On October 7, Ai2 announced new releases accompanying the Nature publication of its Bolmo research, extending an approach for adapting existing language models. The researchers report promising results; they have not established equal access across languages. Yet their work raises a consequential question: should speaking your own language carry an invisible surcharge when you use AI?

The test of multilingual AI should be the cost and quality of a completed task in each language, rather than the length of its language menu.

Most models divide text into predefined word fragments; Bolmo instead works with the underlying bytes that computers use to represent text. That is the technical detail that matters here. It opens another way to handle language, without settling whether people will receive a better or cheaper service.

The inequality is documented, although its exact size changes with the system. A 2023 study by Orevaoghene Ahia and colleagues, covering benchmarks in 22 languages, found that users of several languages faced higher usage costs while receiving poorer results. Those findings describe the systems studied then, not today’s prices.

A separate study by Aleksandar Petrov and colleagues found substantial differences in how much representation translated texts required—and showed that moving to bytes or characters did not automatically remove them. A new architecture is an opportunity to investigate unfairness, not a certificate that it has disappeared.

Imagine a small shopkeeper trying to understand a supplier’s conditions in Tamil. This is an illustration, not a reported Bolmo deployment. A useful assistant could help her compare terms herself instead of depending on an English-speaking intermediary. The benefit would include privacy, confidence and bargaining power.

But if her language consumes more of a service’s allowance, or produces answers that need repeated correction, the apparent equality of access is misleading. A free tier may run out earlier. An institution with a fixed budget may serve fewer people. A worker may spend the saved time repairing mistakes. These are possible consequences of the cost structure, not established effects of this week’s release.

The economic question therefore extends beyond the supplier’s bill. What does one correctly completed interaction cost, including retries, checking and the user’s time? Who absorbs the difference between languages?

India already has practical reasons to ask. On September 30, the government announced CAG–BHASHINI Anuvaad, a tool intended to support translation of official documents and audit reports, with human oversight. The announcement establishes a programme, not demonstrated accuracy or savings. Its purpose nevertheless makes language quality a matter of public participation: citizens need to understand what institutions say about the use of public money.

Translation professionals and frontline staff could become more productive if AI handles routine drafts. They could also inherit responsibility for errors while losing the time and pay needed to find them. Employers should measure that checking work before claiming a productivity gain. Local language expertise belongs in the production budget.

There is a serious counterargument. Falling prices and better translation may expand access faster than architectural reform. Waiting for perfect equality could withhold useful assistance from people who have little today. That is a reason to deploy carefully and compare outcomes, while continuing to improve the underlying systems.

Open research can help universities and smaller firms experiment without starting from nothing. But usable local services still require computing, skilled people, reliable distribution and support. If those remain concentrated, openness alone will not determine who earns the returns.

Three futures follow. In a constructive one, local businesses and public institutions fund language expertise, share independent tests and pass efficiency gains into broader access. In an uneven one, commercially attractive languages improve quickly while smaller communities receive thin, unreliable services. In a harmful one, institutions remove human assistance and pressure people into English because it is cheaper to serve them that way.

Better text processing also leaves other barriers intact. Someone with low literacy, a disability or unreliable connectivity may need voice, accessible design, offline support or a person. Language coverage should never become an excuse to close those routes.

What does a successful task cost in every language you advertise? Who pays for correction, and can the affected user obtain it? Do your margins depend on underfunded human checking?

Buyers should require results by language: completion rates, waiting time, correction effort and total cost. Public institutions can fund shared tests and pay local experts to challenge them. Founders can price the service people actually receive. Investors can ask whether growth creates local capability as well as revenue.

Speaking your own language should remain a way to exercise judgment, not a penalty for needing help.