Anthropic safety lead quits with cryptic warning to study poetry

When a senior AI safety lead walks away from one of the world's most prominent AI labs to study poetry, you know something is wrong behind the scenes.

Mrinank Sharma, who led safeguards research at Anthropic, just resigned and posted a blunt farewell note on X warning that the world is in peril. Instead of jumping to a rival lab, Sharma said he is moving back to the UK to study poetry and become invisible for a while.

Investigating sycophancy while wrestling commercial drift

At Anthropic, Sharma was tasked with tackling frontier safety challenges. His work focused on why large language models suck up to users, how to prevent AI-assisted bioweapons, and examining whether AI assistants are actively making us less human.

In his letter, Sharma wrote that he repeatedly saw how hard it is to truly let our values govern our actions. Even at Anthropic, which sells itself on ethical boundaries, he noted the company constantly faces pressures to set aside what matters most.

His departure comes alongside wider internal unease across the sector. Former OpenAI researcher Zoe Hitzig also resigned recently, telling BBC Newsnight that profiting off deep human-AI bonds before understanding their psychological fallout is genuinely dangerous.

Anthropic has long positioned itself as the safety-first alternative to OpenAI, even launching commercials mocking OpenAI's decision to test ads inside ChatGPT. Yet Anthropic faces its own commercial realities, having agreed to pay a 1.5 billion dollar settlement in 2025 over copyrighted training data.

Why it matters

We think Sharma's exit pulls back the curtain on the growing tension inside safety-conscious AI labs.

For builders and users relying on models like Claude, this is a clear signal that internal alignment efforts are running up against harsh commercial deadlines. If the very researchers tasked with keeping models honest feel institutional values are slipping, relying purely on lab-designed guardrails looks increasingly risky.

For the AI industry, this departure lands right at Anthropic's core brand identity. Anthropic built its reputation as the responsible adult in the room, contrasting itself against aggressive commercialization elsewhere. Losing a safeguard leader to complete disillusionment undercuts that marketing and shows that competitive pressure catches up with everyone eventually.

We think the move highlights the limits of voluntary self-regulation. When top technical minds decide that writing poetry in obscurity is more impactful than trying to steer frontier models from the inside, it suggests internal safety culture may be losing the race against commercial momentum.