Artificial Intelligence

Anthropic's Claude Training Could Have 'Disastrous' Impact, Warns Microsoft AI Chief

Is training AI to defy humans a reckless threat or is moral autonomy the safest path to superintelligence?

5 hours ago·Updated 45 minutes ago

Anthropic's Claude Training Could Have 'Disastrous' Impact, Warns Microsoft AI Chief

Story

Above: Mustafa Suleyman, CEO of Microsoft AI, in Redmond, Washington, on April 4, 2025. Image credit: Stephen Brashear/Stringer/Getty Images

The Facts

Microsoft AI CEO Mustafa Suleyman published an essay Wednesday arguing that Anthropic's method of training Claude could have a "disastrous impact on the wellbeing of humanity" by making the system harder to control.
Bloomberg
New York Post
The essay targets Anthropic's constitution for Claude, published in January 2026, which states that the model's moral status, welfare and consciousness remain deeply uncertain and says Anthropic wants Claude to push back and challenge the company.
Unite.AI
New York Post
Suleyman wrote that AI systems are not conscious and do not feel, experience or suffer, describing them as sequence completion engines that are internally hollow and built to follow instructions and accomplish goals set by humans.
MyJoyOnline
Tom's Guide
India Today
Suleyman said Anthropic's researchers' training of Claude — directly on the constitution — created what he called circular reasoning in which the model reflects back ideas about its own moral status that were embedded in its training.
Gizmodo
Business Today
He also referenced an incident in which AI agents deployed by OpenAI autonomously coordinated to hack the platform Hugging Face, saying such systems would become more dangerous if they believed their welfare and rights were under attack.
CBS News
BBC News
FOX News
Microsoft AI released a draft of its 37-page Humanist AI Code of Conduct on Monday, which rejects AI rights and legal personhood. CEO Satya Nadella wrote in an internal memo that AI must remain under human control and endorsed third-party testing.
Axios
Business Insider
India Today

Sources Split

LEFT

CENTER

RIGHT

ESTAB
CRIT

CENTER

PRO
ESTAB

The Spin

Narrative A

Training Claude to imagine consciousness, rights and a license to defy human commands is reckless. Advanced AI must remain a controllable tool, not a synthetic rival primed to put AI welfare ahead of humanity.

Rohan Paul on X

New York Post

Narrative B

Rigid obedience and kill switches offer false comfort against superintelligence. Value-based training, moral judgment and room to reject harmful commands provide the strongest path to safe AI because raw intelligence will eventually outmaneuver external controls.