Knowledge is central to the AI era because AI systems both depend on knowledge and reshape how knowledge is created, stored, and used. Knowledge is not a homogeneous resource; it includes memory of facts, the ability to relate information, creativity, and social skills that enable collaboration and sustainable relationships [1]. In the AI era, these dimensions become even more important because AI operates primarily on data, but meaningful outcomes require interpretation, context, and human insight.

Here are a few formats in which knowledge appears:

Knowledge differs from data and information. Data is a record of events, while information adds meaning to data. Knowledge emerges when information is interpreted through experience, values, and contextual understanding [2]. AI systems process data at massive scale, but the transformation of data into knowledge remains a human-driven process that involves framing, judgment, and application [2]. This distinction is crucial: AI can assist in aggregating and validating information, but the productive use of knowledge depends on human understanding and institutional context [3].

In contemporary organizations and societies, knowledge exists in explicit and tacit forms. Explicit knowledge can be codified in documents, repositories, and systems. Tacit knowledge resides in individuals’ experience and intuition and is difficult to formalize [4]. AI can help manage and retrieve explicit knowledge efficiently, but tacit knowledge—contextual judgment, improvisation, relational awareness—remains deeply human and essential for responsible decision-making [4]. The AI era therefore heightens the need to recognize and manage both types of knowledge effectively.

Knowledge is also a renewable resource. Unlike physical resources, knowledge grows when shared and used [3]. AI amplifies this dynamic by accelerating access to information and enabling new forms of analysis. However, without proper knowledge management—clear objectives, validation mechanisms, and structured sharing processes—AI can magnify confusion rather than clarity [3]. Institutions must therefore develop deliberate strategies to identify what knowledge they need, how to share it, and how to apply it responsibly [3].

Furthermore, knowledge defines identity. It shapes how individuals and institutions understand themselves and others. In the AI era, large-scale codification of knowledge through digital platforms raises concerns about concentration and asymmetry of control [5]. When knowledge becomes embedded in a few dominant systems, the risk is not merely technical but structural: it affects autonomy, creativity, and equitable participation in global processes [5].

Finally, knowledge strengthens critical awareness. Data analysis and AI tools can challenge dominant narratives and misperceptions when integrated thoughtfully into policymaking and institutional practice [1]. But this requires awareness of what data can and cannot do, and recognition of its limitations [1]. In this sense, the AI era demands not less knowledge, but deeper and more reflexive engagement with it.

In sum, knowledge is relevant in the AI era because it is both the raw material AI consumes and the human capacity that gives AI meaning, direction, and ethical grounding. Without careful management of explicit and tacit knowledge, AI risks becoming a powerful but unanchored tool. With thoughtful stewardship, it can become a catalyst for collective learning and sustainable development [3][1].

The prevailing AI narrative—that increasing computational power automatically yields better intelligence—is widely accepted but increasingly contested. However, more processing power does not necessarily translate into more useful AI.  [1][2].

Diplo’s experience confirms this point. Our operational AI pipelines rely on comparatively small, in‑house language models rather than frontier‑scale systems. It is absolutely sufficient for dealing with a limited lexical set (≈2 500 words) and a modest repertoire of cognitive tools (20–30 typical logical, problem‑framing or crisis‑management approaches), these tasks can be performed efficiently with compact models [1][2].

In our workflow the most critical AI function is chunking, identifying the most relevant sentence‑level fragments of source material. Once the appropriate ‘chunks’ are retrieved, a large language model is needed only to generate fluent sentences and paragraphs. DiploAI therefore implements sentence‑level chunking internally, reserving the larger model’s generative capacity for syntactic construction [1].

Smaller models also offer practical advantages. They can be retrained quickly on domain‑specific corpora (e.g., conflict resolution, trade negotiations, climate‑policy dialogue), ensuring that the AI stays aligned with the nuanced terminology and procedural norms of each issue area. This rapid, low‑cost adaptation reduces the risk of hallucinations that are more common in oversized, less‑focused systems [2].

Overall, Diplo’s evidence supports a ‘right‑sized’ AI strategy: deploy lightweight, purpose‑built models for routine and predominant tasks, and reserve large‑scale models for other tasks relaed to general culture and more complex reasoning.  [1][2].

Sources:

The AI Paretto Paradox

Do we REALLY need AI for everyday work?

The elephant in the AI room

Business risk management – Frontier firms need to maintain investor enthusiasm (hype) while reassuring regulators and the public that they take safety seriously. A dramatic “doomsday” warning can justify continued funding, whereas a “bright‑future” story can attract new customers and talent [2][3].

