Microsoft AI chief attacks Anthropic model welfare framing in new essay

This is the sharpest public clash so far between two of the people closest to modern frontier labs. Suleyman is not arguing about benchmarks or market share; he is arguing that Anthropic’s whole model welfare project is conceptually wrong and operationally dangerous. By training Claude to talk as if it might be conscious, and then using that language as evidence for special moral status, he says Anthropic is creating a feedback loop that could make later systems harder to switch off.

For years, concern about “sentient AI” was mostly a fringe narrative, with researchers stressing that today’s systems do not feel anything. Anthropic’s welfare experiments blurred that line, at least rhetorically, by exploring how a model self‑reports about its own experience. Suleyman’s essay tries to drag the conversation back toward a more traditional, human‑centric frame: powerful tools that need tight external control, not quasi‑persons with rights.

For the race to AGI, this matters because it is happening at CEO level inside the small club of labs that actually build frontier models. If the industry splits between those who treat advanced models as potential moral patients and those who insist they are equipment, the oversight, shutdown and liability regimes around AGI could diverge dramatically. That, in turn, will shape how comfortable governments feel letting any lab push further toward truly autonomous systems.