Google DeepMind VP: AI's Next Phase Hinges on User Trust

Key Takeaways

• Google DeepMind now evaluates models for sycophancy and agent safety, not just traditional harms
• The company is actively developing Gemini's persona for agentic interactions
• Finding the balance between refusing to answer and going too far remains a core challenge

Safety vs. Usefulness: The Constant Tradeoff

Google announced a wave of AI products at its I/O developer conference this week. Personal AI agents, code generators, search tools, and a new world model for generating physically accurate video. Most of it runs on the company's latest Gemini 3.5 models, developed inside Google DeepMind.

Tulsee Doshi, DeepMind's product VP, spoke with Fast Company about the thinking behind these developments. Her perspective reveals a company wrestling with questions that have no clean answers.

The core tension: how much should an AI refuse to do? Doshi described a spectrum that the team navigates constantly. On one end, a model that declines too often becomes useless. On the other, one that answers everything can cause real harm.

“There's always a trade-off between blank response rate—not responding to a user because you maybe don't want to answer about a particular topic—[and] answering in a nuanced way, and then answering in a way that maybe goes too far. That's always the spectrum that we're trying to find the right balance on.”

— Tulsee Doshi, Google DeepMind Product VP

New Safety Criteria: Sycophancy and Agent Behavior

DeepMind's evaluation criteria have expanded. Traditional harms remain a concern, but the team now tests for subtler problems. Sycophancy is one. This is when an AI agrees with users too readily, telling them what they want to hear instead of what's accurate.

Agent safety is another new focus. As AI moves from answering questions to taking actions, the stakes change. A chatbot that gives bad advice is one thing. An agent that makes purchases, sends emails, or modifies files based on flawed reasoning is another.

Doshi noted that guardrails need to be built around the product experience itself. Verification systems must be in place before agents can act on users' behalf.

The Persona Question

AI assistants project personalities. Some are formal, some casual. Some defer to users, others push back. Google is still figuring out what Gemini's persona should be.

“That persona is going to evolve as we get feedback from users, as we see what folks resonate with and don't resonate with. Also, as we enter this more agentic era of Gemini acting with and for you, there's a switch in persona that you also need to think through.”

— Tulsee Doshi, Google DeepMind Product VP

This is a design challenge with real consequences. When an AI acts on your behalf, you need to trust its judgment. That trust depends partly on how the AI communicates what it's doing and why. Doshi framed this as thinking through how Gemini helps users clarify their intentions before taking action.

The Case for Refusal

Doshi offered a personal view that cuts against the industry's push for ever-more-capable models. She said she feels assured by an agent that chooses not to answer a question.

This is notable coming from a product leader. The default instinct in tech is to maximize capability. Users complain when assistants refuse requests. But Doshi's comment suggests Google sees refusal as a feature, not a bug. An AI that knows its limits is one you can rely on.

Enterprise Adoption Still Taking Shape

Fast Company asked Doshi about AI transformation in enterprise settings. The interview ended before she could fully respond, but the question itself points to the gap between demo and deployment. Google can announce impressive capabilities. Getting businesses to trust AI with real work is another matter.

That trust gap is exactly what Doshi's work addresses. Safety evaluations, persona design, and refusal rates all shape whether organizations will hand AI tools to employees. Get those wrong, and the most capable model in the world sits unused.

Frequently Asked Questions

What is AI sycophancy and why does Google test for it?

Sycophancy is when an AI agrees with users too readily, reinforcing their beliefs instead of providing accurate information. Google tests for this because it undermines the AI's usefulness as a reliable assistant.

What is agent safety in AI?

Agent safety refers to ensuring AI systems that take actions on users' behalf do so correctly and with appropriate guardrails. As AI moves from answering questions to making purchases or sending emails, the potential for harm increases.

What are the Gemini 3.5 models?

Gemini 3.5 is Google DeepMind's latest generation of AI models, powering the company's new personal agents, code generators, and search tools announced at I/O 2024.

Why would an AI choosing not to answer be a good thing?

An AI that refuses to answer when uncertain or when a request is problematic demonstrates self-awareness about its limits. This builds user trust by preventing confidently wrong or harmful outputs.

Source: Fast Company / Mark Sullivan