#20 Society and Technology Bulletin

One of the fundamental characteristics of scientific research is that no study exists in isolation; rather, each contributes to a larger body of knowledge. Academic literature functions like a structure built upon individual studies; each new article draws upon, evaluates, critiques, and advances previous research through its own contribution. This process operates through citations. Researchers reference each other’s work to establish a shared discourse and to properly attribute credit for original contributions. Thus, the connection between a work and its author gains meaning within the balance of collective knowledge production.

However, the landscape emerging in recent years—particularly in artificial intelligence research—is challenging the boundaries of this traditional cycle. Models developed by major technology companies require interdisciplinary collaboration on a massive scale. Machine learning researchers, data engineers, hardware designers, ethics experts, and product managers now work together on the same projects. The resulting scientific outputs are published with author lists far longer than conventional norms.

Highly collaborative papers with large author lists are commonplace in the fields of positive science and engineering. Yet, a particularly striking example of this trend emerged recently with Google’s paper on Gemini 2.5. In the article The inclusion of exactly 3,295 authors generated astonishment both within the scientific community and among technology circles. The joke circulating online—“How many Google researchers does it take to change a light bulb?”—has brought to the forefront new questions about the scale of production processes in AI research and how scientific contribution should be defined.

On one hand, the joke invites scrutiny of the functional roles of so many researchers involved in the study. It remains unclear to what extent each of the 3,295 individuals directly contributed to the generation of scientific content. Yet, such an extensive author list suggests that modern AI projects rely on an unprecedentedly layered division of labor. In these projects, researchers are not only those working on model training; teams involved in data collection, data cleaning, infrastructure building, ethical assessments, security testing, and productization are also integral to the process.

Machine learning researcher David Ha revealed a hidden surprise in a post on X (formerly Twitter) in the submission: When the first letters of the authors’ names are read in sequence, they spell out the sentence: “GEMINI MODELS CAN THINK AND GET BACK TO YOU IN A FLASH.”

The paper describes the Gemini 2.5 Pro and Gemini 2.5 Flash models, introduced in March. These models employ a novel reasoning technique that generates text in a “think aloud” style before producing answers, aiming to solve more complex problems. Google’s first Gemini paper in 2023 listed only 1,350 authors; within two years, this number has increased by 144 percent.

Does such an expansive author list truly reflect active participation by every individual in every stage of the research, or is it simply the result of corporate practice—listing everyone who had any involvement in the project?

The answer to this question is not merely an ethical concern; it also shapes how knowledge enters circulation. A paper with thousands of names creates the impression of meticulous, consensus-driven production, yet in reality, it obscures the ability to distinguish individual contributions. This situation blurs recognition of scientific labor and inflates citation counts, misleading metrics of scholarly impact.

Moreover, from the perspective of scholarly integrity, it becomes unclear which findings are attributable to which researchers. Consequently, while massive teams accelerate scientific production in AI research, they also introduce new questions demanding a redefinition of academic norms. It appears that AI research is fundamentally altering the nature of research itself.

In accordance with President Trump’s executive order issued in January titled “Removing Barriers to American Leadership in Artificial Intelligence,” the White House released a comprehensive strategy document titled “Winning the AI Race: America’s AI Action Plan.” announced This plan extends beyond limiting the American AI ecosystem to technological development; it presents a global vision encompassing geopolitical, economic, and cultural dimensions.

The document groups over 90 policy actions to be implemented over the coming weeks and months under three main pillars:

• Accelerating Innovation: Removing regulatory barriers and accelerating private sector R&D activities.

• Building AI Infrastructure: Streamlining permitting processes for data centers and semiconductor factories, expanding vocational training programs for skilled labor.

• Leadership in International Diplomacy and Security: Exporting AI standards and forming security-based technology partnerships with allied and friendly nations.

One of the plan’s most notable features is the design of AI exports as “end-to-end packages.” The combined export of hardware, software, models, and standards aims to integrate U.S. technology more tightly into global networks.