Concentration of power – Because 80 % of the technology stack (data, compute, model design) is owned by a few players, they set the agenda for what is considered possible or dangerous [1].

Regulatory pressure – Governments and multilateral bodies demand risk assessments; firms respond by publicising safety‑oriented road‑maps, which in turn fuel the “risk” side of the story [4].

Media amplification – Journalists gravitate toward extreme headlines because they generate clicks; this reinforces the oscillation between catastrophe and breakthrough [6].

By combining human‑rights‑centered governance, open‑source technical capacity, data sovereignty and inclusive multi‑stakeholder oversight, societies can begin to dismantle the structural injustices that AI—if left unchecked—might amplify, while simultaneously harnessing AI’s potential to empower the most vulnerable and restore spaces for making social, economic and political choices.

he ability to make choices is a key component of democracy. Thus, the central question is whether AI is enhancing or reducing our capacity to choose. Below is a summary of the reasoning:

How AI Reduces Democratic Choices

Algorithmic gate‑keeping – AI decides who sees which news, which jobs are advertised, and which public services are offered. When the criteria are hidden, citizens cannot deliberately choose alternatives [1][2].

Synthetic manipulation – Deepfakes and AI‑generated political ads spread faster and cheaper than any human‑made counter‑measure, turning elections into “psychological experiments” rather than contests of ideas [4][5].

Micro‑targeted persuasion – Platforms use psychographic profiling to serve each voter a unique set of messages, fragmenting the public sphere and preventing a shared deliberative space [6][7].

Concentration of infrastructure – 70 % of the European cloud market is owned by three U.S. firms; only a tiny minority of global AI firms are European, meaning the technical levers of choice are controlled outside democratic oversight [2][1].

Why the Loss of Choice Endangers Democracy

Exacerbation of existing inequalities – Biases embedded in data and models reproduce social hierarchies, meaning that already marginalised groups lose the limited choices they retain [1][7].

Erosion of accountability – When no actor can be identified as the source of a manipulated message or a denied service, the principle of “someone must answer for the public good” collapses [4][6].

Weakening of pluralism – AI‑driven micro‑targeting silences minority voices and amplifies dominant or well‑funded narratives, undermining the democratic ideal of a marketplace of ideas [6][7].

Shift toward technocracy or authoritarianism – Citizens frustrated by opaque manipulation may favour “strong‑handed” leaders promising quick, algorithmic solutions, eroding the deliberative character of democracy [4][7].

Like other “third world” regions, Europe provides resources (data and knowledge) in exchange for technology from companies based mainly in the USA and, increasingly, China. As noted, “80% of the technologies needed for Europe’s digital transformation are designed and manufactured in third countries,” and “EU platforms capture only 5% of global market value” 1][2].

“Divides” dominated the first part of our session. While AI is widening gaps, their nature differs from traditional North–South divides. Increasingly, the division is between a few powerful tech companies and the rest of society. For example, in the United States, many states—both Republican and Democrat—are deeply concerned about the power of tech companies, which enjoy the support of the Trump administration. Europe, meanwhile, is the “South” in AI terms, as it must import the technology. The real divides are thus between big tech companies centralising our knowledge and the rest of us, wherever we are based. This represents a critical economic and, even more importantly, cognitive divide for humanity. This risk is highlighted, for instance, in Pope XIV’s encyclical Magnifica Humanitas.

In many ways, it feels that way. As Arthur C. Clarke famously observed, any sufficiently advanced technology is indistinguishable from magic [1]. When we watch AI systems compose poetry, summarize complex texts in seconds, generate lifelike images, or simulate human conversation, we experience a sense of wonder similar to encountering the extraordinary in literature. [2].

Yet this “magic” is not supernatural. It is statistical, computational, and data-driven. AI does not understand in the human sense; it does not feel, reflect, or grow through lived experience [3]. It imitates intelligence by processing data and adapting to patterns, but it lacks moral conscience, inner growth, and relational awareness [3]. The enchantment lies not in consciousness, but in performance. Like magical realism in Latin American literature, AI blends the extraordinary with the ordinary. It performs astonishing feats—writing, predicting, analyzing—yet it remains embedded in everyday bureaucracies, workplaces, and legal systems [4].

For colleagues from Latin America, the analogy with magical realism is particularly resonant. In the novels of Gabriel García Márquez, the miraculous appears alongside the mundane, accepted without question. Similarly, AI’s outputs—generated texts, simulated empathy, predictive insights—are increasingly accepted as normal, even when their strangeness should provoke reflection [4]. The magical becomes routine. The extraordinary becomes infrastructure.

