The Monoculture Problem in Artificial Intelligence

Sonja Schmer-Galunder uses the historical example of German forestry monoculture to illustrate how optimising a system for a single metric—timber production—by removing all diversity eventually leads to systemic collapse. She draws a direct parallel to AI, arguing that a lack of diversity in models and their outputs makes both human society and the models themselves less resilient and more brittle.

She describes how, at the beginning of the last century in Germany, trees were planted in perfect geometrical positions to maximise timber production and facilitate taxation, while all undergrowth, bushes, and insects were removed . This worked well for about 100 years, maximising revenue, but ultimately led to the complete collapse and dying of the forest . She explicitly draws the analogy to AI, warning of correlated failure risks when models share similar training data, languages, and weights .

Sonja Schmer-Galunder argues that the 2007–2009 financial crisis is a powerful analogy for the dangers of AI monoculture, where reliance on a single flawed mathematical model across all financial institutions wiped out enormous wealth. The model optimised away the very variability that would have allowed the system to absorb shocks, making it non-resilient.

She states that between 2007 and 2009, the United States lost 11 trillion dollars in household wealth due to a financial crisis built on a single mathematical model . Every financial institution assumed housing prices were correlated with local economic conditions, when in fact they had become correlated with mortgage lending norms that the models were not measuring . She references Nassim Taleb's work on how markets became non-resilient by optimising away variability .

Sonja Schmer-Galunder uses the biological process of embryonic development to argue that healthy growth is not about proliferation or size but about differentiation and integration of difference. She suggests this principle—that maturation means specialisation and interdependence—should inform how we think about AI development and diversity.

She explains that in early embryonic growth, the initial phase involves proliferation of cells, but true maturation means differentiation, where cells specialise, form relationships, and become interdependent systems . She concludes that from the earliest beginning of life, healthy development is integration of difference rather than mere size .

Sonja Schmer-Galunder argues that Western bias in AI models produces outputs that are culturally inappropriate for other contexts, even if not catastrophically harmful. She uses concrete examples to illustrate how models trained predominantly on Western data fail to account for the lived realities of people in other parts of the world.

She gives examples of culturally inappropriate AI outputs, such as recommending picnic activities in countries where temperatures reach 50 degrees and nobody goes for picnics, or suggesting poetry based on bird songs in countries where birds are considered annoying . She also mentions her work with Google on an evaluation dataset called AntroBench, designed to provide cultural evaluations for AI models .

Sonja Schmer-Galunder identifies a catch-22 in the debate over data inclusion: on one hand, countries risk having their data extracted without reciprocal benefit; on the other hand, withholding data risks the permanent loss of endangered languages and cultures. This tension is particularly acute for oral data, which is harder to collect and more easily lost.

She describes the dilemma where countries giving up their data for free may have a product sold back to them at a cost, raising concerns about data colonialisation . Conversely, some countries may want their data included precisely to prevent its loss, especially for oral traditions that are more difficult to capture and more easily disappear .

Sonja Schmer-Galunder argues that the culture of data extraction is not a peripheral problem but is embedded at the very foundation of AI development, as demonstrated by the mass scanning of books and collection of personal data without explicit consent, even within Western societies. She questions how this deeply ingrained culture can be changed, particularly for the Global South.

She notes that Stuart Russell, speaking at an earlier panel, acknowledged that every book ever written has been scanned into AI models, which, while providing a large volume of data, does not solve the diversity issue as it remains largely English-language content . She also observes that even within Western cultures, individuals have not been asked for consent before their data was incorporated into AI models .

Sonja Schmer-Galunder argues that tech diplomacy summits are a necessary mechanism to equip Global South countries with the negotiation skills needed to engage with technology companies that now wield state-like power. She frames this as going beyond simply having a voice at the table to developing substantive diplomatic capacity.

She references a global summit on tech diplomacy-described as the second global summit for tech diplomacy-aimed at equipping countries of the Global South not just with a voice but with real negotiation skills and diplomatic tools to negotiate terms with technology companies that act like state actors .

Sonja Schmer-Galunder draws a qualitative distinction between recoverable and irrecoverable forms of systemic collapse, arguing that the loss of epistemic and cultural diversity through AI monoculture may be irreversible in a way that financial crises are not. She uses the contrast between the financial market recovery and the permanent death of monoculture forests to make this point.

She notes that while the financial markets recovered from the 2008 crash, the monoculture forest never recovered, illustrating a special qualitative distinction between these two types of systemic failure . She warns that if current conditions do not allow for sufficient diversity, future societies may lose the adaptability needed to respond to outside shocks, constituting potentially irreversible damage .

Supheakmungkol Sarin argues that AI models developed in the US or China using predominantly Western data are fundamentally ill-suited to respond appropriately to under-resourced cultural and linguistic contexts, creating significant safety and security risks. Because the models are designed to answer questions for a Western cultural context, their outputs in other contexts will be inappropriate or even harmful.

