Finding the balance between innovation and societal impact

Artificial intelligence has become, in just a few years, the fastest-spreading technology in modern history: it took electricity and then the telephone a quarter of a century to reach the majority of households, only a few years for the smartphone, and already more than a billion people interact daily with tools powered by AI models, often without even realizing it. Heat maps of global adoption reveal an unprecedented concentration: small, hyper-connected states in the Gulf and Southeast Asia already boast usage rates close to 60%, while much of sub-Saharan Africa remains below 10%. This enormous gap serves as a reminder that AI is not simply a software layer – it is a socio-technical phenomenon intimately linked to the level of economic development, the state of infrastructure, and the ability of populations to harness these new drivers of productivity.

Three parallel dynamics structure the spread of this technology: first, the technological frontier, where laboratories designing models innovate and experiment; second, the infrastructure, which provides the essential computing power, electricity, and connectivity; and finally, adoption, which is measured by concrete uses in businesses, government agencies, and everyday life. Each of these components influences the other two: without data centers and networks, there is no accessible AI; without users and use cases, the frontier would lack the economic incentive to advance.

Detailed mapping shows that national wealth remains, for now, the best predictor of adoption: where GDP per capita exceeds $25,000, the share of the workforce using at least one AI service quickly climbs above 30%; below this level, the curve often stagnates. However, economic history proves that these trajectories are not fixed. The emblematic example of East Asia illustrates how a country can catch up with, or even surpass, its peers: massive and consistent investments—in education, infrastructure, and trade liberalization—were enough to transform an industrial lag into technological leadership. The moral is clear: democratic access to general advancements, more than pure invention, determines the distribution of gains.

Delving into the "lower layers" of the ecosystem reveals a pyramid of prerequisites. It all starts with electricity: without reliable power, no digital activity can survive. Yet nearly 700 million people worldwide still lack a home connection, with a significant concentration in Central and West Africa. Next comes the internet: in several countries where average bandwidth doesn't exceed basic broadband, the deployment of heavyweight, multimodal models remains a luxury. The physical proximity of data centers constitutes a third, underlying factor: by reducing the network distance between servers and users, latency is lowered, operating costs are reduced, and the path is paved for real-time applications such as code generation or live speech synthesis. Today, 80% of high-performance computing hosting regions are located north of the Tropic of Cancer; this inevitably shapes the geography of innovation.

In addition to this physical foundation, digital skills are essential. Proficiency in office software, information retrieval, and basic digital hygiene—all indispensable prerequisites—remain unevenly distributed: international assessments show that globally, fewer than half of adults know how to complete an online form. For AI, these fundamentals are crucial but insufficient: it is also necessary to understand probabilistic logic, potential biases, and the difference between inference and machine learning. Initial and continuing training programs that incorporate these concepts are already seeing measurable productivity gains, demonstrating that human capital development accelerates the return on investment in hardware.

A more subtle, but potentially decisive, obstacle lies in the linguistic dimension. Major language models, which one might assume to be omniscient, are primarily trained on English, which represents only about 15% of online speakers and barely 5% of the world's population. Of the 7,000 languages ​​identified, coverage remains minimal; some, spoken by tens of millions of people, lack any structured digital resources. This asymmetry reduces the relevance of automated assistants for public services, agriculture, or healthcare in vast areas of the globe, increasing the risk of a cognitive divide. Supporting the creation of corpora, translating interfaces, and funding multilingual open-source projects are becoming strategic levers for more equitable dissemination.

On the technological frontier, the race for performance has intensified: in less than eighteen months, the cost of inference per thousand tokens has been divided by ten, while standardized assessments show a leap in logical reasoning, instruction following, and information retrieval. This acceleration relies heavily on the ecosystem of "builders": large cloud companies, specialized startups, university laboratories, and open-source communities. Their attractiveness is measured by the growing number of published models, but also by the volume of capital and energy mobilized: training a next-generation model can consume several gigawatt-hours, making the convergence between innovation and the energy transition crucial.

Meanwhile, infrastructure builders—low-carbon energy providers, semiconductor manufacturers, and network operators—are shaping the supply curve. Analysis of installed capacity indicates that three countries account for more than half of the megawatts available in high-density data centers; this concentration could become a source of geopolitical vulnerability if supply chains become strained. Diversifying locations, encouraging regional collaboration, and integrating renewable energy sources are therefore key strategies for securing growth while controlling the carbon footprint.

When adoption indicators are combined with the dynamics of professional use, one conclusion is clear: service-intensive sectors—finance, technology, and communications—capitalize on efficiency gains earlier. Companies that integrate writing assistants or chatbots for internal support see a 30% to 40% reduction in time spent on repetitive tasks and a significant increase in employee satisfaction. In industry, predictive diagnostics and energy optimization of production lines represent the first breakthroughs, but their widespread adoption will depend on the availability of compatible connected devices and teams capable of leveraging the data flow.

Because AI is a general-purpose technology, its macroeconomic value stems from its horizontal diffusion far more than from the achievements of basic research alone. Aggregate productivity gains project a contribution of up to one percentage point of annual GDP in the most advanced economies, but only a tenth of that value where adoption remains embryonic. Closing this gap is therefore as much a political as a scientific challenge. Governments have four main levers at their disposal: guaranteeing universal access to quality connectivity; promoting lifelong learning; supporting the emergence of a local ecosystem of data centers and cloud services; and finally, creating a clear regulatory framework that protects fundamental rights without stifling experimentation.

The issue of public trust encapsulates the tension between innovation and regulation. Legal frameworks emerging around the world are seeking a balance between transparency requirements, data traceability, and the imperatives of competitiveness. A risk-based, proportionate, and flexible approach appears best suited to supporting the rapid evolution of these models' capabilities. Regulatory "sandbox" mechanisms, already proven in fintech, could offer entrepreneurs a structured testing ground, accelerating the market launch of useful solutions while ensuring that ethical safeguards keep pace.

For companies aspiring to leverage AI, the roadmap comprises three milestones. The first milestone is a process audit to identify information bottlenecks and low-value tasks that can be automated. The second milestone is data governance, which must integrate quality, security, and transparency, especially when data fuels continuous learning. The third milestone is upskilling business teams, a prerequisite for moving from gimmicky tools to truly differentiating capabilities.

The societal impact cannot be underestimated. AI can catalyze a general rise in living standards, but it can also exacerbate inequalities if access is concentrated in a few hubs. Investing in linguistic inclusion, supporting research on less energy-intensive models, and encouraging international cooperation around open standards are the cornerstones of a responsible dissemination strategy. Furthermore, experience shows that the more governance involves citizens, industry, and academia, the stronger the social mandate innovation benefits from, reducing friction associated with change.

Ultimately, the promise of AI is neither limited to the number of parameters in a model nor to the dizzying curve of benchmarks. It lies in the ability of nations, organizations, and individuals to embrace this technology, to put it to use in service of their priorities—health, education, the fight against climate change—and to share its benefits equitably. Previous industrial revolutions have shown that an invention's impact is only as significant as its widespread adoption. We are collectively at a crossroads: choosing to broaden the base of AI beneficiaries means investing in the physical and human infrastructure that makes it possible, while ensuring that the voices of all cultures resonate in the models and applications that will shape our digital future.

To learn more: discover why The effectiveness of LLM masks a risk of cognitive driftWhat are the economic effects of generative AI and how the rapid rise of artificial intelligence redesigns the world.