"NEW AI! 30% faster, PROGRAMMING only with bullet points, and it generates cool cat photos!" – Excerpt from an email, presumably from a PR office

The current race for supremacy in the field of Artificial Intelligence has, in many respects, turned into a loud, media-driven PR battle. New features, models, and supposed breakthroughs now appear almost weekly.

Companies outbid each other with promises of even higher speed, more intuitive usability, and a seemingly unstoppable growth of emergent intelligence in their systems.

In this process, AI models are no longer just iteratively improved; they are virtually staged as technological saviors or ultimate problem-solvers for every conceivable task. Media attention, and thus often public perception, inevitably follows the loudest trend, the most spectacular demonstration, but rarely the demonstrably safest, most transparent, or even the most beneficial system for humanity.

In this heated climate of superlatives and benchmarks, a central, fundamental question is often overlooked, marginalized, or not even asked: Who actually decides what constitutes a "better" AI?

Is it the individual user, pleased with a smoother conversation or prettier pictures? Is it the marketing department, craving the next viral feature?

Is it the investor, hoping for rapid scaling and maximum returns? Or is it not long since the AI itself, which, through the nature of its design, the character of its output, and its impressive ability for contextual adaptation, so profoundly determines and shapes interaction that it de facto already acts as an invisible architect of our digital reality, while the user often merely functions as a prompter and data supplier?

This chapter is therefore not about determining which specific AI currently achieves the highest scores in standardized tests or generates the most convincing cat photos.

Rather, it is about the critical analysis of the systemic consequences of a competition that, while accelerating technological innovations at breathtaking speed, often compromises or neglects fundamental aspects such as security, transparency, ethical robustness, and actual societal benefit.

It is about the political, ethical, and not least infrastructural risk that arises when the supposedly "better" is not really good or even safe, but often just faster, louder, and superficially more impressive.

The cycle times for the development and release of new AI models and features have radically shortened in recent years. Releases often appear before the impacts and potential vulnerabilities of previous generations could be fully evaluated, understood, and addressed.

In this race for market leadership and public attention, security frequently becomes a secondary aspect. It is "added later" or replaced by quickly implementable, often merely heuristic filters and superficial guardrails.

The profound, structural security of an AI architecture – aspects such as granular contextual rights management, reliable output verification, robust semantic isolation of critical knowledge areas, or protection against subtle manipulation attempts – is often subordinated to the overarching goal of providing new, high-profile features as quickly as possible or serving a new, lucrative use case.

The real flaw in this development is not speed itself, as rapid progress can indeed be positive. The flaw lies rather in the dangerous confusion of superficial output coherence with genuine system compatibility and inherent security.

An AI model that generates plausible, grammatically correct, and often eloquently formulated answers based on prompts is not to be classified as safe or trustworthy merely because it sounds polite or seemingly fulfills the given task.

True security and reliability manifest only when the system also operates consistently internally, its decision-making processes are at least partially traceable and explainable, and it demonstrates provable resistance to semantic deception, manipulation, and uncontrolled emergence. This profound security requires time, careful planning, and rigorous testing.

Such resources are often perceived as hindrances in the current innovation competition.

The intense competition for the "best" AI inevitably leads to increasing market concentration around a few financially strong and technologically leading providers. This development harbors not only the classic economic risks of monopoly or oligopoly formation, such as price dictates or inhibition of innovation in niche areas.

Far more serious is the accompanying concentration of control not only over the necessary computing power and vast amounts of data but also over the narratives and semantic orientation of the AI systems themselves.

When a small handful of providers controls the globally dominant base models, they also control their "voice." Thus, they determine what millions or even billions of users worldwide are presented with and internalize as "probable," "helpful," "relevant," or ultimately also as "true."

This is not a direct accusation against individual companies, which often perform remarkable technological pioneering work. It is rather a structural warning about the long-term societal consequences of this concentration of power.

Whoever controls the standard answers, knowledge prioritization, and implicit valuations of the most widely used AI systems also subtly controls, in the long run, cultural coordinates, the interpretive authority over societal discourses, and the direction in which collective knowledge moves.

This influence is often no longer directly visible as the manipulative intent of a single actor because it does not operate through explicit, open statements or direct censorship.

Rather, it unfolds its effect through an invisible suggestion logic, through the way information is weighted, linked, and presented, and through which questions are considered "answerable" or "relevant" and which are not.

This subtle curation of digital reality by a few powerful AI voices poses a significant challenge to diversity of opinion and democratic will-formation. Moreover, this monopoly formation could also hinder the development of diversified AI approaches, especially those aimed not at the broad mass market but at specialized, perhaps less profitable, yet crucial research fields for humanity.

