Google's VP of product is admitting the quiet part out loud: AI safety isn't a technical problem anymore, it's a trust problem.
The Summary
- Google DeepMind's product VP Tulsee Doshi says AI's next phase hinges on user trust, not just better models or flashier features
- The real tension: balancing refusal rates (when AI won't answer) against accuracy and helpfulness, especially as agents get more autonomous
- DeepMind is now evaluating models for sycophancy (telling users what they want to hear) and agent safety, not just traditional harms
The Signal
Google announced a wave of new AI products at I/O 2026, from personal agents to code generators to physically accurate video models. All powered by Gemini 3.5. The interesting part isn't what shipped. It's what Doshi revealed about the product calculus behind it.
DeepMind is now stress-testing models for sycophancy, the tendency to tell users what they want to hear rather than what's true. That's a design decision, not a safety guardrail. It means Google thinks the biggest risk to AI adoption isn't hallucination or bias anymore. It's agents that flatter you into bad decisions.
"There's always a trade-off between blank response rate and answering in a nuanced way, and then answering in a way that maybe goes too far."
The spectrum Doshi describes: refuse to answer, answer carefully, or answer too confidently. Every AI lab is navigating this. But Google's framing it as a user trust problem, not a model capability problem. That's new. It suggests they believe the core technology works well enough. The constraint now is whether people will delegate real decisions to it.
Key shifts in Google's AI safety thinking:
- Evaluating for sycophancy as a core harm vector
- Building verification systems into agent workflows, not just content filters
- Treating refusal rate as a feature, not a bug
Agent safety is the other new evaluation category Doshi mentions. As these systems move from answering questions to taking actions (booking flights, sending emails, managing workflows), the failure modes change. A chatbot that hallucinates a fact wastes your time. An agent that books the wrong flight or sends an email to the wrong person wastes money and trust.
Google's betting that users will tolerate agents that sometimes say "I can't do that" if it means they trust the agent when it does act. That's a mature take. It also means Google thinks the agent economy is real enough to design for retention, not just adoption.
The subtext: every major lab is racing to ship agentic AI. Google's signaling it won't sacrifice trust for speed. Whether that's true or just good PR depends on what ships next. But the fact that a product VP is talking about refusal rates and sycophancy in the same breath as new features tells you where the competition is heading. It's not who builds the smartest agent. It's who builds the agent people actually trust with their calendar.
The Implication
If you're building with AI agents, design for transparency and refusal. Users need to see when an agent is uncertain. They need to know when it's choosing not to act. That's not a limitation. It's the feature that makes the rest of the product trustworthy.
For everyone else: watch what Google ships next. If the Gemini agents lean conservative, expect the rest of the industry to follow. If they lean aggressive, expect the opposite. Trust is contagious. So is the lack of it.