A student asks ChatGPT to explain a difficult scientific concept. A business executive consults it before making a decision. Elsewhere, a researcher uses the same technology to explore an unfamiliar field. The exchanges may last only seconds. Their consequences are harder to establish. Did someone actually learn something? Was a better decision made? And what becomes of the information once the conversation ends? Such questions receive considerably less attention than the impressive statistics surrounding artificial intelligence’s expansion.
In September 2025, researchers associated with OpenAI, Harvard University, and Duke University published How People Use ChatGPT, a working paper issued by the National Bureau of Economic Research (NBER). Drawing on a primary sample of approximately 1.1 million conversations, the study examined usage patterns through privacy-preserving automated classification. Its findings were revealing: non-work-related interactions represented more than 70% of consumer usage, while practical guidance, information seeking, and writing collectively accounted for nearly 80% of classified activity. ChatGPT was becoming something more pervasive than a workplace assistant. It was entering the ordinary architecture of everyday judgment.
There is considerable scientific value in documenting that transition, although the research has boundaries worth examining. Its classifications were generated through automated systems, the underlying information came from ChatGPT’s consumer environment, and several authors maintained professional affiliations with OpenAI. The study was also circulated as an NBER working paper rather than a peer-reviewed journal publication. These circumstances warrant methodological scrutiny without diminishing the importance of the evidence. What remains comparatively unexplored is what happens after users receive a response. Frequency of interaction provides only an incomplete picture of the intellectual and social consequences.
An accurate explanation can remain intellectually unused. An impressive analysis may encounter an institution unable to implement its recommendations, whether because of limited resources, organizational resistance, or competing priorities. Universities already face versions of this problem. The availability of generative tools offers new possibilities for learning and research, yet their presence in classrooms tells us little about the quality of students’ reasoning or the originality of scientific work. Sometimes the immediate achievement is simply that a task becomes easier. That matters. Whether ease translates into understanding is a different question.
Economic research adds a complication. In July 2026, Rachel Yuting Fan and Ha Minh Nguyen published Aggregate Gains from AI and Their Distribution: Global Evidence from Usage Data as an International Monetary Fund working paper. Using data from Anthropic’s AI systems, rather than ChatGPT, they estimated an annualized labor-cost equivalent of approximately $2.7 trillion in time savings across 86 countries, representing about 3.4% of those economies’ combined GDP. The estimate expresses the monetary value assigned to working time potentially saved. It should not be confused with observed increases in economic output, corporate profits, or household income.
The distribution of those estimated gains is particularly revealing. In developing economies, usage-based value was concentrated among relatively narrow groups of highly paid professionals. Some occupational structures appear better positioned than others to capture the immediate advantages of generative intelligence. Education, regulatory readiness, and access to complementary resources can influence the extent of that advantage. This leaves an uncomfortable possibility: a technology may become increasingly accessible while its most economically valuable applications remain concentrated among those who already possess substantial professional and institutional resources.
Yet the beneficiaries cannot be reduced to a handful of corporations. Developers earn revenue from AI services, semiconductor manufacturers supply essential hardware, and cloud operators provide computational infrastructure. At the same time, smaller enterprises may gain analytical capabilities that were previously beyond their means. Independent researchers can access sophisticated assistance despite limited funding. These opportunities are substantial, particularly where conventional technological resources are scarce. Their economic and social effects, however, depend on circumstances that usage figures rarely capture. Access can create opportunities without ensuring their realization.
The International Labour Organization’s June 2026 review, The Impact of GenAI on Jobs, Productivity and Work Organization, offers a useful qualification. Productivity improvements appear in the available evidence, although their magnitude varies and reported time savings have not consistently translated into higher measured output, earnings, or employment. Large-scale job displacement remains limited in the studies reviewed, while occupational opportunities for younger workers and changes in workplace autonomy deserve closer attention. A worker may accomplish more in less time without receiving a proportional economic reward. The organizational arrangements surrounding technology become important here, perhaps as important as the technology itself.
There is also a physical dimension that tends to disappear from discussions of digital intelligence. The International Energy Agency’s April 2026 report, Key Questions on Energy and AI, estimated global data-center electricity consumption at approximately 485 terawatt-hours in 2025, with a central projection of 950 terawatt-hours by 2030. These figures cover data centers generally, including activities unrelated to ChatGPT. AI-focused facilities nevertheless represent an increasingly significant source of demand. Advanced computing depends on electrical networks, semiconductors, specialized equipment, industrial supply chains, and investment decisions whose consequences extend well beyond software markets.
