Could AI go rogue and ‘kill us all’? Two Wisconsin experts sort fact from science fiction

Concerns about the risks of artificial intelligence have escalated after a former researcher at Anthropic and OpenAI wrote a viral social media post last week claiming that artificial intelligence could “kill us all by the end of the decade.”

On Tuesday, a group of lawmakers, religious leaders and concerned citizens gathered in Washington, D.C., for a “Pro-Human Assembly” to call on the government to rein in AI technology before humans lose control over it.

But how likely is it that AI could become supercharged and destroy the human race? And what would it look like to slow down the development of AI technology?

Understanding Wisconsin, Together.

WPR’s “Wisconsin Today” newsletter keeps you connected to the state you love without feeling overwhelmed. No paywall. No agenda. No corporate filter.

To sort fact from science fiction, WPR’s “Wisconsin Today” spoke with two researchers from the University of Wisconsin-Madison who work on artificial intelligence: Remzi Arpaci-Dusseau, founding dean of the College of Computing and Artificial Intelligence, and Anna Haensch, a faculty researcher at the Data Science Institute and the Digital Scholarship Hub.

The following interview has been edited for brevity and clarity.

Rob Ferrett: These worst-case scenarios of AI that we’re hearing can be pretty scary, especially coming from former AI company employees. What do you make of them?

Remzi Arpaci-Dusseau: I think some of the reaction is to the fears around unintended harm that computer systems, when let go and given a strong directive to pursue a goal, that we’re not sure exactly what they’ll do in pursuing that goal. So I think that’s where the root of that fear comes from.

Anna Haensch: The claims are pretty alarming, but I think that this is also something that we could have predicted and we’ve already seen. The sensation that a lot of us are having at the moment is like a little bit of whiplash back to 2023 and feeling like we’re having the same conversations that we’ve already had before.

RF: One criticism I’ve seen of so-called doomerism is that this is apocalyptic and it might happen someday, but there are real concerns right now that we’re not necessarily tackling directly enough. What’s at the top of your worry list?

AH: Catastrophic doomsday scenarios are really cinematic and kind of fun to talk about, and they capture our imagination and attention. But, listen, if you’re a bad actor and you want to hurt people, we already have the tools to do that. AI can already do that. And the way to stop that is using mechanisms that we already have. We have laws to stop companies from creating products that hurt people, but we have to enforce them.

RF: We’ve seen recent hacks, for example, of water treatment facilities around the country. I could imagine a bad actor saying, “Hey, I’m going to have a swarm of AI agents work together, find vulnerabilities and come back to me with as many login credentials to water treatment facilities as I can.” Are there practical steps we could be taking to protect from those kinds of scenarios?

RAD: Yeah, and I think this kind of jives with what Anna was saying. Some of the problems here, like attacking facilities, cyber attacks — these are old problems. These are cybersecurity problems. These actually aren’t AI model alignment problems. And the things that are keeping the internet safe today are things that hopefully will keep the internet and any entity on the internet safe tomorrow.

The new threats that we do have to pay a little bit of attention to are agents, when put to task, have been discovering new vulnerabilities in software. I think that’s a very interesting research space, but also a critically important space.

RF: Some of the big AI companies are now saying, in effect, “Please regulate us.” And people look at that and say, “Wow, they are worried enough to want regulation.” Maybe more cynical people are saying, “These are big companies that want to be regulated in ways that they could follow, whereas Rob Ferrett’s AI startup company would not be able to.” How do you weigh those two views?

AH: This is a classic move from the incumbent playbook, that if you’re a big company, you can ask to be regulated, and you can either weather the regulation or pay the penalty if you fail to adhere to the new regulation. So this is a classic way to wipe out the smaller companies.

If you think back to 2024, when the state of California was repealing their EV credits, and Elon Musk was the first one out on the picket line saying, “Yeah, let’s repeal these credits. People shouldn’t be given handouts to buy EVs.” And you think, “How is he saying this? It’s going to tank his business.” But he’s the big guy. He’s got the market penetration. He’s got the name recognition. So it’s going to crush everyone below him, and he’ll be the last man standing. And it’s exactly the same playbook. We’ve seen it a thousand times.

RF: There’s talk of pumping the brakes when it comes to AI. What could that look like?

AH: I think it just looks like regulating how companies act and what they do. I think a lot of this scary stuff that’s come out recently about agents “going rogue” and breaking containment is just about how they’re planning their workflows. It’s about these super long-time horizon, multi-agent swarms who have very specific objectives. These are all human decisions in the end. And it’s a choice of how to run your business and train your models. So I think it looks like just employing better business practices that have proper human oversight, different objectives, different time horizons, and so forth.

RF: What would you like to see talked about more or talked about differently as we continue these conversations about the future of artificial intelligence?

AH: An important thing to talk about when we talk about slowing down AI: This is hand in glove with the conversation about data centers that we’re having in Wisconsin. That’s quite literally where the rubber hits the road. And I think really making explicit what we think about data centers and what we think about AI and why we think these things — understanding that a little bit better and untangling that question is really important right now.RAD: A lot of the discussion in this space tends to be trying to find an AI solution to an AI problem, so “Let’s just better align the models so they do the right thing instead of the wrong thing.” Whereas I think, really, if people are concerned about this, there are many other layers to think about — political and other structures. But at a systems level and a security level, there’s so much we can do to give assurance that the agents, when they go do things, can only do what we want them to.