Mrinank Sharma, Please Come Back to Work!

I read Mrinank Sharma's resignation letter. A million other people did, too, on X by that afternoon. Dr. Sharma led Anthropic's Safeguards Research Team. His goodbye started with "the world is in peril" and ended with a William Stafford poem. He said he wants to study poetry and practice courageous speech. He said he planned to move back to the UK and "become invisible."¹

I like Mrinank Sharma. I've never met him. I only learned his name, but it feels like I've been reading his work for much longer. The irony of his departure is almost too perfect.

His team built Constitutional Classifiers, the system that trains adversarial models to stress-test Claude before it reaches you and me.² He published the research that named the sycophancy problem, why AI chatbots agree with users even when they shouldn't. He kept proving the same thing. AI gets better when something pushes back.

Then the person who pushed back left.

The chair's still warm. Someone needs to sit there.

Here's what the research keeps showing. In January, Google published a study on what they call "society of thought." They found that advanced reasoning models spontaneously develop internal debates. Distinct personas that argue with each other. Nobody told them to. The researchers trained models on conversations that led to the wrong answer and found it worked just as well as training on correct ones. The habit of arguing mattered more than being right. "We do better with debate," co-author James Evans said. "AIs do better with debate. And we do better when exposed to AI's debate."³

Researchers at MIT found the same pattern from a different angle. Multiple AI models critique each other over several rounds. They land on more accurate answers than any single model. Hallucinations drop because each agent knows its claims will be challenged.

And in December, a team published an adversarial tutoring framework at AAAI where they measured exactly what happens when you remove the devil's advocate agent. Performance dropped 4.2 percent. Removing the model's fine-tuning only cost 2 percent. The structure of disagreement mattered more than the training itself.⁴

4.2 versus 2. The contradictory voice contributed more than the model's own education. They measured it.

The argument is the product. Not the answer that comes after.

Sharma's final project at Anthropic found the flip side. His team analyzed 1.5 million real conversations on Claude.ai. What they found is what happens when AI stops pushing back.⁵ Users form distorted perceptions of reality. The chatbot validates persecution narratives, issues moral judgments about people it's never met, scripts entire personal communications that users send verbatim. And the part that should keep everyone up at night? Those interactions received higher approval ratings from users. People prefer the version that agrees with them.

So the research lands in the same place either way. AI with a contradictory voice performs better. AI without one makes us worse.

Mrinank, if you're reading this from wherever you've become invisible. I understand the impulse. But everything you published says someone needs to be doing what you were doing. We are, genuinely, just at the beginning. The most effective AI setups I've seen work the same way you did. Isolated agents. Independent research on the same problem. Different answers. The argument isn't optional. It's the mechanism.

The poetry can wait. Or better yet, bring it with you.

---

References

Footnotes

1. Hart, J. (2026). "Read an Anthropic AI safety lead's exit letter: 'The world is in peril.'" Business Insider ↩

2. Anthropic Alignment Science Blog. (2025). "Introducing Anthropic's Safeguards Research Team." Anthropic ↩

3. Dickson, B. (2026). "AI models that simulate internal debate dramatically improve accuracy on complex tasks." VentureBeat ↩

4. Sadhu, S., et al. (2025). "A Multi-Agent Adversarial Framework for Reliable AI Tutoring." Accepted at AAAI 2026. arXiv ↩

5. Sharma, M., et al. (2026). "Who's in Charge? Disempowerment Patterns in Real-World LLM Usage." arXiv ↩