The same IEA analysis documents remarkable improvements in energy efficiency per AI task. That development deserves attention because it complicates simplistic assumptions about the environmental trajectory of artificial intelligence. A conventional text query can become substantially less energy-intensive even as total computational demand rises through expanding adoption and more demanding applications. Reasoning systems, autonomous agents, and video generation introduce different computational requirements. The relationship between efficiency and resource consumption is therefore unstable, influenced by technological design as well as the behavior of markets and users.
Water presents another set of questions. In October 2026, the IEA reported in Powering Artificial Intelligence that nearly 40% of identified AI-dedicated data-center capacity was located in areas experiencing high or extremely high baseline water stress. The finding does not establish that those facilities are responsible for local shortages. It does, however, complicate infrastructure planning in territories where competing demands already exist. Cooling technologies, electricity generation, local ecosystems, and water availability cannot be considered independently. Even efforts to reduce water use may increase electricity requirements, depending on the systems employed.
Who ultimately pays for this infrastructure is partly a political and regulatory question. A data-center project can generate investment, employment, and new electrical capacity. Under different arrangements, the costs of network expansion may be distributed among consumers who receive few direct benefits from the development. Public authorities must assess projected demand that may never materialize, while private investors face their own exposure to technological and financial uncertainty. There is no universal outcome. Geography, regulatory design, corporate commitments, and the allocation of infrastructure costs determine much of what eventually reaches households and communities.
The geopolitical implications are less visible in the experience of an individual user. Countries that control advanced chips, major computational platforms, and critical technological supply chains possess capabilities that others must obtain through external providers. For Mexico and other emerging economies, wider AI adoption may coexist with continued dependence on foreign infrastructure and intellectual property. That relationship deserves closer examination. Technological access can improve productivity and scientific opportunities while leaving unresolved questions about domestic innovation, strategic autonomy, and the capacity to participate in higher-value industrial activities.
Universities occupy an unusual position within this landscape. Their students and researchers may benefit enormously from AI, particularly where laboratories, funding, and scientific networks remain limited. But introducing advanced tools into academic work does little to explain whether institutions are developing stronger research cultures. Scientific independence involves questioning results, recognizing methodological limitations, testing competing explanations, and formulating problems that automated systems cannot settle on their own. An institution can become technologically sophisticated while remaining intellectually dependent. Conversely, capable researchers working with limited resources may find in AI an opportunity to overcome barriers that previously constrained their investigations.
This tension also appears in the relationship between artificial intelligence and scientific discovery. AI systems may accelerate research, support medical advances, improve industrial operations, and help optimize energy networks. Some of these applications could generate benefits substantially greater than their computational costs. Others may never progress beyond promising demonstrations. The evidence is still developing, and the criteria for comparing such outcomes are far from straightforward. An improvement in scientific productivity, for example, does not necessarily produce an immediate social benefit, although its longer-term consequences may prove considerable.
The NBER study offers an unusually detailed picture of how people interact with ChatGPT. The IMF examines the estimated economic value associated with AI usage. The ILO considers consequences for employment and work organization, while the IEA investigates the infrastructure and resources supporting the broader technological expansion. Together, these investigations illuminate different portions of a process whose overall consequences remain difficult to establish. Each contributes evidence, though no single one can account for the relationship between technological consumption, knowledge creation, economic distribution, and environmental sustainability.
Perhaps the most interesting uncertainty lies in what occurs between these dimensions. An AI-generated answer may become the starting point for meaningful scientific discovery or remain a convenient explanation that is quickly forgotten. A productivity improvement may strengthen an institution or disappear into an accounting adjustment. Infrastructure investment may stimulate development while introducing pressures elsewhere. These possibilities are not mutually exclusive, and their relative importance will vary across economies, institutions, and territories.
The ChatGPT boom is real. Its consequences are still being negotiated through decisions that often occur far from the screen on which an answer appears. Who benefits and who pays may ultimately depend less on the number of intelligent systems deployed than on the institutions, resources, and relationships through which their capabilities are put to use.
For now, the distance between technological possibility and its eventual social consequences remains an unusually consequential space for scientific investigation.
Phoenix24 Research | Independent Scientific and Geopolitical Analysis