AI on the Job training in Geneva · canton of GenèveDoes AI really understand when your team thinks in two languages?

In organizations where managers alternate French and English within a single sentence, generative AI appears at ease. It translates, summarizes, formulates. But does it genuinely absorb the cultural and professional context that gives meaning to language mixing? Research on cognitive load offers a more nuanced answer.
What research establishes about multilingual processing
John Sweller (1988, Cognitive Science) demonstrated that the human brain has limited capacity to process multiple information streams simultaneously. When a person switches between two languages, she already uses this resource to arbitrate between language codes — a real cognitive load. Generative AI displays unified performance that can mask this tension: it treats French and English as text strings, not as two distinct cognitive and professional universes. Therefore, it can propose formulations that are technically correct but culturally misaligned or contextually fragile.
The mechanism of lost nuance: the hidden cost
Fred Paas, Alexander Renkl, and John Sweller (2003, Educational Psychologist) deepened this phenomenon: when a tool reduces cognitive load at an early stage, the user invests less attention in qualitative filtering. In a multilingual context, this means the team more readily accepts an AI proposal — for example, an English email drafted for a bank — without verifying whether the register truly matches the client relationship or internal tone. The machine lightens surface work and conceals the cognitive burden that humans should retain: judgment about linguistic and cultural fit.

What we conclude too quickly: is accuracy enough?
It is tempting to believe that if AI produces a grammatically correct sentence in English or French, it has understood the context. Yet Avraham Kluger and Angelo DeNisi (1996, Psychological Bulletin) show that perceived performance and actual performance regularly diverge: we judge fluent output as good. In a multilingual environment, a 'correct' formulation may be inappropriate — too formal for a colleague with whom you mix languages, or too casual for an external partner. AI does not see this boundary, and the person validating it risks overlooking it too, precisely because the tool has already reduced their mental load. (our executive and employee training programmes)
Limits of this finding: context and experience
Studies on cognitive load and multilingualism arise largely from controlled laboratories or educational settings. The reality of international organizations is more composite: some teams develop genuine fluency in code-switching and instinctively detect contextual errors from AI. Conversely, newcomers to multilingual environments may be more easily seduced by machine output that seems to master both codes. The effect is neither universal nor predictable without observing the actual team.
A practical check in Genève
Before entrusting AI with composing a message for a mixed network (French-English), identify three criteria: (1) the register of language expected by the recipient (formal/informal, French/English); (2) if the message mixes languages in the original, check whether AI has unilaterally decided; (3) ask a peer from your team, native or deeply embedded in the cultural context, to validate a first draft. Over two weeks, document cases where AI proposed 'technically correct but contextually off-target.' This is a tangible indicator of the tool's limitation when facing authentic multilingualism — where languages are negotiated, not merely juxtaposed. AI training in the canton of Geneva. To go further: explore the AI on the Job training in Geneva in the canton of Geneva, or browse our executive and employee training programmes in Switzerland.
In pictures: AI on the Job in Geneva



- ai on the job
- Geneva
- research
- training geneva