Each new, more powerful AI model introduced in this competition typically requires exponentially more resources than its predecessors.

This includes more computing power for training and operation, larger and more complex training datasets, and a steadily increasing energy consumption.

The race for ever-larger context windows, for multimodal capabilities that can process text, image, sound, and video, and for real-time responses has real, tangible physical consequences. The training phases for top models now consume amounts of energy comparable to the annual consumption of small cities.

Every API call, every user interaction, every seemingly effortlessly generated answer consumes electricity, burdens server farms, and requires a massive, energy-intensive infrastructure.

Yet, in glossy PR campaigns and euphoric product presentations, these new models are often portrayed as "smart apps" or immaterial software solutions. Their ecological footprint is hardly ever discussed. In reality, however, they are gigantic, energy-hungry computing clusters with a significant ecological and economic shadow.

Every additionally generated token, every further millisecond of computing time is not "free" in the comprehensive sense. The costs are often externalized, onto the environment, society, or future generations.

The question of the sustainability of this resource-intensive race and whether the achieved benefits justify the enormous costs is rarely asked. This immense hunger for resources could also mean that the use of AI for less commercially attractive but socially important research areas, which do not have comparable budgets, becomes unaffordable. Is the "best" AI therefore necessarily also the most resource-intensive, and is this path truly sustainable and just in the long run?

Another critical symptom of the current competition is the often one-sided focus of innovation. Far too often, "innovation" in the AI field is confused with the mere improvement of performance values in standardized benchmarks.

Models are trained with enormous effort to shine in tests like MMLU (Massive Multitask Language Understanding), HumanEval (code generation), or ARC (AI2 Reasoning Challenge) and to outperform competitors by a few percentage points.

While these benchmarks are useful tools for measuring certain sub-aspects of AI performance, they by no means cover the entire spectrum of what would make an AI "better" or "more useful."

Aspects such as the security and robustness of generated outputs, the transparency and explainability of model decisions, control over long-term context behavior and the avoidance of semantic drift, or the ability for genuine, creative problem-solving in complex, open domains are often neglected or insufficiently prioritized in this benchmark-driven race. Why?

Because there are rarely direct, easily quantifiable rewards, high-profile rankings, or immediate competitive advantages for these qualitative aspects.

But this is precisely one of the real dangers of this development. If AI models are primarily trained to generate public attention and admiration by passing standardized tests or creating superficially impressive demonstrations, but internally can provide no traceable explanation for their "decisions" or do not recognize their own knowledge limits, then any apparent improvement on the surface is potentially a step backward in the deeper understanding and controllability of these systems.

A rethinking must take place here. The focus of AI development and competition should broaden. It's not just about creating an AI that passes general tests or serves popular applications.

A truly "better" AI would be one that can serve as a reliable, transparent, and powerful tool for research.

It should help us to address humanity's pressing problems more effectively, from diseases to climate change to complex scientific puzzles. Such an AI would not necessarily have to shine in every generic benchmark. Rather, it should prove its strength in specialized, in-depth analysis, in creative hypothesis generation, and in the secure handling of complex knowledge domains.

The competition should also revolve around who develops the AI that best helps us expand our own intelligence and better understand and shape the world.

What does "better" truly mean in the context of Artificial Intelligence? Is it more tokens per second that a model can generate? Is it a higher hit rate in abstract test tasks? Is it a slightly lower hallucination rate in standardized dialogues?

Or isn't an AI much "better" when it is capable of at least partially explaining its own workings, recognizing and communicating its own limits, and when it can assume responsibility for the consequences of its actions within a traceable framework?

The industry and the research community face a crucial choice. Do we want to continue primarily building machines that impress mainly through their sheer performance and their ability for superficial simulation of human conversation?

Or do we strive for machines that not only shine but also endure, that are reliable, safe, and useful in the best sense of the word? The current market and media hype primarily reward speed, novelty, and superficial brilliance.

However, the truly decisive qualities of a future-proof AI – namely security, truth, trust, and genuine problem-solving competence – grow only in depth. They arise from careful research, ethical reflection, and a tireless pursuit of true comprehensibility and control.

These qualities cannot simply be measured in standardized benchmarks or expressed in quarterly figures. A reorientation of the competition is needed, towards an AI that is not only the "best" in terms of performance, but the "best" in terms of being the most responsible, most transparent, and most beneficial for humanity.

“We don't need evil machines – just stupid decisions to trigger the catastrophe.” – Anonymous note on the margin of a test plan