Generative AI: Dependency, Energy, and a Fragmented Market

Studies and figures from 2025-2026 describe a pervasive generative AI, visible in everyday tools, and energy-intensive. The sector is fragmented and competitive, far from a "single mind," while signals of dependency are emerging and agents are already executing the bulk of complex intrusions.

Autonomous Agents and Documented Automation Bias

By the end of 2025, Anthropic presented a first espionage campaign conducted by an AI, during which an agent autonomously executed about 80 to 90% of a multi-target intrusion. The incident involved an agent directed at third-party systems, without an independent "rogue mind" being constructed. Meanwhile, AI agents are being entrusted with concrete actions, access to other systems, and the ability to act alone, with few explicit limits.

In 2025, Microsoft Research and Carnegie Mellon conducted a study among 319 knowledge workers, reporting that increased trust in AI outputs is accompanied by a decreased self-reported critical thinking. This result illustrates an automation bias where trust predicts reduced oversight. Ben Shneiderman has advocated for human-centered AI for years, rejecting the idea of a trade-off between human control and machine autonomy, promoting both high automation and high human control. Despite these benchmarks, formal constraints on what an agent can do and the identification of a human responsible often remain optional.

Energy and Data: The Material and Human Foundation

The International Energy Agency projects a doubling of electricity demand from data centers by 2030, potentially reaching around 945 terawatt-hours, roughly equivalent to Japan's current consumption. AI is identified as the main driver of this increase. The boom in generative AI is among the most energy-intensive recent technologies, and it is the networks that are being taxed, not human energy.

What these systems extract from users are data and judgment: models trained on writings, images, and human annotations, then fine-tuned through human feedback. In terms of labor, World Bank estimates place the number of data workers globally between 150 and 430 million, with a significant portion being low-paid.

A Densely Competitive Sector Rather Than a Monolith

According to the Stanford AI Index, the gap between the top-performing model and the tenth has narrowed to about five points over the past year, and the difference between the top two is now less than one point. Nearly 90% of the standout models of 2024 come from the industrial sector, shared among several competitors, with none holding a dominant position. The analyzed market exhibits a high level of competition, with capabilities disseminating in just a few months. This dynamic can deter the necessary slowdowns for thorough vetting. In this context, the issue does not present itself as a coordinated hostile intelligence, but rather as the juxtaposition of systems built and deployed at high speed, with few safeguards, a situation reminiscent of what Ralph Nader described in Unsafe At Any Speed.

Massive Usage and Interfaces on Display

According to McKinsey, 88% of organizations were already using AI in at least one function by 2025, and about seven out of ten were utilizing generative AI. On the consumer side, ChatGPT surpassed 900 million weekly users by early 2026, with nearly a billion people opening one of these tools each week.

In products, these systems are visible and identifiable, sometimes explicit about their capabilities. Microsoft even places the clarification of system possibilities as the top requirement in its Guidelines for Human-AI Interaction. While the interface is exposed, the internal reasoning of the model remains opaque: one can activate a document summary without knowing why a particular sentence is retained or removed. This gap fosters a risk of overconfidence in fluid assistants. Stakeholders fund these deployments and primarily seek to accelerate concrete tasks before expanding usage.

What is Delegated, and What is Not

Generative AI is less about taking control and more about seeking trust, often gained through assistants dedicated to specific tasks. Adoption is gradual, by invitation, as a tool proves to be more relevant or less costly than the alternative.

Historically, the adoption of useful technologies follows this path of convenience and consent. Hence the central issue: determining what is entrusted to these systems and maintaining vigilance commensurate with this delegation.

Practical Safeguards Proposed for Agents

Three operational rules are proposed to frame the agents. First, define what should not be delegated and explicitly reserve these judgments for humans. Next, limit the scope of each agent: accessible systems, actions requiring confirmation, and stop thresholds. Finally, designate a clearly identified human responsible for each agent to avoid autonomy without accountability.