He explains that models are currently developed in the US or China with dominant data, and will therefore not be designed to respond or give appropriate answers for other contexts . He states that in most cases, outputs will be catered towards dominant probabilities suitable for other contexts, and that this creates security and safety issues because the model is not designed to fit the purpose of under-resourced communities .

Supheakmungkol Sarin highlights that of approximately 7,000 languages in the world, only around 100 are represented in current AI models, meaning the vast majority of the world's linguistic and cultural heritage is entirely absent. He emphasises that languages are not merely communication tools but carriers of distinct ways of thinking, values, and cultures.

He references a discussion at the Tech Diplomacy Conference in Paris where the question of language representation in AI models arose, noting that of approximately 7,000 languages, only about 100 are represented in models . He stresses that the missing languages carry not just linguistic content but distinct ways of thinking, values, and cultures that are entirely unrepresented .

Supheakmungkol Sarin argues that the solution to cultural under-representation in AI is not post-hoc fine-tuning or localisation but ensuring that diverse cultural data, thinking, and values are incorporated at the very design stage of model development. He contends that localisation—making a model speak a language—does not equate to genuine cultural representation.

He argues that the way to solve the problem is not fine-tuning at a later stage or localising the model, but ensuring that the data of diverse cultures is represented when the model itself is being developed, at the very design stage . He illustrates the inadequacy of localisation by noting that a model can speak a language while still representing the ideology or context of another culture .

Supheakmungkol Sarin argues that communities must first engage in design thinking—identifying what data, language, and cultural values they want represented—before contributing to model development. Simply adding more data without this prior reflection does not address the underlying problem of cultural misrepresentation.

He states that communities must start from design thinking: identifying what data represents their community and culture, and how to bring that knowledge into a model . He contrasts this with the current approach of simply trying to get more data in and make it work, arguing that speaking a language alone does not solve the problem of cultural representation .

Muhammadou Kah argues that a significant portion of valuable knowledge assets in many parts of the world, particularly in Africa, is oral rather than written, making their inclusion in AI models especially challenging yet critically important. The exclusion of this oral knowledge means that entire bodies of cultural, sectoral, and heritage knowledge are absent from AI systems.

He notes that quite a number of useful knowledge assets in his part of the world are not even written, raising the question of how verbal and visible data can find its way into learning models . He emphasises that these oral knowledge assets span sectoral areas including health, agriculture, wisdom, and values, and that they need to be captured in AI models .

Muhammadou Kah uses the example of Cameroon, which has over 200 languages, to illustrate that the scale of linguistic and cultural diversity within a single nation far exceeds what current AI models capture. Each of these languages carries distinct cultural heritage and sectoral knowledge that must be incorporated into AI systems.

He uses Cameroon as an example, noting that it has over 200 languages, each representing a distinct culture and heritage with embedded knowledge assets . He explains that these knowledge assets extend beyond mere exchanges into sectoral data covering health, agriculture, wisdom, and values .

Muhammadou Kah argues that building capacity and competence in the Global South is essential to enable informal and non-literate populations—who represent the majority in many countries—to contribute their data to AI models. Without this investment in human capacity, the data of the majority will remain excluded.

He states that the centrality of building capacity and competence needs to be revisited in ways not previously considered, and that there is an informal population with data that needs to get into models . He points out that in many Global South countries, the majority of the population is not formally educated, meaning literacy rates skew towards those without formal education, making capacity building all the more critical .

Muhammadou Kah argues that the Global South must firmly reject data extraction and instead negotiate rules of engagement that retain data ownership while establishing equitable, win-win economic models for data utilisation and benefit-sharing. He frames this as a matter of sovereignty and economic justice.

He states that data extraction is an evolving reality that must be rejected, at least from the Global South's perspective . He argues that the trade-off between extraction and preservation depends on negotiating rules of engagement that retain data ownership and establish an equitable, fair model for sharing the value extracted from data .

Muhammadou Kah draws a direct parallel between historical natural resource extraction and the emerging risk of data extraction, arguing that the Global South must not allow the same pattern to repeat itself with data. He contends that the Global South's data and natural resources give it non-financial intangible assets that can be leveraged as negotiating power.

He describes how historically, natural resources were extracted from Global South countries only for the value to be sold back to them at a higher cost, and insists this must not happen with data . He argues that the Global South's inability to convert data into value should not result in giving it away for free, and that sophisticated AI models still need Global South data and natural resources to function, providing negotiating leverage .

Muhammadou Kah argues that multilateralism must play a central role in norm-setting and codes of conduct to deter extractive data practices, since many Global South countries lack the regulatory capacity to do so unilaterally. He sees multilateral forums as the key arena for establishing fair rules around AI and data governance.

He acknowledges that some argue the Global South has no choice because data can be taken without their knowledge or capacity to prevent it, which is precisely why the ethics, transparency, and responsibility of AI matter . He argues that multilateralism can encourage states and private sector actors to think differently about these issues, and that the Global Dialogue on AI exists to ensure inclusivity and that all voices and perspectives are heard .

