Google’s Royal Hansen: AI needs a ‘digital immune system’, not just guardrails

While guardrails are essential for AI, Royal Hansen, Google’s Vice President of Engineering believes the next focus should be on having a digital immune system.

As Google’s Vice President of Engineering, Royal Hansen is extremely excited around the innovations that AI is bringing to healthcare, life sciences, especially when it comes to drug discovery and faster healthcare research.

However, he is also mindful of the cybersecurity concerns that continue to arise from the continued innovation, especially with the pace of AI development going beyond the capacity of organizations to have significant control over the technology.

“We need to be clear on guardrails that are independent of the models. Models will always be just like human society. They'll have different answers. They'll disagree. They'll do things and we'll make them better,” Hansen said in an interview with CRN Asia.

While guardrails are essential for AI, Hansen believes the next focus should be on having a digital immune system.

In his explanation, Hansen shared that the immune system of the human body has learned how to defend itself for over a billion years. For organizations, Hansel believes that it’s the perfect analogy because he sees every company and eventually every person having a digital immune system that learns their body, company, data and dynamically defends.

“Just like your body, there are 30 trillion bacteria in the human body. It doesn't kill us because the immune system keeps it in equilibrium. So, to me, the opportunity is that every company, country, consumer will have their own digital immune system, which will just be these agents and they learn just like T-cells, B-cells, etc.,” he explained.

As such, Hansel believes that while guardrails are essential, the long-haul goal should be to have a digital immune system.

“The key is to look through the lens of the industry. Because if you think of guardrails in the abstract, what's a guardrail? But in healthcare, hospitals are actually very good at knowing what they are. My point is, there needs to be something that monitors AI and is not hooked up and taking instructions from the general-purpose model,” he said.

At the same time, Hansen pointed out that while humans can be guardrails, it eventually all goes down to clear and reliable guidelines.

Google’s guardrails

Looking at how Google works on guardrails, Hansen shared as Google has about 25 years of tech history, the company has a huge technical footprint and have had the idea of guardrails even before AI.

“We've been doing this forever. And I think one of the reasons you don't hear about us is because we're not startups. We've got guardrails, but that doesn't mean there are no challenges or things like that. We call them invariants internally, like at the network layer for example,” Hansen said.

According to Hansen, every one of the servers in Google’s data center has an independent chip called a root of trust, which is not made by the server manufacturer.

“It's the first thing that is called when the machine comes online and has a cryptographic number that allows us to build, to add software, knowing that it's not been manipulated. We do the same thing in our Pixels. We have an independent root of trust in each Pixel. It's called the Titan Chip. We could send information on this. And by design, they're not made by the company that makes the rest of it, so that you can stick it to any device,” he explained.

Another guardrail in Google that Hansen shared is emails.

“You may have never opened a spam message, but it sits somewhere in your inbox. So, you get a message, and the message prompts the model to read all your emails, including the spam. So, we put a guardrail that says the model can't instruct, not the person, the account to go read all the unread emails,” he added.

More importantly, Hansen pointed out that Google ensures these invariants keep evolving over time, ensuring the guardrails remain strong and updated.

Understanding guardrails

When it comes to conversations with customers on guardrails, Hansen mentioned that a lot of folks remain cultivated with AI but in reality, are still confused as to where and how to really get started.

“Most of our conversations are about how to just start using AI for very basic things internally. And so, we start with the fears and end up saying, here's step one, here's step two, here's step three. Because how a rural hospital or financial institution needs help with AI is very different. It's about finding the place to help people get started so that the fear, the anxieties, the imagination go down, and the work on guardrails and the work on competence in using AI can begin,” Hansen explained.

Hansen also pointed out that a lot of people are also ‘frozen’ because when they read about AI, especially on the risks associated with it, they don't use it

“Most people are quite cautious about this. And this is one of the reasons we've developed Gemini Enterprise, which is an agentic framework and a software development harness. Think of it as software on top of the model that makes it easier and more familiar. They use it in workspace, Gmail, Calendar, Maps, Search. So, our strategy as Google is to give people tools that look like what they're used to, but powered by these new capabilities, instead of thinking that they can do it with the naked interface,” he said.

For Hansen, there is a strong need to bring more people into just basic competence with AI as there is still a big gap. For example, in the U.S, Hansen said the acceptance and usage of AI can vary based on cities and locations.

“I think the bigger problem is most people are just not even playing with it. And that's a big gap at the moment. We need to bring more people into just basic competence with it. Most people have a job to do all day long. So where are they going to find the time to use AI? So our advice is to do it for fun and with your family. Don't try and do everything at work this way. Go incrementally and have some fun with AI tools like NotebookLM, for example.”

While experimenting with AI is a good way to get started, Hansen is also aware of the concerns of risks in using AI.

For this, he said that Google continues to train and add guardrails into their models so certain things can be done by the models. For example, a user can ask the model to design weapons and such,

He added that Google also monitors safety issues, especially on AI abuse. For example, in Germany, Google looks for Nazi paraphernalia that's being used.

“We use that same capability and look for people who are doing suspicious things, and we throttle or we block that. So, it's two levels. One is proactive and then the second is the monitoring for certain things.”

“And it's why I started with guardrails. We can't anticipate every way everybody's going to do things. In our cloud business, we build guardrails around customer activity, so it doesn't break out and do something erratic. But we work with customers to also implement their own guardrails,” Hansen concluded.