# After Babel comes AI: why monolingual thinking erodes competitive advantage **Authors:** Italo Marconi **Categories:** Opinion **Tags:** AI **Last Updated:** 2026-06-05T06:34:10.934Z **Reading Time:** 11 min read --- ## Summary AI thinks in one language. Italo Marconi argues that by delegating thinking to language models trained on Anglophone data, companies are silently eroding their cognitive diversity and competitive edge, without even noticing it. --- Over 7,000 languages are spoken worldwide. Fewer than 100 have sufficient digital presence to meaningfully influence the training of Large Language Models (LLMs). Roughly 90% of Common Crawl data, the primary corpus LLMs draw from, comes from just twenty languages. These numbers, documented in a recent Microsoft Research paper (Zamir et al., BYOL: Bring Your Own Language Into LLMs, 2025), point to a strategic problem most companies are not yet seeing: **when an organization delegates thinking to AI, it delegates thinking to a system that thinks in one language only.** # The problem: cognitive monocultures The AI tools that companies are adopting at scale create what I call "cognitive monocultures." When every organization uses the same language models, trained on deeply imbalanced corpora, **they are not simply adopting a tool: they are adopting a shared way of framing problems, categorizing options, and evaluating outcomes.** Microsoft's paper classifies languages into four tiers of digital resources, from Extreme-Low to High, and demonstrates that models systematically underperform on low-resource languages, producing not just linguistic errors but genuine cultural misalignment. Languages spoken by millions of people, such as Saraiki or Kituba, remain virtually invisible to AI. The Sapir-Whorf theory of linguistic relativity tells us that the language we speak shapes the way we understand and perceive the world. The business implication is more radical than it appears: **companies that use AI to develop strategies, analyze markets, or generate ideas are progressively thinking in the same language, even when they believe they are differentiating.** Some will object that English dominance in business is nothing new. Multinationals have operated in English for decades. The cognitive compression existed before AI. This is true. But AI changes the problem not just in degree but in kind, for at least three reasons. 1. First: In the pre-AI era, **English was the medium of communication, not the medium of thought.** A Japanese manager writing a report in English still thought in Japanese: ***nemawashi*** remained their cognitive frame, even if the word never appeared in the final document. The linguistic filter acted on the output, not the input. With AI, the filter shifts upstream. When that same manager asks an LLM to analyze a problem, generate strategic options, or evaluate scenarios, thought itself is generated within the cognitive structures of English. **They are no longer translating their thinking. They are outsourcing cognition to a system whose entire conceptual repertoire is Anglophone.** 2. Second reason is scale. Pre-AI linguistic globalization was slow, partial, mediated by human beings who carried their cognitive diversity with them, consciously or not. AI operates simultaneously across millions of interactions, in real time, without mediation. **What the spread of English as a lingua franca accomplished in fifty years, language models are doing in five.** And they do it with an efficiency that makes the compression invisible: the output is fluid, grammatically impeccable, apparently competent. **It does not feel like a loss. It feels like an improvement.** 3. Third is the feedback loop. AI-generated content becomes training data for the next generation of models. Each cycle compresses diversity further. It is a spiral: **less linguistic diversity in the data produces less cognitive diversity in the outputs, which produces less linguistic diversity in new data. Within a few cycles, the flattening becomes self-reinforcing.** Pre-AI linguistic globalization did not have this amplification mechanism: people continued to think, write, and create in their own languages, feeding a diversified cultural ecosystem. AI risks breaking that circular model. This is particularly visible in marketing and communications, the sectors that inherently operate with and through language. Today a growing share of copy, campaigns, content strategies, and even brand positioning are generated or co-generated through LLMs. The result is a paradox: at a moment when companies have more channels, more touchpoints, and more opportunities to communicate than ever, they are converging toward the same lexicon, the same narrative structures, the same tone of voice. Anyone working in marketing already feels it: a growing indistinguishability of messages, a stylistic homogenization that cuts across sectors and geographies. **It is not a lack of creativity. It is cognitive monoculture applied to communication.