Muhammadou Kah clarifies that the Global South is not seeking to exclude itself from AI development or to prevent data sharing, but rather to co-shape the rules governing AI on fair, equitable, and transparent terms. He argues that inclusive models will ultimately produce better outcomes for all, including more adaptable AI for local health and agricultural contexts.

He states that the Global South is not saying the data is none of others' business, but rather calling for fairness, equity, transparency, and shared benefits in a win-win situation . He argues that models with more representative data will be more adaptable to local health and agricultural realities, rather than relying on synthetic data that creates problems in specific communities .

Muhammadou Kah argues that the loss of epistemic diversity in AI leads to a lack of innovation, the introduction of bias, and systems that appear functional but produce non-optimal outputs, creating systemic risks that can spread globally with serious safety consequences. He frames this as a fundamental threat to the integrity and utility of AI systems.

He argues that lack of representation introduces biases, and biases produce dysfunctional systems that are seemingly functional, with outputs that become non-optimal and create systemic risks that can spread globally with safety concerns and dire consequences . He also notes that lack of diversity triggers a lack of innovation in the system .

Muhammadou Kah argues that true resilience in AI systems can only be achieved if models are comprehensive, systematically inclusive, and broadly representative. He warns that a model fed flawed or geographically narrow data will produce harmful outcomes when applied in different contexts, using the example of medical devices applied in rural African communities.

He states that resilience can only be achieved if it is comprehensive, systematically inclusive, and broadly representative, enabling better innovation, optimisation, and adaptation . He gives the example of a medical device using non-representative data that may work well in one geography but create havoc when applied to a rural village in Gambia, Tanzania, or Nigeria .

Wallace Cheng identifies three interconnected problems arising from AI market concentration: a handful of predominantly English-language AI tools dominate globally; collective human judgment, creativity, and local knowledge may diminish; and small biases in foundational models become global defaults when deployed across critical sectors. He frames these as compounding risks that must be addressed collectively.

He describes the first problem as the dominance of fewer than ten AI tools globally, the majority of which are in English . The second problem is that while individuals may become more efficient, collectively humanity may become less creative and rely less on local knowledge and regional experience . The third problem is that foundational models set up as infrastructure are widely deployed in healthcare, education, military, and defence, meaning a small bias becomes a global default .

Wallace Cheng argues that biased AI models deployed in military or healthcare contexts can cause irreversible harm, such as misidentifying civilians as soldiers or providing clinically inappropriate recommendations across different health systems. He distinguishes between the universality of medicine as a science and the non-universality of healthcare systems, which require local cultural and infrastructural knowledge.

He notes that in healthcare, while medicine itself may be universal, healthcare systems are not, requiring knowledge of different infrastructures, cultural preferences, and economic situations . In military and security contexts, he warns that AI models with biases against certain groups, or that cannot distinguish civilians from soldiers, will create irreversible damages . He states he is focusing specifically on these last two sectors in his work .

Speaker 1 argues that the challenge of AI monoculture goes beyond language to encompass fundamentally different cultural value systems, and that AI must be able to respond according to the culture of the user. Different cultures assign different moral and social weights to demographic groups, and AI systems must reflect these differences rather than imposing a single cultural framework.

He gives the example that in some distant cultures, older people are considered more relevant than babies, while in Western cultures the reverse is true, and this makes a significant difference in AI responses . He also references the historical example of Microsoft's attempt to create a universal encyclopaedia in the 1990s, which immediately encountered the problem that the telephone was attributed to different inventors-Graham Bell in the US, Meucci in Italy, and Popov in Russia-illustrating how cultural framing shapes knowledge .

Speaker 1 argues that rather than providing raw oral or cultural material to AI developers, communities should package their data together with its full cultural context, thereby increasing its value and improving their negotiating position. He warns that providing raw material alone is a win-lose approach that cedes control once the data is extracted.

He argues that transferring raw material is a win-loser approach because once data is extracted, it is very hard to maintain a balanced position . He suggests that communities should provide a package that includes not only raw oral traditions but all the cultural context that improves the value of those traditions, making it easier to achieve a balanced relationship between the two sides . He draws on the European experience of monuments being digitised without fees, resulting in lost commercial opportunities .

Speaker 2 argues that the 2008 financial crash was not merely an accident of model error but was structurally engineered through misaligned incentives and deliberately non-orthogonal financial products, resulting in a planned transfer of wealth from pension funds to hedge funds. This challenges the framing of the crisis as an unintended consequence of monoculture and suggests intentional design.

He explains that the crash occurred because mortgage sales incentives were aligned with volume rather than ability to repay , and that financial products were structured in a deliberately non-orthogonal way, which was clearly known at the time and resulted in a transfer of money from pension funds to hedge funds .

Speaker 2 argues that if only one dominant AI system exists, the diverse cultural data fed into it will ultimately serve a single political or ideological purpose, endangering those who provide that information. He proposes that having multiple AI organisms—rather than a single dominant one—is a preferable and safer architecture.

He suggests that if only one organism exists, the data provided by various cultures will serve one political or ideological purpose, and the more information there is about different ideologies, the greater the danger for those who provide that information . He therefore proposes multiple AI organisms as a solution .