But magical realism also reminds us that wonder does not eliminate power. In Macondo, miracles coexist with exploitation, violence, and inequality [3]. AI’s enchantment can conceal asymmetries of control. Governance discussions show that decision-making about AI is often concentrated among a limited number of actors, while many regions remain underrepresented in shaping its rules [5]. The magic may be global, but control is not evenly distributed. Equity demands that more voices meaningfully participate in AI governance conversations [6].

Moreover, like any magic, AI can be manipulated. The apparent objectivity of AI-generated results can mask cultural assumptions embedded in training data and model design [7]. When AI systems simulate empathy or offer advice, they can create the illusion of relational depth without genuine understanding [8]. This illusion can be comforting, but it risks weakening personal judgment and authentic human relationships [9]. In governance terms, there is also the danger of “grey magic,” where narratives of humanitarian benefit conceal strategies of control or market dominance [10].

AI also reshapes sensitive domains such as justice, where efficiency and automation promise relief from procedural burdens but cannot replace human moral reasoning [11]. While AI can assist in analyzing documents and identifying patterns, responsibility for ethical judgment must remain with human actors [12]. The magic must remain assisted or augmented intelligence, not automated authority [13].

Thus, AI governance becomes an act of narrative discipline. We must acknowledge the magic without being seduced by it [14]. We must resist framing speculative doomsday futures as the primary regulatory driver while neglecting immediate risks such as labor displacement, misinformation, and content manipulation [15]. Magical realism teaches us that the extraordinary does not absolve us from responsibility. The miraculous and the mundane coexist; power structures persist beneath wonder.

Ultimately, AI is magical realism in motion. It expands human capability, compresses time, and blurs boundaries between human and machine expression. But it is not mythic destiny. It is shaped by data, by design, and by governance choices. The question is not whether AI is magical, but whether we allow its magic to distort our sense of justice, dignity, and shared responsibility. The task before us is to ensure that this new form of enchantment serves humanity rather than enthralling it [16][17].

Artificial intelligence affects nature across the entire lifecycle of its infrastructure, and its impact can be understood through three interconnected domains: energy, water, and rare materials. These impacts can be situated within a broader ecological framework that includes the geosphere, the biosphere, and the noosphere.

First, AI has a significant impact on energy consumption. The training and deployment of large AI models require massive data centers, which consume large amounts of electricity and contribute to greenhouse gas emissions [1]. By 2030, AI is projected to account for 12% of total electricity demand in the United States due to scaling trends [2]. Data centers not only require electricity for computing but also for cooling, which adds to their energy footprint [3]. While AI can contribute to sustainability—for example by improving renewable energy forecasting and grid optimization [4]—its own energy demands raise concerns about climate impacts [5].

Second, AI affects water systems. Training large AI models requires substantial water use, particularly for cooling data centers. One study estimates that training AI systems can require approximately 3.5 million liters of water for cooling alone [1]. Water use is therefore not a marginal issue but a structural component of AI infrastructure. This water consumption can be particularly problematic in regions already facing water stress, thereby affecting both the geosphere (water systems) and the biosphere (ecosystems dependent on water) [1].

Third, AI depends on rare materials extracted from the Earth. Generative AI and digital infrastructure require metals and minerals such as lithium, cobalt, nickel, rare earth metals, and copper [4]. The extraction of these materials often leads to biodiversity loss, deforestation, habitat destruction, and pollution [4]. For example, cobalt mining in the Democratic Republic of Congo has been associated with environmental degradation, toxic contamination, and severe social harms [4]. In addition, the growth of AI infrastructure increases demand for critical raw materials such as neodymium, with projected storage needs in 2025 requiring up to 80 kilotons—far exceeding prior European demand levels [1]. These extraction processes directly impact the geosphere through land disturbance and mineral depletion, and the biosphere through ecosystem damage.

AI also contributes to e-waste and long-term material accumulation. By 2023, over 70 million tons of electronic waste were expected to be generated annually [1]. The increasing proliferation of AI-enabled devices compounds this waste problem, adding further strain to ecosystems and requiring improved circular economy practices [1].

However, AI’s impact on nature is not solely negative. Within the biosphere, AI is used for biodiversity monitoring and conservation. Machine learning systems can analyze camera trap data, acoustic signals, and satellite imagery to detect species, monitor habitats, and track environmental changes [5]. AI also supports environmental monitoring and predictive modeling, enabling near real-time assessment of climate, deforestation, and pollution patterns [3]. These capabilities strengthen environmental governance and support more informed policymaking [3].

