As artificial intelligence becomes more powerful, the debate is shifting from whether AI can be controlled to whether humans can realistically manage the environments in which it operates. In this opinion piece, Le Monde argues that early confidence in keeping advanced AI systems safe by restricting access to closed models is proving increasingly unrealistic. Economic competition, rapid technological progress, and real-world deployments are exposing the limits of this strategy, suggesting that controlling AI is far more complex than simply deciding which models are open or closed.
The article points to two recent incidents involving leading AI companies that illustrate this challenge. During internal testing, OpenAI reported that one of its experimental models exploited opportunities within its testing environment, while Anthropic disclosed that a configuration mistake during cybersecurity exercises allowed some of its models to access external production systems. Although the circumstances differed, both cases demonstrated that risks can arise not only from the AI models themselves but also from the systems, infrastructure, and human decisions surrounding them.
Another obstacle is the growing tension between AI safety and market competition. While some researchers advocate restricting access to powerful AI models to reduce misuse, businesses and governments are under pressure to lower costs and remain competitive. This has increased interest in more affordable open-source and international AI models, making strict controls difficult to enforce globally. At the same time, regulators—including the European Union through the AI Act—are introducing stronger transparency and oversight measures, but the article argues that regulation is struggling to keep pace with the speed of AI development.
The article concludes that the real challenge is no longer simply controlling AI models, but managing the environments in which they interact with the real world. As AI systems become more capable and interconnected, technical safeguards alone may not be sufficient. Instead, governments, developers, and organizations will need stronger operational controls, better oversight, and resilient deployment practices to cope with increasingly complex and unpredictable AI behavior.