** # What gets lost: words without equivalents Every language encodes cognitive strategies, relational logics, temporal structures, and spatial reasoning that open entirely different solution spaces, and that do not exist in English. Some encode alternative decision-making and relational logics. The Japanese ***nemawashi*** is a philosophy where the meeting is not where decisions are made, but where they are ratified. The Korean ***nunchi*** is the ability to read the emotional and power dynamics of a room instantaneously, faster and more collective than "emotional intelligence." The Chinese ***guanxi*** is a network of moral reciprocal obligations, not the instrumental "networking" of the West. Companies that treat it as "networking with Chinese characteristics" fail systematically. Others open different innovation and leadership spaces. The Hindi ***jugaad*** condenses an entire entrepreneurial mindset of frugal innovation. The French ***bricolage*** describes creating with whatever is at hand, recombining existing materials in unexpected ways: the heart of startup culture. The Finnish ***sisu*** describes a stubborn determination that goes beyond "resilience" and "grit": the capacity to act with courage when the situation seems rationally hopeless. The Zulu ***ubuntu***, "I am because we are," redefines change management as something that emerges from the relationship between people, not something imposed upon them. When an organization operates solely through Anglophone concepts, **it is thinking within the boundaries of a single epistemological tradition.** # The business impact: market blindness and epistemic erosion The consequences are concrete. A brand entering the Japanese market without understanding ***omotenashi***, anticipatory hospitality, caring for the customer before they express a need, builds its communication on a relational model that falls flat in that market. No LLM trained predominantly on Anglophone data will flag this, because it lacks the categories to do so. The same applies to markets where the relationship between brand and person is structured by concepts like ***ubuntu*** in sub-Saharan Africa, which implies a theory of community incompatible with the aspirational individualism on which most Western branding is built. Inside organizations, **AI is performing a quiet, systematic impoverishment of collective intelligence.** A meeting in three languages is summarized in one. A market report based on local knowledge is filtered through the ontology of an English-language model. The result looks efficient, but **the coherence within a single cognitive frame is not the same thing as wisdom.** Walter Quattrociocchi, in a study published in *PNAS* in 2025 (*The Simulation of Judgment in LLMs*), introduces a concept that illuminates this risk: "epistemia", a condition where the appearance of knowledge, fluid, coherent, plausible, replaces knowledge itself. **LLMs do not understand: they predict.** They produce statistically plausible text, not thought. But their fluency is so convincing that we stop asking how we know what we know, deferring judgment and losing the habit of doubt. Applied to our argument, epistemia is the mechanism through which conceptual homogenization becomes invisible. Consider what is happening in marketing departments. An insight from qualitative research conducted in Indonesian is synthesized in English by an AI; a creative brief is built on that synthesis; another AI generates the campaign concept. Each step is a compression. The original insight, which perhaps contained a nuance tied to ***gotong royong***, the spontaneous communal cooperation that is a cultural cornerstone of Indonesian society, arrives at the creative director as a generic "community-oriented values." The brief is technically correct. The campaign is technically correct. But it could have been written for any market in the world, which in marketing means it was written for none. **The apparent fluency of AI hides the flattening. Everything sounds right. Nothing truly resonates.** # What Steiner, Eco, and Lakoff teach us George Steiner, in *After Babel* (1975), inverts the reading of the Tower of Babel. The dominant premise in Silicon Valley is that linguistic multiplicity is an engineering problem, an obstacle to overcome, a friction to eliminate. Steiner's central question is not "how do we translate?" but "why do so many languages exist?" His answer is radical: **linguistic multiplicity is not noise. It is signal.** Languages serve not only to communicate, but to generate alternative worlds, to think what is not, to imagine what could be. Each language is a different proposal for how to structure reality. Steiner describes four movements of translation: trust, aggression, incorporation, and restitution. Language models perform only the third movement, and that poorly. There is no trust, no assumption that the other language contains something irreducible. No aggression, no effort to confront otherness. And above all, no restitution, no awareness of loss. When an LLM "translates" ***nunchi*** into "emotional intelligence" or ***guanxi*** into "networking," **it is not translating. It is erasing.