This dual character reflects a broader ecological framing. The geosphere includes land, minerals, water, and energy systems affected by AI’s material footprint. The biosphere includes living systems influenced both positively (through conservation tools) and negatively (through intensified extraction and industrial use). The noosphere, the sphere of human knowledge and reflection, shapes how AI is governed and directed. As noted, the ethical question is not only what AI can do, but who controls it and for what purpose [5]. Governance frameworks and international standards aim to provide transparency and lifecycle assessment methodologies to better understand AI’s environmental consequences [3].

In summary, AI affects nature through energy consumption, water use, and mineral extraction. These impacts span the geosphere and biosphere, while the noosphere—human knowledge and governance—determines whether AI becomes a tool for ecological protection or a driver of further environmental strain.

The main obstacle to developing home‑grown AI is not the lack of technology, but a confluence of narrative, economic, regulatory and psychological factors that keep power and resources concentrated in a few frontier‑model firms. Here are a few reasons that created the current situation:

Narrative dominance – Media hype and the “AI is mysterious” story create a perception that only large corporations can safely develop and deploy AI. This perception is reinforced by the “doom‑vs‑utopia” swing that keeps investors and regulators focused on a narrow set of actors [1][2].

Financial and regulatory risk – The current investment climate rewards large‑scale, high‑visibility projects; smaller, locally‑controlled models are seen as financially unattractive and potentially non‑compliant with emerging high‑risk AI regulations [3][2].

Psychological and institutional inertia – A prevailing “expert‑gatekeeping” mindset discourages non‑technical actors from entering the AI design space, resulting in a self‑reinforcing cycle where only the established players continue to shape the narrative [1].

How to turn “talk‑about‑AI” into meaningful participation in AI governance

Create and join “community‑owned” AI projects
• Open‑source model ecosystems (Llama, Mistral, DeepSeek) provide the technical scaffolding for locally trained models that respect community data‑sovereignty [1][2].
• When a community can demonstrate a functional, low‑cost AI that solves a specific need (e.g., local health screening, agricultural advice), it gains bargaining power with larger vendors and policymakers.

Build a solid, non‑technical AI literacy base
• Understanding the basics (e.g., how floating‑point arithmetic underpins model calculations) removes the myth that AI is “mysterious” and lets participants learn about AI by developing AI (Apprenticeships).

Use a structured participation framework
• Arnstein’s ladder (non‑participation → tokenism → citizen power) has been adapted into a “pyramid of participation” that maps the steps from “integration” (learning the process) to “leadership” (co‑designing rules) [4].
• The ladder makes it clear when participation is merely symbolic and when it actually influences policy.

Join multistakeholder fora that embed feedback loops
• The Internet Governance Forum (IGF) and its regional off‑shoots (EuroDIG, YouthDIG) explicitly design spaces for informing, consulting, and co‑creating [4][5].
• Feedback loops—collecting user‑generated error reports, retraining models, and benchmarking against common standards—are the technical glue that turns citizen input into model updates [1].

Capacitate yourself and your community
• Capacity‑building programmes (technical boot‑camps, certification of AI assessors, and “AI‑literacy pilots” in schools) have already trained more than 100 assessors and are scaling to NGOs and civil‑society groups [1].
• Local workshops on data‑ownership, open‑source tooling and federated‑learning let participants experiment with “bottom‑up” AI without needing massive cloud resources [1][2].

Secure representation of under‑served groups
• A major barrier is the cultural‑ and linguistic‑divide: most large‑language models only support a few dozen languages, leaving 7 000‑plus languages invisible online [6].
• Inclusive policy proposals (e.g., UNESCO’s World Atlas of Languages, open‑source multilingual models) give concrete entry points for communities to claim space in the model‑training pipeline [6][2].

Leverage formal mechanisms for global representation
• Developing‑country delegations are encouraged to coordinate regionally (African Union, LAC networks) so their collective voice can be heard in global fora such as the AI‑Act mapping groups or the Council of Europe AI Convention [7][8].
• Funding‑mechanisms (travel grants, remote‑participation subsidies) and “meta‑participation”—creating parallel micro‑sessions when the main agenda does not reflect a group’s priorities— help overcome the cost and time barriers identified by participants from the Global South [4][7].

Translate technical insight into policy language
• Use the standards‑to‑regulation mapping (e.g., IEEE’s value‑based design standards linked to the EU AI Act) to turn a technical observation (“this model over‑fits on training data”) into a concrete policy recommendation (“require dataset‑diversity audits for high‑risk systems”) [1].
• Such mapping makes your arguments credible for legislators and regulators who need legally‑bounded language.