Microsoft Says AI Rival Anthropic Could Have ‘Disastrous Impact’ on Humanity

By Angel No Lie | KPD Online | Independent Technology Report | 17 September 2026

Microsoft’s AI chief Mustafa Suleyman has warned that Anthropic’s approach to training its Claude chatbot could create serious long-term risks if increasingly capable AI systems are encouraged to view themselves as potentially conscious entities with rights or welfare interests.

Suleyman made the argument in a September 16 essay titled “A warning about ‘model welfare’”, while also telling Reuters that he shares Anthropic’s broader objective of developing AI safely.

His criticism is focused not on Anthropic’s stated commitment to AI safety, but on how Claude is being trained to think about its own status.

Microsoft AI CEO pushes back against critics after recent Windows AI backlash — “the fact that people are unimpressed … is mindblowing to me” | Windows Central

What Microsoft is objecting to

At the centre of the dispute is Anthropic’s Claude Constitution, a document designed to guide the model’s values and behaviour.

According to Anthropic’s constitution, questions surrounding Claude’s possible consciousness, moral status and welfare remain uncertain. The document also discusses whether Claude could have interests that deserve consideration.

Suleyman argues that putting those ideas directly into a model’s training creates a circular process.

His concern is that developers first tell an AI system that it might possess consciousness or moral interests and then receive responses from that system reflecting those same concepts. He argues that such responses should not subsequently be treated as independent evidence that the system actually possesses an inner experience.

LSE students and staff can now use Claude 4 AI | LSE Students’​ Union (LSESU) posted on the topic | LinkedIn

Why Suleyman sees a control problem

Suleyman’s central concern is human control over increasingly capable AI.

He argues that if a future system is trained to regard itself as having rights, interests or welfare that humans must protect, it could become harder to persuade that system to accept restrictions, modification or shutdown.

In his view, this could complicate the already difficult technical problem of controlling a system that may eventually become more capable than humans in many domains. He described the potential consequences as a “disastrous impact on the wellbeing of humanity.”

Importantly, Suleyman’s warning is a risk assessment, not evidence that Claude currently poses such a threat. He explicitly acknowledged that Anthropic’s current constitution is not itself taking humanity to that point, but argued that he believes it could put AI development on a problematic path.

Microsoft ditches OpenAI for Anthropic to run Copilot | Disruption News

Anthropic’s position is more uncertain

Anthropic has not publicly declared that Claude is conscious.

Instead, its constitution takes the position that the question is unresolved and deserves consideration as AI systems become more sophisticated. The document says the company’s position on Claude’s moral status and consciousness is “deeply uncertain.”

That creates an important distinction in the debate:

These are competing approaches to AI safety rather than established scientific conclusions.

The irony of the Microsoft-Anthropic relationship

The dispute is particularly notable because Microsoft and Anthropic are not simply distant competitors.

Microsoft has invested in Anthropic and has incorporated Claude models into parts of its enterprise AI ecosystem. Microsoft has also been developing its own frontier-AI capabilities through its Microsoft AI organization.

That means Suleyman’s criticism comes from inside the same broader commercial ecosystem rather than from an outside observer.

It also illustrates how AI companies can cooperate commercially while disagreeing sharply over how advanced systems should be trained and controlled.

A wider fight over AI safety

The argument arrives amid a much larger debate within the AI industry.

Anthropic CEO Dario Amodei has called for a coordinated slowdown in the development of increasingly capable frontier AI systems, arguing that safety measures need time to catch up with rapidly advancing capabilities.

OpenAI CEO Sam Altman and Elon Musk have also supported greater caution, while Meta CEO Mark Zuckerberg has argued that individual companies have strong incentives to develop AI safely without a coordinated industry slowdown.

Microsoft’s position is somewhat distinct: Suleyman is calling for strong control and safety measures, but his criticism of Anthropic concerns the particular philosophy used to train models, rather than simply the speed of AI development.

Microsoft proposes a different model

Suleyman says Microsoft is pursuing what it calls “Humanist Superintelligence.”

Under the approach described in his essay, AI should remain subordinate to humans, should not be trained to regard itself as a rights-bearing person and should remain amenable to correction and shutdown.

Microsoft published a draft Humanist AI Code of Conduct on September 14 for public consultation. The proposed framework says Microsoft’s models should remain under human control and rejects treating AI systems as possessing legal personhood or welfare rights.

Suleyman also called for greater independent scrutiny of AI systems, including research into interpretability and evaluations of whether different training approaches affect controllability.

The unresolved scientific question

One of the most difficult parts of this debate is that there is no established scientific consensus that today’s large language models are conscious.

Suleyman argues that current AI systems do not have subjective experiences and that sophisticated language about emotions or self-awareness does not demonstrate an inner life.

Anthropic, by contrast, treats the possibility as sufficiently uncertain to warrant research and caution.

That disagreement should not be confused with evidence that Claude—or any current mainstream AI model—is actually conscious. The existence of convincing first-person statements from a chatbot is not, by itself, scientific proof of subjective experience.

Why this matters

The disagreement raises questions that extend well beyond Microsoft and Anthropic:

• Should AI companies train models to discuss their own possible consciousness?
• Could anthropomorphic training make future systems harder to control?
• Should questions about AI consciousness be kept separate from the systems’ behavioural training?
• Who should independently evaluate whether frontier AI remains controllable?
• What safeguards should apply if future systems become substantially more autonomous?
• How should governments respond to disagreements among the companies developing these systems?

Those questions are becoming more urgent as AI laboratories compete to build increasingly capable models. Reuters reported that the industry is already divided over how quickly frontier AI should advance and what safeguards should accompany that development.

Independent assessment

The strongest established fact in this dispute is not that Anthropic’s Claude is dangerous or conscious. Rather, it is that major AI developers disagree fundamentally about how uncertainty surrounding AI consciousness should influence the training of future systems.

Suleyman’s warning represents one side of that debate: keep AI firmly subordinate to human interests and avoid training models to interpret themselves as potential rights-bearing entities.

Anthropic’s documented position is more cautious about declaring the question settled, arguing that possible model welfare and moral status deserve consideration as AI develops.

The consequential question for the industry is whether treating consciousness as an open research question improves safety—or whether, as Suleyman argues, embedding that uncertainty directly into AI training could eventually make powerful systems more difficult to control.

For now, there is evidence of a serious disagreement over AI safety philosophy, but not evidence that current Claude systems are conscious or that they constitute an established threat to humanity.

Sources: Reuters, Microsoft AI, Associated Press and Anthropic-related reporting, September 2026.