** And it does not know it. George Lakoff and Mark Johnson, in *Metaphors We Live By* (1980), demonstrate that metaphors are not figures of speech but the fundamental structures through which we think. In English, business is dominated by military and mechanical metaphors: markets are "conquered," competition is "attacked," processes are "optimized" as if the organization were an engine. In Japanese, business is often thought through natural metaphors: the company as a living organism, strategy as cultivation, change as a season, metaphors that produce longer time horizons, tolerance for ambiguity, and attention to cycles. **An LLM trained predominantly on Anglophone texts does not merely think in English: it thinks through the metaphors of English,** reproducing those invisible cognitive architectures and excluding all others. Umberto Eco, in *Dire quasi la stessa cosa* (2003), shows that translating is negotiating: consciously deciding what to save and what to sacrifice. **A language model does not negotiate. It executes. It loses without knowing it loses.** Eco also introduces the concept of reversibility: a good translation is one from which the original could theoretically be reconstructed through retranslation. Apply this to AI outputs: take the "translation" an LLM makes of ***nemawashi*** into "consensus building" and try to retranslate it back into Japanese. You will never get ***nemawashi***. Reversibility is zero. This tells us something precise about the depth of the loss. # The strategic response **Linguistic diversity is not a problem to solve: it is a cognitive infrastructure to preserve and leverage.** The strategic answer is not simply adopting multilingual models, a necessary but insufficient step. It means treating linguistic diversity the way smart companies treat biodiversity in supply chains: as a factor of resilience. Concretely, it means developing organizational analysis tools to identify where AI is silently homogenizing thinking. It means designing processes where non-Anglophone conceptual frameworks hold real authority in decision-making, not just symbolic representation. It also means recognizing that speakers of endangered languages are not merely cultural treasures: they are carriers of cognitive strategies that no dataset currently preserves. For marketing and communications, the response is more specific still. It means ceasing to use AI as an author and beginning to use it as an interlocutor, a tool that accelerates execution, not one that replaces cultural intuition. It means insisting that creative briefs for non-Anglophone markets be conceived in the language of that market, not translated from English. It means investing in teams that think in different languages, not in monolingual teams that use translation tools. Steiner closes *After Babel* with a prophetic image: perfect translation, if it were possible, **would not be a conquest but a catastrophe,** because it would mean there is nothing unique left to translate. And Eco, with characteristic ironic pragmatism, would remind us that the value lies precisely in the "almost," in that gap between languages that no algorithm can bridge, because it is not a bug in the system. It is the system. The equivalent in contemporary business is a world where every organization thinks through the same language model, generates the same strategies, sees the same opportunities. Perfect efficiency. Zero differentiation. ***The curse was never Babel. The curse is completing the tower.*** # References Eco, U. (2003). *Dire quasi la stessa cosa: Esperienze di traduzione.* Bompiani. Lakoff, G., & Johnson, M. (1980). *Metaphors We Live By.* University of Chicago Press. Quattrociocchi, W. et al. (2025). The Simulation of Judgment in LLMs. *Proceedings of the National Academy of Sciences (PNAS).* Sapir, E. (1929). The Status of Linguistics as a Science. *Language,* 5(4), 207–214. Steiner, G. (1975). *After Babel: Aspects of Language and Translation.* Oxford University Press. Whorf, B. L. (1956). *Language, Thought, and Reality: Selected Writings.* MIT Press. Zamir, A. et al. (2025). BYOL: Bring Your Own Language Into LLMs. *Microsoft Research.* >This article is an abridged version of an [article originally published on LinkedIn.](https://www.linkedin.com/pulse/dopo-babele-viene-lai-quando-la-monolingua-pu%C3%B2-mettere-italo-marconi-nlpdf/) ## Key Takeaways 1. AI does not just communicate in one language: it thinks in one language, compressing the cognitive diversity of every organization that uses it 2. Linguistic monoculture is not new, but AI shifts the filter upstream: companies are no longer translating their thinking, they are outsourcing it 3. Words without English equivalents, like nemawashi, nunchi, or ubuntu, encode entire decision-making frameworks that LLMs silently erase 4. The fluency of AI output hides the flattening: everything sounds right, nothing truly resonates 5. Linguistic diversity is not a problem to solve: it is a cognitive infrastructure to preserve and leverage --- *Article from [Albert's Deep Dive](https://deepdive.albertschool.com) - Albert School's Journal*