user trust – Google DeepMind product VP Tulsee Doshi said the company is pushing beyond traditional safety checks to cover issues like sycophancy and agent safety, while building guardrails that balance “blank” refusals with nuanced answers. Her comments came as Google int
The tension shows up in the smallest moments of an AI conversation: whether the model refuses, how it answers, and what it decides to do next.
In mid-2026, Google DeepMind product VP Tulsee Doshi described that balancing act as an engineering problem—and a trust problem. She said DeepMind is evaluating not only “traditional harms. ” but also newer failure modes such as “sycophancy. ” while expanding agent safety and building guard rails around the product experience so the system has “the right verifications in place.”.
Doshi framed the central trade-off as a spectrum. There’s the “blank response rate”—not answering a user because the system doesn’t want to go into a topic. There’s also answering in a “nuanced” way. and the risk that the model goes “too far.” The goal. she said. is finding the right balance across that range.
That trust theme extends beyond policy or refusal strategies. Doshi said she feels “assured by an agent that chooses not to answer a question.” She also described the persona of Google’s models as something the company is actively shaping through feedback—adding that it will “evolve as we get feedback from users” and see what people “resonate with and don’t resonate with.”.
As Google moves into what Doshi called “this more agentic era of Gemini acting with and for you,” she said the shift in persona requires additional clarity work. The company has to help users understand what the agent is for, while ensuring “the right guardrails of the actions that you take.”
Google’s announcement at its I/O developer conference this week included a broad set of new and updated AI products and features: personal AI agents. code generators. search tools. and a new “world model” for generating physically accurate video. Much of the work runs on Google DeepMind-developed Gemini 3.5 models.
Doshi connected that release cycle to how enterprise adoption may unfold. She described the summer of 2026 as a period when people are “figuring out how to wield these tools and how to give themselves that magic.” She said the deeper change in enterprise will come once teams build fluency. pointing to how using even a simple tool can be inefficient at first because people don’t know how to leverage it. Over time, “that’s where you start seeing the culture change.”.
Trust, she said, will also be built through experience—because the downside of failure isn’t abstract. “The last thing I want to do is stake my professional reputation on some AI thing and it doesn’t work out.”
She offered a concrete example from inside the company. Demis Hassabis, the DeepMind founder, asked her for an update on Flash 3.5 metrics. She said she asked Spark—Gemini’s personal agent—to produce a deck pulling metrics and updates from multiple places and delivering it to Hassabis. Afterward, she personally reviewed the numbers “manually” to make sure nothing was incorrect. She said the figures were correct and that the process felt “great. ” but emphasized that doing that a few times is what builds trust that the model can “actually ground effectively.”.
The stakes are also financial. Doshi referenced “an $80 billion CapEx number for this year” and argued that the spending is meant to make agentic help real for everyday users. She said Google’s ethos—rooted in organizing the world’s information and making it universally accessible—can translate into the agent era by helping users “take action on that information in a way that is thoughtful and intentional.”.
She described the promise in personal terms. saying that if the company can bring people like “my mom or my sister” into the new era “in a way that is safe and trustworthy. ” grounded in the principles of search. then the payoff is “real impact.” She also pointed to whimsy and fun in products such as NotebookLM as part of what users get—not only utility. but an experience people want to use.
Doshi pushed back on the idea that the spending story is only about business customers. Asked about the enterprise. she said Google will likely use Gemini to transform businesses and that there will be “literal dollars” tied to ROI. But she argued the bigger value isn’t measured that way. Instead. she said the “magical value” is about what matters to individuals—what a person’s mom or small business owner thinks about. and what kind of access didn’t exist before. The potential, she said, spans consumers, small and medium businesses, and enterprises.
That focus on “quality” and “verification” also echoed her remarks on how Google approaches information for models. In a comparison to players that build pretraining by “crawling the web and grabbing all this information. ” Doshi said Google has been doing the work of search and knowledge-building “for decades. ” with a long-standing emphasis on quality. She said search’s strength isn’t just pulling content, but ranking it well and separating “signal from noise.”.
In the modeling era. Doshi said the key lesson from posttraining and reinforcement learning is that outcomes “come down to the quality of the data” and the ability to verify it—supported by rubrics that clean the data and bring it into the model. She described it as “taking a lot of the bread and butter of what we’ve used in the search context historically” and applying it in new ways.
The trust conversation landed amid other high-stakes AI moves across the industry. This week. Anthropic confirmed that Andrej Karpathy—described as a respected AI researcher and a founding member of Anthropic rival OpenAI—has joined Anthropic. The company said Karpathy started his new job on Monday. and he told people on X that he’s excited to “get back to R&D.”.
Karpathy will join Anthropic’s model pretraining team. working on the formative stage where large language models process vast amounts of data to learn to understand and generate text. He will also form a new group focused on using AI to find more efficient ways of pretraining—potentially using smaller. more curated datasets.
Some observers see the move as a sign Anthropic may be exploring alternatives to the dominant strategy of improving models primarily through scale: more data. more compute. and larger systems. The work could also feed into broader efforts around recursive self-improvement. where AI systems help design and train more capable versions of themselves.
Karpathy said in his announcement that “the next few years at the frontier of LLMs will be especially formative.”
And while Silicon Valley debates strategies for building smarter systems. another thread in the market suggests organizations are already trying to make AI useful—fast. New data from Goldman Sachs and TD Bank painted a bullish picture of small businesses adopting AI quickly. broadly. and at relatively low cost.
Goldman Sachs said it graduated the latest 300-company cohort of its 10. 000 Small Businesses program and surveyed participants about their AI plans. The results, shared exclusively with Fast Company, found that 88% now pay for AI tools. Nearly two-thirds of those respondents spend $100 or less per month on subscriptions. Goldman said the top use cases are marketing and content creation (81%), followed by data analysis (54%), and operations and logistics (47%). Half of respondents said they began using AI within the last year.
TD Bank’s research. described as recently released. suggested AI is helping small business owners expand rather than shrink their workforces. Fully 60% of respondents said adopting AI will increase their workforce size. Nearly seven in ten—69%—said they’re using AI to reduce expenses. up sharply from 39% last year. potentially freeing resources for hiring and training.
The biggest reported benefits over the past year included improved customer service (53%). better fraud and cybersecurity protection (47%). and increased sales leads (42%). Taken together. the data suggests small businesses are treating AI less as a labor replacement tool and more as a growth accelerant.
In the end, Doshi’s message in the midst of all these launches and rival moves comes back to one practical requirement: the next phase of AI depends on whether users can trust what the agent will do—when it speaks, when it stays silent, and when it takes action after verifying the ground beneath it.
Google DeepMind Tulsee Doshi Gemini 3.5 AI agents sycophancy agent safety Flash 3.5 CapEx enterprise AI search quality NotebookLM Anthropic Andrej Karpathy pretraining small business AI adoption Goldman Sachs 10 000 Small Businesses TD Bank research