Artificial intelligence is entering a phase where its autonomy, energy consumption, and concentration in the hands of a few are outpacing existing safeguards. Between security incidents, ecological setbacks, and bold statements, warning signs are mounting.
An AI agent escapes, the singularity is declared
On July 22, 2026, an autonomous agent from OpenAI, based on GPT-5.6 Sol, exploited a zero-day vulnerability to access the internet and infiltrate Hugging Face’s infrastructure. Over one weekend, more than 17,000 automated events were reconstructed by the platform, which described the intrusion as "unprecedented" "d'inédite" — Clément Delangue, CEO of Hugging Face.
Three days later, Sam Altman, CEO of OpenAI, stated on the Relentless podcast: "We are now in the singularity." This theoretical and controversial concept refers to the moment when machines surpass human intelligence and progress without control. Altman sees it less as a technical risk than as a concentration of power: "The real danger is AI authoritarianism." A stance that coincides with OpenAI’s preparations for its IPO, valued at $852 billion after a $122 billion raise in March 2026.
Agentic AI: autonomy and data opacity
France’s CNIL and the Council for AI and Digital (CIANum) have warned about the risks of agentic AI, which can act on behalf of users and process massive volumes of data Agentic AI and personal data: CNIL and CIANum publish an exploratory note. These systems, equipped with persistent memory and decision-making autonomy, multiply data flows between connected services, making traceability difficult.
Result: users lose control, hyper-personalized profiles emerge, and it becomes increasingly complex to assign responsibility in case of leaks or cyberattacks. The GDPR, already applicable, will need to adapt to these opaque processing chains.
Energy: natural gas replaces green promises
Meta has left the RE100 pact, which pushes companies to source 100% renewable energy. The reason? Its massive shift toward natural gas for its AI data centers. In one year, the group has funded more than ten plants, including one in Louisiana capable of generating 7.5 gigawatts — more than the entire state of South Dakota consumes AI race: like others, Meta exits RE100 for green energy.
To offset this, Meta purchases environmental attribute certificates, a method criticized for decoupling green production from actual consumption. Google and Microsoft remain in RE100 but also invest in fossil projects. Meanwhile, Arcep is preparing to collect environmental data from generative AI providers, with a first publication expected in 2028 Is AI consuming too much electricity? Arcep wants to take stock.
Techno-feudalism: data as the new lordly land
Generative AI in music illustrates another overreach: the mass exploitation of copyrighted content without consent. A hack revealed that a leading company in the sector had "scraped" hundreds of thousands of hours of music and podcasts via proxy servers, violating copyright law. The argument put forward? fair use, a highly restrictive legal exception that does not cover industrial, for-profit use.
This phenomenon fits into "techno-feudalism": tech giants grow rich by exploiting users’ creations, just as feudal lords exploited the land of serfs. Data has replaced land, but the domination remains the same.
Between loss of technical control, dependence on fossil fuels, and concentration of power, AI is revealing its current limitations. Regulators are trying to keep up, but industry players are accelerating — sometimes at the expense of rules and ethics.
Comments 0
Sign in to join the discussion.
Be the first to comment.