As the world’s tech leaders prepare to gather for the India AI Impact Summit on February 16, one topic is no longer optional: language equity. In a country like India, where linguistic diversity isn’t just a feature but a way of life, the "language gap" in AI isn't just a technical glitch—it’s a barrier to safety, participation, and economic opportunity.

Of roughly 7,000 global languages, only a handful are thriving digitally, with a mere 10 languages making up 82% of all internet content. This massive resource disparity is the core challenge for content moderation in non-English languages. Before Large Language Models (LLMs) emerged, moderation tools and human reviewers already struggled, often failing to capture the linguistic complexity of transliteration, code-switching, and sophisticated algospeak, which is common in non-English content.

The rise of LLMs presents a paradox. While these systems are often hailed for being language-agnostic, research shows they are built on a foundation that heavily favors English and a few other dominant languages, creating a typological echo-chamber. This creates a "poor get poorer" cycle in the digital space. High-resource languages get the best moderation tools, the most accurate chatbots, the safest filters and dominate performance leaderboards. Meanwhile, communities speaking "low-resource" languages are left with tools that don't understand their slang, their cultural nuances, or their safety risks.

How should platforms distribute finite investment resources across different languages? Most platforms default to a utilitarian or welfare maximizing approach. They look at where the most users are and put their money there. On paper, this makes sense for shareholders: invest in English and Spanish first because that’s where the volume is. However, this approach entrenches existing digital exclusion, fails to mitigate risks for smaller, highly vulnerable communities and leaves user trust and safety gaps unaddressed. In the long run, it also limits market expansion and user retention.

An alternative approach: focusing on capabilities

Instead of just counting heads, we propose a capabilities approach. This means shifting the focus from inputs—how much money we spend, to outcomes- what users can actually do. Language "support" should not mean clunky or tone-deaf translation. Real equity means a user in a rural village in South East Asia should have the same level of safety and agency on a platform, as a user in New York. For example, relying on machine translation for low-resource languages might look like language coverage, but it fails to capture critical contextual, cultural, and idiomatic nuance—and is sometimes just plain wrong.

We believe that bridging the language divide isn't just a matter of how much money you throw at it. It’s about where that money goes. "Investment should be deployed toward strengthening digital capabilities of language communities, rather than brittle off-the-shelf translation tools that don't expand what communities are capable of. A capability-enhancing investment would open doors for people to actually participate in the digital spaces—whether that’s finding reliable info, joining a conversation, or sharing their own culture. Such investments might look like: adapting a model to encode local cultural awareness (making sure a model understands local social norms, traditional foods, or the importance of a specific festival) and engaging community partners (working with people on the ground to create high-quality data that reflects how they actually speak). Since no budget is bottomless, we suggest a "risk-based" approach. By targeting communities that need it most, we can find cost-effective ways to make a huge difference, backed by clear metrics that prove the value to both society and the business.

Risk-based targeting

We suggest using a simple formula to find out who needs help most: R = f(Vulnerability, Service Impact), where

• Vulnerability: Does this community face social instability, conflict, or high numbers of younger users?
• Service impact: How much do they rely on the platform, and how bad is the current quality of the AI they are using?

Populations that are highly vulnerable and experience high service impact present the potential for the greatest 'capability gains'. This allows platforms to make continuous, incremental, and targeted investments. Examples of cost-effective interventions include:

• Community-in-the-loop: micro annotation sprints with local speakers and cultural experts to create small, high-quality fine-tuning datasets for priority low-resource languages.
• Artificial corpora generation: Using LLMs to annotate and parse cultural datasets, Creating datasets that replicates low-resource language features using generative models and culturally curated seed data
• Multi-agent interaction: Simulating cross-cultural dialogue to generate diverse, high-quality cultural datasets that improve understanding and alignment
• Leveraging regional models: Utilizing smaller, specialized multilingual models developed by local actors (e.g., AfriBERTa, SEA-LION, IndicBERT), which are more cost-effective, require fewer computational resources, and avoid the biases of larger models

Measuring progress: core metrics for language equity

How will platforms know they’re making progress? We suggest measuring at both ends of the investment: baseline assessments on the language equity gap (which should guide investments and be monitored for improvement over time) and post-facto assessments that show impact in quantifiable business outcomes and societal justice. Detailed tables are attached to this article with full descriptions of suggested metrics, alternatives and their limitations, but here is the gist of what we should be looking at:

1. State of Equity, or Estimating the Language Equity Gap (Table 1 below). In order to measure how “digitally ready” a language is, we suggest three “buckets” of metrics to track.
   1. Resource parity, using tools like the Digital Language Equity (DLE) metric. This helps us see which languages are starving for data. If a language has a lot of speakers but a low DLE score, that’s a red flag for investment.
   2. Model performance and fairness, to compare moderation outcomes across languages. Metrics like F1 Parity compare how well an LLM moderation model works in English versus another language. F1 score is a standard accuracy metric adopted by most model providers to balance precision and recall. F1 parity requires the same standard of performance across languages. We also look for equalized odds—basically, is the model accidentally flagging "safe" content as "harmful" more often in some languages than others?
   3. Cultural competence: This is the most overlooked part. Does the model understand local food, festivals, or social norms? New benchmarks (like DOSA for Indian cultures) help us test if an LLM is actually "smart" or just repeating Western biases.
2. Measuring the real world impact, in business and societal terms (Table 2 below). We look at four main areas to see if changes are actually working. While some of these are hard to measure and require expert audits, they provide the return on investment (ROI) story that set up language equity as a core business priority rather than a side project.
   1. Reducing real-world risk: We track the capability gain—basically, how much better the model performs after we give it a boost with local community data. We also measure if the overall "Risk Function" for a community is dropping over time.
   2. Encouraging user participation: One of the biggest hurdles for non-English speakers is false positives, that is, when the model removes a harmless post because it doesn't understand the slang or the script (like transliteration). If persistent, such removals can have a silencing effect on a language community. We also look at appeal parity: if a user in Ethiopia appeals a deleted post, do they have the same chance of a fair review as someone in Germany?
   3. Building transparency: When content is removed, users deserve to know why. This metric tracks whether platforms are providing clear, culturally relevant explanations in the user's native language rather than just a generic template.
   4. The business bottom line: Language equity is good for growth. We track "engagement rate lift" (do people post more when they feel safe and understood?) and "safety-related churn" (how many users are leaving the platform because they’re tired of being misunderstood or harassed?).

Conclusion

The current utilitarian approach to internationalization is regressive. It keeps the best tools in the hands of the few while leaving the rest of the world to hobble along. But as we head into the India AI Impact Summit, the goal is clear: we need to move from an "English-first" mindset to an "equity-first" reality.

Using our more progressive, risk-based strategy, even platforms with modest budgets can stop the language gap from widening. The Risk Function R = f(Vulnerability, Service Impact) enables focus on where the vulnerability is highest through cost-effective solutions like smaller regional models and local community partnerships. In future work, we will explore next-gen AI solutions like using generative models for artificial corpora and multi-agent dialogue simulations to augment and fill gaps with sourcing community partners.

Ultimately, investing in language equity isn't a passion project. We believe language equity is good business. When platforms work better for more people, user churn goes down (people don't leave because of trust and safety issues) and engagement goes up. The platform is “sticky” and the community thrives. Tools for LLM-driven content moderation are becoming more widely adopted everyday and this article proposes a set of metrics to help make them work better than traditional review methods. Now, we just need the collective will to point these tools in the right direction.

* * *