Google DeepMind VP on balancing AI safety with product quality
Google announced new AI products at its I/O conference this week, including personal agents, code generators, and video generation tools. Most run on Gemini 3.5 models developed inside Google DeepMind. The company's product VP discussed how the team approaches safety and design as AI agents take on more autonomous roles.
Safety now means more than preventing harm
Tulsee Doshi, Google DeepMind's product VP, said the safety evaluation process has expanded beyond traditional risks. Teams now assess for sycophancy-when models simply agree with users-and agent safety, the risks that arise when AI systems take actions on behalf of users.
The work involves building guardrails into the product experience itself, not just the model. Doshi described a constant tension between refusing to answer questions and providing nuanced responses that might overstep.
"Personally, I feel assured by an agent that chooses not to answer a question," Doshi said.
Persona design is still evolving
Google's models project a specific persona-the way they communicate and interact. Doshi said this remains an active area of development and will shift based on user feedback.
The shift toward agentic AI introduces new persona considerations. As Gemini moves from answering questions to acting on behalf of users, the company must rethink how the model communicates and what guardrails it needs around the actions it takes.
Enterprise adoption timeline unclear
Doshi did not provide specifics on how quickly enterprises will adopt these models, though he acknowledged the question. The pace will likely depend on how organizations solve their own safety and integration challenges.
For product teams working with generative AI and LLM technology, these considerations around safety, persona, and agent behavior are becoming central to development decisions. Learn more about AI for Product Development to understand how these principles apply across industries.
Your membership also unlocks:
