Facts are friendsA commentary on …
Generative AI has access to a small slice of human knowledge | Aeon Essays -
https://aeon.co/essays/generative-ai-has-access-to-a-small-slice-of-human-knowledge
Key Arguments
This essay challenges the claim that AI will create ‘global knowledge collapse’ by showing:
History matters: Humanity marginalised Indigenous and local knowledge centuries before AI existed through colonialism and institutional dominance
AI as curator: New technology can process non-textual data (sensors, video, social media) to recover knowledge human institutions dismissed
Economic reality: The next billion users live in India, Southeast Asia and Africa – companies must build multilingual, culturally aware AI to succeed
The real risk: Knowledge abandonment through political inaction, not technological inevitability
Introduction
Deepak Varuvel Dennison’s article republished as the long read ‘What AI Doesn’t Know’ London’s The Guardian newspaper (18 November 2025) articulates a concern that resonates across technology circles and amongst those discussing the impact of technology: that generative AI threatens a ‘global knowledge collapse’ by marginalising traditional wisdom, privileging Western ways of understanding the world and flattening human understanding into algorithmic homogeneity.
The article follows an Indian-American researcher’s journey from dismissing his father’s traditional Siddha medicine to confronting the profound gaps in AI’s knowledge base. His father’s tumour, which shrank, apparently, through herbal treatment (but equally perhaps through positive thinking) after doctors recommended surgery, becomes a metaphor for what AI cannot yet capture.
The concerns are legitimate. Large language models train predominantly on English text. Common Crawl, one of the largest public training datasets, contains 45% English content whilst Hindi – spoken by 7.5% of the global population – accounts for just 0.2%. Some 97% of the world’s languages are classified as ‘low-resource’ in computing contexts.
However, this pessimistic forecast – that GenAI is destined to flatten knowledge into homogeneous, Anglo-centric mush – overlooks three critical counter-forces already reshaping the landscape:
Historical Context: It attributes to AI a knowledge collapse that humanity achieved centuries ago through colonialism and institutional dominance and centuries before that as we lost the knowledge of the Inca Empires and before that Atlantis (if it ever existed)
AI’s Curative Capacity: It underestimates AI’s unique capacity to synthesise novel insights from previously disconnected knowledge domains
Economic Imperative: It ignores the powerful economic incentives driving AI companies to serve the next billion users in India, Southeast Asia and Africa
The article’s central thesis holds that as GenAI becomes the primary way people access information, mechanisms like ‘mode amplification’ - where AI models disproportionately emphasise the most statistically prevalent responses – will amplify dominant ideas whilst marginalised knowledge fades. Researcher Andrew Peterson warns this could produce ‘knowledge collapse’: a gradual narrowing of accessible information. As I say, fair challenges.
The real danger, however, is not AI-driven knowledge collapse but knowledge abandonment – communities lacking resources to digitise their traditions, researchers failing to secure funding for multilingual AI, or societies neglecting to create frameworks ensuring AI serves diverse populations. These are political and economic choices, not technological inevitabilities. The ubiquity of mobile ‘phones is likely to reduce this risk, too.
Far from a destructive force, AI possesses both the technical capacity and economic motivation to become a curator and catalyst for the world’s forgotten wisdom – provided we make deliberate choices about how to deploy it.
AI as Curator: Why the Knowledge Collapse Predates AI
The original knowledge collapse predates AI by centuries
The Guardian article correctly observes profound power imbalances in the digital corpus, where English and Western sources dominate. However, this observation attributes the symptom – data asymmetry – to AI, rather than the historical cause: human systems. For centuries, colonisation, the standardisation of education and the privileging of print over oral traditions have been the primary engines driving the marginalisation of Indigenous, local and low-resource knowledge systems.
The digital gaps in Hindi, Tamil and Siddha medicine merely reflect this pre-existing, human-enforced cultural hegemony. They existed (or, perhaps better stated, were already disappearing) way before the development of AI. No evidence is presented there decline has been accelerated by AI - just a fear that it may be lost in the noise.
As Māori scholar Linda Tuhiwai Smith emphasises in ‘Decolonizing Methodologies’, colonialism profoundly disrupted local knowledge systems by severing ties to land, language, history and social structures. These processes form part of a broader legacy that continues to shape how knowledge is produced and valued today.
The herbal concoction that cured the Guardian author’s father was dismissed by doctors and his own sister – a trained physician – long before ChatGPT existed. Bangalore’s devastating water crisis stems from decades of sidelining the Neeruganti community’s traditional water management expertise, not from AI’s current emergence in 2022. The glass-facade buildings that overheat in tropical climates were built by architects who ignored centuries of local building knowledge, chasing Western modernism’s aesthetic rather than climatic appropriateness.
The flattening of knowledge was a human achievement carried out over centuries through deliberate (if ill informed) institutional choices.
AI as curator of the long tail
Paradoxically, AI now offers practical tools to reverse this trend. Whilst LLMs currently train on a limited, text-heavy subset of the world’s knowledge, the rapidly expanding Internet of Things and electronic monitoring are fundamentally changing the data landscape. AI possesses a unique capability that distinguishes it from earlier technologies: the capacity to process vast streams of non-traditional, non-textual data.
This matters because traditional knowledge often exists in forms AI can now capture – not just in written texts, but in agricultural practices monitored by soil sensors, in architectural cooling techniques validated by temperature data, in water management systems tracked through environmental monitoring. AI can now assimilate hugde amounts of data: sensor readings from local ecosystems, video recordings of embodied practices, sentiment analysis from multilingual social media, real-time agricultural data from thousands of smallholder farms and patient-reported outcomes linked to traditional remedies. It can capture and interpret exactly those things which sit in the descent – and without record will disappear. Technology can preserve them, not destroy them.
By applying sophisticated pattern recognition to these heterogeneous inputs, AI can synthesise and validate the very local knowledge systems that humans have allowed to fade. The efficacy of a traditional herbal remedy, once dismissed as anecdotal, could be rediscovered and proven by AI cross-referencing global biomedical literature with local ecological data, chemical compound databases and patient outcomes. So, Mr Dennison, I argue that with AI we have a better prospect of actually discovering what it was that cured your father.
Research published in 2025 demonstrates how AI-driven environmental monitoring systems integrate data from IoT sensors to transform raw information into actionable intelligence, enabling real-time decision-making previously impossible for human researchers.
AI becomes a curator of the long tail – not merely an amplifier of the popular. Where human institutions marginalised knowledge through deliberate exclusion, AI can recover it through systematic inclusion.
Concrete evidence of preservation in practice
The transformation is already underway.
Te Hiku Media in Aotearoa New Zealand has developed Papa Reo, the first automatic speech recognition system specifically designed for te reo Māori. The system achieves 92% transcription accuracy, rapidly digitising oral traditions and preserving elder knowledge that might otherwise disappear with their last speakers. This groundbreaking Māori-led initiative demonstrates what happens when technology is shaped by those who hold the knowledge, rather than imposed from outside.
Dartmouth University researchers have built NüshuRescue, an AI-driven framework for preserving Nüshu – a rare script historically used by women in China. Using minimal data of as few as 35 sentence-pairs in Chinese and Nüshu, they trained a large language model to expand the database of text in the endangered script. As Professor Soroush Vosoughi notes, ‘Our work demonstrates that generative AI and large language models significantly lower barriers to revitalising endangered languages, rapidly producing valuable linguistic resources even from minimal data’.
Across Africa, the Masakhane project strengthens natural language processing for over 35 African languages, many of which are endangered or under-resourced. Their community-led approach demonstrates how collaborative efforts can amplify preservation across entire regions. Similar initiatives are emerging across the Pacific, Australia and Indigenous communities in the Americas.
These are not marginal experiments but deployed systems serving real communities. Multilingual models and community-owned datasets can rapidly revive non-English knowledge when given focused attention and resources. The key requirement is that communities control how their knowledge is digitised and deployed.
Beyond Reproduction: AI as a Generator of New, Diverse Knowledge
Synthetic data addresses statistical imbalances
The critique that GenAI merely reproduces existing statistical prevalence misses AI’s evolution into a powerful knowledge generator. Synthetic data – information algorithmically created to model realities and fill gaps – represents a direct counter-force to knowledge homogenisation. AI developers can deliberately create synthetic datasets for low-resource languages, traditional building codes or under-documented agricultural techniques, effectively manufacturing the training data that was missing.
Recent breakthroughs demonstrate this approach. Microsoft’s Phi-4, a 14 billion parameter model, relies heavily on high-quality synthetic data crafted using methods such as multi-agent prompting and instruction reversal. Despite its relatively modest size, Phi-4 achieves performance comparable to much larger models through targeted synthetic data that encourages systematic reasoning.
DeepSeek-R1 and similar reasoning models generate verified mathematical proofs, scientific hypotheses and logical frameworks that humans never explicitly articulated. These systems expand knowledge through novel combinations and insights. As synthetic data techniques mature, the reliance on entirely human-curated datasets will decrease, replaced by algorithmically generated content that can deliberately address gaps and biases.
This capability depends on intentional design choices. Synthetic data generation can either reinforce existing biases or actively counterbalance them – the outcome depends on developer priorities and community involvement in defining what knowledge matters. The point is that it is not necessarily “mush”.
Crossing disciplinary boundaries
AI crosses disciplinary boundaries that inherently constrain human research. Human knowledge acquisition remains siloed – separating hydrology from sociology from Indigenous languages from urban planning. Individual researchers rarely possess expertise across such disparate domains, and institutional structures reinforce these divisions through department boundaries and specialised journals.
Modern AI systems face no such constraint. They can link traditional Neeruganti water management practices with modern micro-climate sensor data and contemporary urban planning models, enabling novel research beyond human cognitive bandwidth.
Research published in 2025 on precision agriculture demonstrates this potential. AI systems integrating IoT sensors with machine learning algorithms achieved over 30% water savings by synthesising real-time soil moisture data, weather forecasts, crop physiology models and historical yield patterns – connections that would require teams of specialists working across multiple disciplines. The system operates continuously, learning from outcomes and adapting strategies in ways that traditional siloed research cannot match.
Environmental monitoring provides another example. AI-driven systems now combine water quality sensors, climate data, satellite imagery, social media sentiment analysis and public health records to identify pollution sources and predict environmental risks. This integration of diverse data streams – spanning natural sciences, social sciences and real-time community reporting – represents a form of knowledge synthesis that was previously impossible.
When quantitative sensor data is combined with traditional farming wisdom encoded in local languages, the result is hybrid knowledge systems richer than either source alone. Farmers in Karnataka who blend inherited wisdom about local soil conditions and seasonal patterns with AI-powered recommendations for irrigation timing and pest management achieve higher yields with lower resource consumption. This is augmentation – the enhancement of human knowledge through technological tools that can process information at scales beyond individual human capacity.
The Global Market: Economic Forces Driving Diversity and Inclusion
Market saturation drives linguistic inclusion
The argument that commercial pressures will reinforce Western bias overlooks a fundamental economic reality: market saturation. Whilst early GenAI models targeted the most lucrative English-speaking professional markets, the next frontier for growth lies in emerging economies with rapidly expanding digital populations.
The mathematics are straightforward. By 2025, it is estimated that 750 million apps will be using large language models, with 50% of digital work automated through AI technologies. The English-speaking professional market – whilst valuable – represents a fraction of potential global users. The next billion AI users will come from India, Southeast Asia and Africa, regions characterised by linguistic diversity and strong cultural traditions.
To penetrate these vast markets, AI companies face an overwhelming commercial imperative for linguistic and cultural inclusion. A model that cannot accurately advise a farmer in Karnataka on local crop rotation in Kannada will fail to capture that user. An AI that cannot assist an artisan in Lagos with market analysis in Yoruba loses that customer to a competitor who can. A system that misunderstands cultural context when providing healthcare advice in Vietnam or Indonesia risks not only poor user experience but potential harm and regulatory sanction.
This economic driver forces a necessary pivot. To succeed globally, AI must become culturally resonant and multilingual. The quest for billions of new customers will prove more powerful for addressing historical biases in data pipelines than any ethical mandate alone.
Evidence of the market shift
The market response is already visible. Switzerland’s Apertus model, developed by EPFL, ETH Zurich and the Swiss National Supercomputing Centre, was trained on 15 trillion tokens across more than 1,000 languages, with 40% of the data being non-English. This includes many languages previously underrepresented in LLMs, such as Swiss German, Romansh and numerous others. As Imanol Schlag, the project’s technical lead, states: ‘Apertus is built for the public good. It stands among the few fully open LLMs at this scale and is the first of its kind to embody multilingualism, transparency and compliance as foundational design principles’.
Meta’s ‘No Language Left Behind’ initiative produced NLLB-200, a model designed specifically for translating low-resource languages. Supporting over 200 languages – including many Indigenous and underrepresented ones – the system addresses a critical gap in global communication. This was not primarily a philanthropic venture but a strategic investment to ensure Meta’s platforms remain relevant in diverse linguistic markets. It may also result in enhanced logic and thinking - the benefits to the human mind of multilingual thinking are well documented, so why might it not be beneficial an AI, too?
The global large language model industry is projected to grow from $8.07 billion in 2025 to $84.25 billion by 2033, with a compound annual growth rate of 34.07%. Companies positioning themselves to serve only English-speaking markets will capture a shrinking share of this growth. Those investing in multilingual capabilities, cultural adaptation and local knowledge integration will dominate emerging markets. And with simultaneous translation, the need to learn English will diminish.
Studies published in 2025 demonstrate that AI translation and multilingual capabilities are improving rapidly, with particular advances in neural machine translation for less common languages and dialects. The improvement is making a dramatic difference in healthcare, customer service and legal contexts – sectors where precision matters and cultural nuance affects outcomes.
Community ownership aligns incentives
Crucially, economic incentives align with ethical imperatives when communities maintain control over how their knowledge is digitised and deployed. Te Hiku Media’s success demonstrates this principle. As a Māori-led initiative, it ensures that te reo Māori language data remains under community control, respecting cultural protocols around knowledge sharing and ensuring that AI tools support rather than supplant traditional teaching methods.
Peter Lucas Jones, the founder named by Time Magazine as one of the most influential people in AI in 2024, states: ‘I’m fighting to ensure that the Indigenous languages of the Pacific are not counted amongst those that will become extinct’. His team travels around New Zealand capturing people talking in te reo Māori, which is then added to Papa Reo to power Kaituhi, an automatic transcription tool. The focus is on accuracy and precision, ensuring the system serves the community’s needs rather than external commercial interests.
This model – community ownership combined with commercial viability – represents a path forward. When those who hold knowledge also control the technology that preserves and disseminates it, AI becomes a tool for cultural sovereignty rather than appropriation. The economic incentive ensures continued investment and development, whilst community governance ensures cultural integrity and appropriate use.
Taken together, these forces show that AI’s trajectory is not inevitably towards collapse but potentially towards restoration – provided we make deliberate choices about governance, funding priorities and community control. The technology itself is neutral; the outcomes depend on human decisions about how to deploy it.
Addressing the Mechanisms of Bias
Mode amplification can be counteracted
In his article Dennison correctly identifies ‘mode amplification’ as a legitimate concern – the tendency of LLMs to over-produce the most frequent responses in their training data. If training data includes 60% references to pizza, 30% to pasta and 10% to biryani as favourite foods, the model might generate pizza more than 60% of the time, whilst under-representing or omitting biryani entirely.
This occurs because LLMs are optimised to predict the most probable next token, leading to disproportionate emphasis on high-likelihood responses. When combined with reinforcement learning from human feedback- which embeds the values and world-views of those providing the feedback - the skew can become more pronounced.
However, this technical challenge is neither insurmountable nor inevitable. Several approaches are already mitigating these effects. First, synthetic data generation can deliberately counter-balance statistical imbalances. If biryani appears in only 10% of natural training data, developers can generate synthetic examples that increase its representation to 30% or more.
Second, advanced training techniques such as those employed in Apertus and similar multilingual models specifically target balanced representation across languages and cultural contexts. By structuring the training process to ensure minimum thresholds of exposure to diverse content, these models avoid the extreme amplification effects that plague earlier systems.
Third, the integration of diverse feedback sources during reinforcement learning can counter-act the bias introduced by homogeneous human feedback. Projects that incorporate feedback from multilingual communities, diverse cultural backgrounds and varied socio-economic contexts produce models with more balanced outputs.
In addition to these mitigation strategies, a recent study from Anthropic (in collaboration with the UK AI Security Institute and the Alan Turing Institute) uncovered a related and more troubling vulnerability: as few as 250 malicious documents injected into a training corpus can produce a “back-door” vulnerability in models ranging from 600 million to 13 billion parameters, regardless of model size or the volume of clean data. This finding reframes the risk: it is not simply about balancing representation, but also about ensuring provenance and integrity of training data. If a tiny number of skewed or malicious examples can pivot a model’s behaviour, then any synthetic-data strategy must be accompanied by rigorous filtering, provenance tracking and adversarial resilience.
The English-centric processing limitation
Research from the University of Oxford and Google DeepMind suggests that multilingual LLMs may process non-English languages through an internal representation space closest to English before translating into the target language. This finding challenges assumptions that LLMs operate in a truly language-agnostic way.
However, research on Claude 3.5 Haiku demonstrates that newer, more capable models exhibit a higher degree of generalisation and use genuinely multilingual features, especially in middle layers of the network. The model shows language-independent circuits becoming more prominent as AI capabilities advance, suggesting that the English-centric processing observed in earlier models may diminish as the field progresses.
More importantly, the direction of travel is clear. Researchers are exploring new architectures to decentralise language representations, developing methods for cross-lingual steering that improve performance in non-English languages, and training on more diverse multilingual datasets to promote balanced representations. The identification of the bias is the first step towards eliminating it.
Conclusion: Restoration, Not Collapse
The article’s warning serves as a vital ethical compass, reminding us that technology does not develop in a vacuum but reflects and amplifies existing power structures. The concerns about data asymmetry, linguistic marginalisation and the dominance of Western ways of understanding the world are entirely valid and demand continued scrutiny.
However, by focusing exclusively on AI’s capacity to reproduce existing biases, the article overlooks three powerful counter-forces already reshaping the landscape. AI’s ability to process non-traditional data streams – from IoT sensors to social media sentiment to embodied practices captured in video – offers unprecedented tools to curate and validate knowledge that human institutions marginalised. AI’s capacity to generate novel synthetic data and synthesise insights across disciplinary boundaries enables the creation of knowledge that did not previously exist in any cultural tradition. The economic imperative to serve the next billion users in linguistically diverse markets creates commercial pressure for cultural inclusion that exceeds the force of any ethical mandate alone.
The flattening of knowledge was a human achievement, carried out over centuries through colonialism, institutional standardisation and the deliberate privileging of certain ways of knowing over others. The digital gaps that concern the Guardian’s author are digital shadows of this older, deeper loss.
With intentional design, community ownership and the inherent power of its generative and analytical capabilities, AI may yet become a powerful tool for knowledge restoration and resilience. The choice before us is not between human wisdom and artificial intelligence – it is whether we use AI to recover knowledge our institutions already marginalised, or whether we allow centuries of human understanding to disappear through neglect masquerading as technological scepticism.
Traditional knowledge holders should lead AI preservation projects, not watch from the sidelines. When they do – as demonstrated by Te Hiku Media, Masakhane and similar initiatives – technology becomes a tool for cultural sovereignty rather than appropriation. When economic incentives align with cultural preservation through market expansion, commercial viability reinforces ethical imperatives rather than undermining them.
AI will not dumb down human knowledge. With appropriate governance, community control and intentional design, it will extend our understanding far beyond today’s limitations – deriving new insights from the Internet of Things, from our electronic discourse, from business data and academic research, identifying correlations in datasets that humans simply cannot access or conceive of being related.
The real risk is not knowledge collapse through AI but knowledge abandonment through inaction. The opportunity is to use AI to widen our understanding, to restore what was lost and to create knowledge that transcends the limitations of both traditional wisdom and modern science alone.
That opportunity requires us to act with urgency and intention – but it remains open, waiting for us to grasp it.
References and Sources
Primary Source Article: Dennison, D. V. (2025). ‘What AI Doesn’t Know: We Could Be Creating a Global ‘Knowledge Collapse’‘. The Guardian, 18 November 2025. https://www.theguardian.com/news/2025/nov/18/what-ai-doesnt-know-global-knowledge-collapse - this is the article being refuted. Provides the core statistics on language representation in training data, explains mode amplification concerns and introduces the knowledge collapse thesis.
Historical and Theoretical Context: Smith, L. T. (1999). Decolonizing Methodologies: Research and Indigenous Peoples. Zed Books. Foundational text explaining how colonialism disrupted Indigenous knowledge systems. Establishes the historical precedent for knowledge marginalisation that predates AI.
IoT and Environmental Monitoring: Giacobbe, M., Tricomi, G., & Puliafito, A. (2025). ‘From Sensors to Data Intelligence: Leveraging IoT, Cloud, and Edge Computing with AI’. Sensors, 25(6), 1763. Demonstrates how AI integrates IoT sensor data for environmental monitoring. Supports argument about AI processing non-textual data streams.
MDPI (2025). ‘Integrating Artificial Intelligence Agents with the Internet of Things for Enhanced Environmental Monitoring: Applications in Water Quality and Climate Data’. Electronics, 14(4), 696. Shows real-world applications of AI-IoT integration for water quality and climate monitoring. Evidence for cross-discipline knowledge synthesis.
PMC (2023). ‘Integration of IoT-Enabled Technologies and Artificial Intelligence for Smart City Scenario: Recent Advancements and Future Trends’. Provides statistics on IoT device proliferation. Documents AI’s role in smart city applications.
