Bias isn’t just in algorithms but in the words we choose

The launch of Switzerland’s Apertus model highlights how even a Latin ending can carry centuries of cultural weight, argues inclusive communication specialist Claudia Vaccarone. 

Switzerland has just taken a bold step in artificial intelligence (AI). The federal technology institutes EPFL and ETH Zurich as well as the Swiss National Supercomputing Centre (CSCS) have launched ApertusExternal link, the country’s first large-scale, open-source, multilingual language model.

The model is truly unique for being multilingual by design: Apertus was trained on 15 trillion tokens (words or parts of words) from over 1,000 languages (out of the 7,000 spoken today in the world).  With approximately 40% of sources being non-English, the model incorporates languages that are generally underrepresented in LLMs, such as Swiss German, Romansh, and many others. 

From a cultural diversity standpoint this is a very important detail, as the vast majority of the main LLMs are still trained primarily in English. That’s because the internet has far more English content than any other language and because the biggest AI companies are US/UK-based, building first for English-speaking markets. In the current iteration of these services, this has consequences: the 2023 report “Lost in TranslationExternal link” by the Center for Democracy & Technology External linkstates that models trained mostly on English-language text “end up transferring values and assumptions encoded in English into other language contexts where they may not belong. For example, a multilingual model might associate the word ‘dove’ in all languages with ‘peace’ even though the Basque word for dove (‘uso’) can be an insult.”

When using LLM-based tools in low-resource languages, more hallucinations and biases can be produced. There is a clear risk of linguistic homogenisation if smaller languages aren’t deliberately included – which is why projects like BLOOMExternal link and Apertus, explicitly designed to prioritise linguistic diversity, are very important. 

Latin grammar and modern bias

Apertus is Latin for “open” – in theory a perfect name for a tool that promises radical transparency: unlike most commercial AI models, Apertus’s code, data and methodology are fully accessible.

Using Latin also gives the name a timeless, academic feel, appropriate for a model developed by leading Swiss institutions such as EPFL, ETH Zurich and CSCS. It conveys digital sovereignty and accessible AI.

However, something intrinsic to the name deserves our attention, especially if we are truly serious about openness.

In Latin, nouns and adjectives take different endings depending on gender: Apert-a is feminine; Apert-um is neutral; Apert-us is masculine.

The Swiss LLM carries the masculine form. Dictionaries may list it as the “default”, but the choice is not neutral. It reflects a long-standing convention in European languages that treats the masculine as universal – a convention so ingrained that we often don’t notice it. This has contributed to making invisible the feminine (think of professions only known in the masculine form) and the crystallisation of a patriarchal hierarchy between the sexes. 

Recent solid sociolinguistic researchExternal link has proven how the human brain cannot deal with the ambiguity of the masculine as universal and interprets the masculine literally.

Inclusive language movements have been working at highlighting the need to dismantle the universal use of masculine forms, favouring gender-neutral forms.

Why the choice of Apertus matters

Choosing Apertus instead of Aperta or Apertum is not a mistake, per se. 

It is unfortunately a reflection of how language itself encodes gender hierarchy, especially in the scientific world, presenting the masculine as the norm and everything else as a deviation. It reflects the broader linguistic habit of treating the masculine as universal. That habit itself is a form of embedded bias – it naturalises male-coded language as “neutral”.

It is a missed opportunity to extend gender inclusion to the brand of this innovative and linguistically inclusive AI tool and to acknowledge the impact and power of language.

One could assume that the choice of the masculine form in Apertus aligns with the masculine of the word “model” in gendered languages (le model in French, il modello in Italian). But when I read Apertus, as a multilingual, European classicist, I see more than grammatical gender; I see a pattern of linguistic choices that reflect a societal bias.

It would be interesting to know more about the creative process leading to the choice of this brand name: was the ideation process fully human, or was it a suggestion from one of the existing LLMs? Did Apertus, perhaps, name itself? 

Out of curiosity, I asked two LLMs – ChatGPT and Gemini – what they thought of the name. Both praised the elegance and clarity of purpose of a Latin word. But when pressed on gender implications, their responses diverged. ChatGPT acknowledged the bias as “subtle and systemic”. Gemini, at first, dismissed the issue as irrelevant because “Apertus is a name, not a person” and “the majority does not know Latin”. Only after further prompting did it concede that the masculine default could reinforce exclusionary norms.

The tech world has a well-documented history of gender imbalance. A name like Apertus subtly reinforces the idea that technology is a masculine domain. It’s a small signal, but it’s part of a larger pattern. Every time a new product or project defaults to a masculine name, it contributes to the narrative that men are the primary creators and users in this field.

A gender-neutral and human rights-based approach

Had the model been named using a gender-neutral and human rights-based approach, the name might have resonated differently. Aperta could have conveyed openness also as welcoming, inclusive and human, whereas Apertum could have underlined neutrality and abstraction, echoing Switzerland’s own tradition of neutrality.

One of the biggest risks and fears around AI technology and tools is that they can reproduce the inequalities of the societies and organisations that build them. But bias does not live only in algorithms or datasets – it lives in language itself.

Acknowledging this reality and reflecting intentionally on our use of language provides a unique chance to model a different path. In this way, Apertus can become more than a tool. It can become a lens through which we examine our society’s defaults.

Openness, after all, is not just about transparency. It is about the courage to question the norms that often make half of the population feel invisible. Switzerland has the chance to show that true innovation means opening not just our code, but our minds.

Edited by Veronica DeVore/ts