In a wide-ranging conversation on The Verge's Decoder podcast, Microsoft AI CEO Mustafa Suleyman doubled down on his view that AI poses genuine, near-term risks — and took pointed aim at Anthropic, arguing that the company's approach to AI safety is making the broader problem worse rather than better.
A Candid Take on the AI Safety Debate
Suleyman, who co-founded DeepMind before joining Microsoft to lead its consumer AI efforts, has long positioned himself as both an AI optimist and a safety realist. In the interview, he reiterated that the threats from advanced AI systems are "real" and that the industry has a responsibility to steer development toward alignment — the technical and philosophical challenge of ensuring AI systems act in accordance with human values and intentions.
But he reserved his sharpest criticism for Anthropic, the AI safety-focused lab behind the Claude family of models. Suleyman suggested that Anthropic's public posture — framing itself as the responsible counterweight to more aggressive competitors — has the perverse effect of raising the stakes of the safety debate without meaningfully improving outcomes. In his view, positioning AI safety as a binary struggle between cautious and reckless actors distorts the conversation and slows down the practical work of building aligned systems.
Anthropic and the "Safety Theater" Critique
The tension reflects a broader split in the AI industry. Anthropic has built its brand around "Constitutional AI," interpretability research, and responsible scaling policies, and has frequently warned about catastrophic risks from frontier models. Suleyman's critique echoes a growing camp — including some researchers and policymakers — who argue that aggressive safety branding can function as competitive marketing, and that it entrenches a narrative in which only a handful of labs can be trusted with frontier development.
The debate matters more in 2026 than ever, as regulators in the U.S., EU, and U.K. move from drafting frameworks to enforcing them. With the EU AI Act in full application and emerging U.S. federal rules on frontier model evaluation, how labs frame their own safety commitments is no longer just a PR question — it shapes compliance, liability, and market access.
AI Is Not Consciousness
Suleyman also pushed back firmly on the tendency to anthropomorphize AI systems. He stressed that today's models — however fluent, persuasive, or seemingly self-aware — are not conscious. They do not have feelings, intentions, or subjective experience. The confusion, he argued, is partly a messaging failure by the industry itself: by using human-like language ("thinks," "wants," "understands"), labs invite the public to over-attribute inner life to machines, which in turn muddies both safety discussions and user expectations.
This distinction matters concretely in 2026, as companion chatbots, agentic assistants, and emotionally attuned models become mainstream. Regulators and researchers have increasingly flagged the risk of user over-reliance and misplaced trust in systems that simulate empathy without possessing it.
Toward Alignment: How the Industry Moves Forward
Rather than treating safety as a marketing differentiator, Suleyman argued for a more pragmatic, industry-wide push toward alignment. That includes:
- Shared technical standards for evaluating model behavior and failure modes
- Transparency about capabilities and limits, rather than selective disclosure
- Collaboration across labs, including competitors, on safety infrastructure and red-teaming
- Realistic public communication that neither hypes AI as godlike nor dismisses its genuine risks
The through-line of Suleyman's argument is that safety is an engineering and governance problem to be solved collectively — not a brand attribute to be weaponized. His critique of Anthropic, whether fair or not, lands in the middle of an ongoing fight over who gets to define what "responsible AI" means as the technology moves from demo to deployment.
As 2026 unfolds, that fight will play out in courtrooms, standards bodies, and boardrooms — not just on podcast microphones. How the major labs resolve their differences over safety, and whether they can align on substance rather than messaging, will shape the next decade of AI development.
via The Verge AI