal rather than a baseline requirement.

The deeper issue is that linguistic equity is still regarded as a technical inconvenience instead of a governance priority. Yet the ability of AI systems to understand diverse languages is directly tied to democratic participation, the right to information and equitable access to essential services.

Emerging research on moral reasoning reinforces this point: large language models (LLMs) exhibit systematically different ethical judgements across languages, with the largest deficiencies in low-resource ones. Safety failures are not just about toxicity, but about whose values and voices are heard by the system at all.

Make linguistic equity a hard safety requirement

If AI is to support rather than undermine democratic participation, governments and platforms must stop treating multilingual capability as a bonus feature. Linguistic equity must be understood as a safety principle in its own right — as fundamental as robustness, transparency and accountability.

Policymakers and regulators can move this from rhetoric to practice by:

• Mandating language-specific performance reporting: Model cards and safety evaluations should disclose accuracy and safety metrics for every language the system claims to support, especially when dealing with high-stakes tasks like moderation and civic information. If a model cannot meet minimum thresholds, it should not be approved for those applications.
• Building language datasets as public infrastructure: Investments in digital public infrastructure must include long-term support for datasets for indigenous, minority, regional and signed languages. These datasets should be treated as digital public goods.
• Funding community-led dataset creation: Communities must be compensated and empowered to lead dataset creation, annotation and evaluation. No AI system can meaningfully represent a language without its speakers.
• Enforcing multilingual performance in public procurement: Governments should refuse to procure AI systems that cannot demonstrate reliability in the language the public actually uses. Multilingual capability should be a gatekeeping condition, not a checkbox.
• Recognizing linguistic bias as ad democratic risk: Election commissions and digital regulators should treat linguistic inequity as a vector for misinformation and unequal civic access.

No language should be digitally invisible

AI is becoming the interface between many people and the institutions that shape their lives. But it cannot serve humanity if it only understands a fraction of it. To build AI that is genuinely safe and inclusive, we must align behind a simple commitment: no language should be invisible.

Linguistic equity is not peripheral to AI safety. It is a democratic imperative. And unless regulators and platforms act now, AI will deepen the existing divides — not because of malicious actors, but because entire languages remain invisible to the systems that are shaping the future.