One of the plan’s striking elements is its emphasis on preserving freedom of expression in “frontier” models. The federal government’s announcement that it will only contract with large language models “free from objective and ideological bias” marks a turning point in political debates over content generation and oversight in AI systems. This approach reopens questions about the very concept of neutrality by highlighting both technical concerns in AI safety and the political implications of value conflicts embedded in technology design.

The question of whether a model truly free of ideological bias is even possible will resurface. When we recognize that neutrality is impossible on the linguistic level, the current crisis becomes even more apparent.

This action plan also reaffirms the U.S. position of treating AI as a central element of its national security strategy and global influence policy. Thus, technological advancement is no longer merely a requirement for corporate growth agendas but has become a vehicle for national interests. It particularly embeds the political hegemonic effects of American technology exports. Therefore, the AI initiative can be viewed as a whole as representing America’s national policy interests and as an enhancement of soft power capacity against primary technological rivals such as China.

On the other hand, filtering AI projects based on the criterion of “freedom of expression” may ignite debate over who sets the rules between technology companies and federal governance. Since this implies direct state intervention in the design of technological products, it raises contentious issues not only regarding content moderation but also concerning market competition and innovation. These developments signify federal government interference in independent technology companies.

Finally, the over 90 policy actions in the plan indicate not only a comprehensive bureaucratic mobilization but also that the U.S. is increasingly framing the AI race as a “national security competition.” This perspective strongly suggests that American AI policy in the coming period will be shaped along trade, diplomacy, and defense axes. Thus, the prevailing naive notion that AI is a universal technological advancement is losing ground.

The International Mathematical Olympiad (IMO), the world’s most prestigious competition for young mathematicians, has been held annually since 1959. National teams composed of the top six pre-university students compete by solving six challenging problems covering algebra, combinatorics, geometry, and number theory. Approximately 8 percent of participants receive gold medals. The difficulty level and associated prestige are exceptionally high.

In recent years, the IMO has become not only a test for young mathematicians but also a benchmark for evaluating the advanced reasoning and problem-solving capabilities of AI systems. Last year, Google DeepMind’s AlphaProof and AlphaGeometry 2 systems solved four out of six problems, achieving the silver medal standard.

This year, a higher threshold was surpassed. The advanced version of Gemini Deep Think solved five out of six problems flawlessly solved, earning 35 points and reaching the gold medal level. IMO President Professor Gregor Dolinar described Gemini’s solutions as “clear, precise, and often easy to follow.” Thus, solving sophisticated mathematical problems at the Olympiad level has now become part of the AI agenda.

Last year’s systems translated problems into specialized mathematical languages (such as Lean) before solving them, a process that took days due to technical steps like transforming all variables—essentially handling the task in a computer language. Gemini Deep Think, however, demonstrated the ability to work directly in natural language, starting from official Olympiad problem statements and generating formal mathematical proofs. Moreover, it accomplished all of this within the Olympiad’s 4.5-hour time limit.

This development demonstrates that AI models are increasingly capable of handling more complex tasks. However, interpreting such achievements directly as “AI now understands mathematics” is problematic. There is a significant distinction between solving mathematical problems and understanding mathematics. A machine’s ability to solve these problems does not imply it grasps the conceptual unity of mathematics. Such a claim risks reducing mathematical thought to mere mechanical processes.

This reduction may strip mathematics of its qualitative richness by redefining it as a sequence of mechanical operations, erasing its historical roots in human perception of natural forms and a priori logical principles. When the reduced version replaces the original and becomes the new norm, the intrinsic richness of mathematics becomes invisible.

Moreover, inferring from AI’s functional outputs in mathematics or language that it possesses genuine comprehension constitutes another illusion. The outputs of summarizing or inferential language models often mask the fact that they do not truly understand but instead skillfully generate statistical correlations. This creates the danger of shaping our perceptions of AI based not on reality itself but on these rapid and fluid outputs.