Indigenous Language Preservation: World Economic Forum (2025). ‘Why Generative AI Needs to Be Trained on More Languages’. Documents Te Hiku Media’s Papa Reo achieving 92% accuracy for te reo Māori. Profiles Peter Lucas Jones and the language preservation movement.
Dartmouth University (2025). ‘Language Preservation Efforts Get an AI Boost’. Details the NüshuRescue project using just 35 sentence pairs. Includes Professor Vosoughi’s quote on lowering barriers to language revitalisation.
Smithsonian Magazine (2025). ‘Can A.I. Help Revitalize Indigenous Languages?’ Broader overview of Indigenous language preservation initiatives across multiple communities. Discusses ethical considerations and community control.
Gummerus, C. (2025). ‘How AI is Preserving Endangered Languages and Cultural Heritage in 2025’. Comprehensive survey of language preservation projects including Masakhane in Africa. Documents community-led approaches.
Multilingual AI Development: ETH Zurich (2025). ‘Apertus: A Fully Open, Transparent, Multilingual Language Model’. Documents Switzerland’s Apertus model trained on 1,000+ languages with 40% non-English content. Includes Imanol Schlag’s quote on public good principles.
Polilingua (2025). ‘The Best LLMs for AI Translation in 2025’. Details Meta’s NLLB-200 model supporting 200+ languages. Provides overview of multilingual AI landscape.
Azumo (2025). ‘Top 10 Multilingual LLMs for 2025: AI Tools for Global-Scale Language Tasks’. Industry analysis showing LLM market growth from $8.07B to $84.25B by 2033. Documents commercial drivers for multilingual AI. Statistics on 750 million apps using LLMs.
Cell Press (2025). ‘A Survey of Multilingual Large Language Models: Patterns’. Patterns. Academic survey of multilingual AI development. Discusses alignment strategies and performance improvements across languages.
EzDubs (2025). ‘2025 AI Trends For Live Language Translation’. Documents rapid improvements in AI translation for less common languages. Evidence of market-driven multilingual development.
AI Technical Capabilities: Dataconomy (2025). ‘Do Multilingual AI Models Think In English?’ Oxford/Google DeepMind research on English-centric processing in multilingual models. Discusses the limitation and potential solutions.
Anthropic (2025). ‘On the Biology of a Large Language Model’. Research on Claude 3.5 Haiku showing genuinely multilingual features. Evidence that newer models are becoming more language-agnostic.
Synthetic Data: MarkTechPost (2024). ‘Microsoft AI Introduces Phi-4: A New 14 Billion Parameter Small Language Model Specializing in Complex Reasoning’. Documents Microsoft’s Phi-4 using synthetic data for reasoning. Shows how targeted synthetic data enables smaller models to match larger ones.
Prime Intellect (2025). ‘SYNTHETIC-1: Scaling Distributed Synthetic Data Generation for Verified Reasoning’. Details DeepSeek-R1 and synthetic data generation for verified reasoning. Evidence that AI generates novel knowledge rather than just reproducing existing content.
IBM (2025). ‘Examining Synthetic Data: The Promise, Risks and Realities’. Industry perspective on synthetic data applications. Discusses how synthetic data addresses underrepresentation and bias.
Netguru (2025). ‘Synthetic Data: Revolutionizing Modern AI Development in 2025’. Comprehensive overview of synthetic data generation techniques. Explains how it counteracts statistical imbalances in training data.
Souly, A., Rando, J., Chapman, E., Davies, X., Hasircioglu, B., Shereen, E., Mougan, C., Mavroudis, V., Jones, E., Hicks, C., Carlini, N., Gal, Y., & Kirk, R. (2025). ‘Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples’. arXiv pre-print, https://arxiv.org/abs/2510.07192. Demonstrates that as few as ≈250 malicious documents can back-door large language models irrespective of size, reinforcing the argument about vulnerabilities in training-data provenance and the need for balanced synthetic data strategies.
Precision Agriculture: Springer (2025). ‘Integrating IoT Sensors and Machine Learning for Sustainable Precision Agroecology’. Discover Agriculture. Research showing 30% water savings through AI-IoT integration in agriculture. Evidence of cross-discipline knowledge synthesis producing superior outcomes.
Ethical Considerations: USC Viterbi (2025). ‘Preserving the Past: AI in Indigenous Language Preservation’. Viterbi Conversations in Ethics. Discusses ethical considerations in AI-powered Indigenous language preservation. Emphasises community ownership and cultural sovereignty.