The Risks of Language Models in Minority Languages

The dual-edged nature of developing a language model for the Balochi language, weighing potential benefits like improved communication, accessibility, and language preservation against serious risks such as misuse by state actors for surveillance and power consolidation, and the unintentional promotion of linguistic monoculture.

Alex Strick van Linschoten


May 22, 2023

In thinking about my work to put together a language model or some utilities relating to the Balochi language, I thought a fair bit about whether I should even start. At a very high level, we can look at general risks that comes from language models, as highlighted in the 2022 Deepmind paper entitled “Taxonomy of Risks posed by Language Models” which covers

“six risk areas: I. Discrimination, Hate speech and Exclusion, II. Information Hazards, III. Misinformation Harms, IV. Malicious Uses, V. Human-Computer Interaction Harms, and VI. Environmental and Socioeconomic harms” (p.214)

I’ll leave you to check out the paper if you wish. I can think of a number of specific risks and harms that could be connected to developing a language model for the Balochi language, most of which relate to state actors and their desire for more power and better surveillance capabilities.

The communities speaking Balochi have historically and currently been subject to more monitoring than many, either from central governments in Iran and Pakistan or from European and American intelligence agencies, for a variety of reasons. In this context, language models can potentially fit into a system which is geared towards maximising power and influence among the powerful, enhancing state control and surveillance. In the longer term — and I haven’t seen to much by the way of research on this, but I’m going to take a leap — I can imagine that language models could well have the effect of creating a kind of linguistic monoculture. (Just think about the kind(s) of language that you read in default responses from ChatGPT and extrapolate from there.)

My assumption is that large, well-funded intelligence agencies already have strong capabilities for Balochi. Indeed, Balochi is one of the language specified as qualifying for the CIA’s ‘Foreign Language Incentive Program’ that offers cash bonuses for new and current employees with foreign language skills in Balochi. There is almost certainly a team working on — among other things — language models that allows for the better monitoring of communications in the Balochi language. (To get more of a sense of what such models and capabilities might be used for, check out this article.)

What, then, are the positive uses of such technology? In no particular order, some things I thought of in the context of Balochi:

I don’t believe that there’s a simple calculation than can be made, putting potential harm on one side and potential benefits on the other. Given that these models surely do already exist or are being developed by state actors, I also don’t think it’s a matter of staying away from the area entirely. That said, I do think it’s important to have these questions present when doing this kind of work as well as to involve and work within the context of pre-existing community efforts. I’ll turn to other resources and previous work in my next post to give a sense of what low-hanging fruit remains for the Balochi language.