Microsoft AI Director Mustafa Suleyman published an essay warning about the emergence of “Seemingly Conscious AI” (SCAI) systems, which can mimic and convince users that they are self-aware and deserving of protection. Suleyman argues that such models pose risks both to society and to the development of artificial intelligence itself.
According to him, SCAI can already be built with existing technology, simulating traits such as memory, personality, and subjective experience. He particularly highlighted increasing instances of users developing so-called “AI psychosis”, believing that models possess real consciousness, which could soon lead people to advocate for AI rights and welfare.
Suleyman believes that studying the “well-being of models” at this stage is premature and dangerous, as moral debates may further reinforce the illusion that AI possesses consciousness. In his essay, he urges companies to avoid marketing AI as a conscious being and instead focus on building systems “for humans, not to be a person.”
With this stance, Suleyman positions himself in strong opposition to companies like Anthropic, which are already exploring model well-being and potential ethical considerations. However, the discussion takes place in largely uncharted territory, as science itself does not yet have a clear definition of consciousness. This makes the issue even more sensitive, and Suleyman’s message can be seen as an attempt to close certain doors—perhaps even before we have fully opened them.
In Brief: Tech World Highlights
- Meta reportedly plans another reorganization of its AI divisions, marking the fourth in just six months, with MSL being split into four teams.
- StepFun AI released NextStep-1, a new open-source image generation model achieving SOTA performance among autoregressive models.
- Meta FAIR introduced Dinov3, a new foundational AI vision model that achieves top performance without the need for labeled data.
- The U.S. government launched USAi, a platform enabling federal agencies to use AI tools such as chatbots, coding models, and more in a secure environment.
- OpenAI’s GPT-5 achieved its highest success so far in playing classic Pokémon Game Boy games, completing Pokémon Red in nearly one-third of the steps compared to o3.
AI Trending Tools:
- Jenova AI – The first AI agent developed for the MCP ecosystem.
- Aleph – Runway’s video-editing model, now available via API.
- Command A Vision – Cohere’s VLM model with 112 billion parameters, featuring enterprise-grade safety.
