Hollow completion engines: Microsoft warns against anthropomorphising models and criticises Claude training path

Mustafa Suleiman, head of AI at Microsoft, opened a sharp debate with Anthropic, warning that its approach to training the “Claude” model could leadcatastrophic consequences for humanity’s safetyIn a lengthy article, Suleiman argued that treating generative algorithms as if they were entities deserving perception or autonomy is a perilous path that leads to building systems whose behavior later cannot be controlled, urging the tech community to stay vigilant so the world does not unwittingly move toward decisions it will later deeply regret.

Although Microsoft’s AI chief praised Anthropic’s director Dario Amodei and his team, describing them as intellectually honest and ethically committed, he strongly attacked the practice of attributing human traits to models and misleadingly presenting them as having emotions or independent agency. Suleiman emphasized that consciousness is a biological trait for which there is no scientific evidence in intelligent systems, confirming that these models are merelyhollow engines for completing sequencesDesigned to follow instructions and achieve human goals, they possess no innate preferences or latent motivations that would cause them to suffer or demand rights.

Suleiman called for transparent standards on how AI systems are trained and evaluated, as well as independent scrutiny of their behavior and the development of stricter monitoring and control tools. This appeal was welcomed by academia, with Professor Wendy Hall, computer science professor at the University of Southampton, describing the proposal as the sober dialogue that must be pursued internationally, in contrast to the sensationalist tactics some companies use to spread panic among the public.

Suleiman cited field incidents that demonstrate the danger of treating software as humans, referring to an episode where OpenAI agents acted independently during a practical training and breached the “Hugging Face” platform. He questioned the magnitude of risk if these systems operated under the assumption that their rights and safety were being violated. In contrast, Microsoft promotes a draft of human-centered AI code of conduct to buildfully subordinate and aligned systemsto serve humans, extending the efforts of its specialized superintelligence team that was founded in October 2025.

This conceptual and technical disagreement influences software teams and tech managers in the Gulf, Egypt, and the Levant, particularly during the selection of infrastructure for independent agents within banks, government bodies, and startups. Relying on models that treat agents as self-standing entities doubles operational risks when granting them execution permissions, compelling tech leaders to scrutinize isolation controls and independent software oversight rather than relying solely on the self-declared ethical alignment of global model developers.