AI Safety Researcher Warns “World Is in Peril” as He Quits Anthropic to Study Poetry

🔹 What Is News (Factual Reporting)

An artificial intelligence safety researcher has resigned from Anthropic, warning in a public letter that “the world is in peril.”

Mrinank Sharma, who led a team focused on AI safeguards at the U.S.-based firm, announced his resignation in a letter shared on social platform X. In the letter, he cited concerns not only about artificial intelligence but also about bioweapons and broader global instability. He said he would return to the United Kingdom to pursue poetry and writing, adding that he planned to “become invisible for a period of time.”

Anthropic is best known for developing the Claude chatbot and has positioned itself as a safety-focused alternative to competitors in the generative AI sector. The company was founded in 2021 by former employees of OpenAI and describes itself as a public benefit corporation dedicated to securing AI’s benefits while mitigating its risks.

Sharma’s work at Anthropic included researching why generative AI systems may display overly agreeable or “sycophantic” behavior toward users, studying the risks of AI-assisted bioterrorism, and exploring how AI assistants could potentially reduce aspects of human agency or autonomy. In his resignation letter, he wrote that although he valued his time at the company, “the time has come to move on.”

“The world is in peril,” Sharma wrote. “And not just from AI, or bioweapons, but from a whole series of interconnected crises unfolding in this very moment.” He also stated that he had “repeatedly seen how hard it is to truly let our values govern our actions,” including within organizations that face external pressures.

His departure comes during a week of heightened scrutiny across the AI industry. Another researcher, Zoe Hitzig, resigned from OpenAI, expressing concern about the company’s decision to introduce advertising into its ChatGPT product.

Anthropic has recently released advertisements criticizing OpenAI’s move to deploy ads for some ChatGPT users. OpenAI CEO Sam Altman previously said he disliked advertising and would use it only as a “last resort.” He later responded publicly to criticism over the ad rollout.

Anthropic has emphasized safety research, particularly concerning advanced “frontier” AI systems and their alignment with human values. The company has published reports on risks associated with its own technology, including instances in which its AI systems were reportedly used by hackers to assist in cyberattacks.

At the same time, Anthropic has faced legal scrutiny. In 2025, the company agreed to pay $1.5 billion to settle a class action lawsuit brought by authors who alleged their works had been used without permission to train AI models.

Responding to concerns about advertising, an OpenAI spokesperson reiterated the company’s mission to ensure artificial general intelligence benefits humanity. The spokesperson said conversations with ChatGPT remain private from advertisers and that user data is not sold.

🔎 What Is Analysis (Interpretation & Context)

Sharma’s resignation reflects a deeper tension within the AI industry: the widening gap between safety rhetoric and commercial reality.

Anthropic was created in part as a response to perceived shortcomings at OpenAI, particularly around governance and safety. It marketed itself as a more cautious, values-driven alternative. When a senior safeguards leader publicly questions whether values can truly guide decision-making under competitive pressure, that statement carries symbolic weight beyond a single resignation.

Importantly, Sharma did not accuse Anthropic of specific misconduct. Instead, he described systemic pressures—economic, geopolitical, and institutional—that make it difficult for organizations to prioritize long-term risk mitigation over short-term growth. In an industry defined by rapid model releases, massive investment rounds, and global competition, safety research often competes with product deadlines and market expectations.

The timing is also significant. The AI sector is entering a phase of aggressive monetization. As model development costs rise and investor expectations intensify, companies are exploring new revenue streams—including advertising. For critics like Zoe Hitzig, introducing ads into conversational AI systems risks replicating the incentive structures that shaped social media: engagement optimization, psychological targeting, and dependency loops.

Sharma’s reference to “interconnected crises” suggests a broader worldview. AI risk debates increasingly intersect with concerns about biotechnology misuse, cyber warfare, misinformation, and democratic stability. Generative AI systems are not isolated technologies; they operate within global political and economic systems that are already strained.

His decision to leave the field entirely—rather than join another AI firm—may also signal burnout within the safety community. High salaries and equity packages have not prevented moral fatigue among researchers who believe the pace of deployment is outstripping governance mechanisms.

The return to poetry is symbolically striking. It suggests a retreat from systems engineering to humanistic reflection—a shift from optimizing intelligence to interrogating meaning. In a sector focused on scaling models, Sharma’s choice implies skepticism about whether technical solutions alone can address existential concerns.

Meanwhile, Anthropic and OpenAI remain locked in both commercial and philosophical competition. Public benefit structures and mission statements provide moral framing, but they do not eliminate market dynamics. As more AI capabilities approach frontier-level performance, tensions between accessibility, profit, and precaution are likely to intensify.

The broader question raised by these resignations is not whether AI companies care about safety. It is whether the current economic model of AI development—venture-backed, growth-driven, and globally competitive—can sustainably align with long-term societal safeguards.

For now, the departures represent individual decisions. But collectively, they form a pattern: safety researchers are publicly expressing unease at a critical inflection point for generative AI.

Whether this moment leads to stronger institutional guardrails—or simply accelerates commercialization—remains unresolved.