A former researcher at the artificial intelligence company Anthropic announced his resignation on social media last week with a stark warning: The people building AI models believe it could wipe out humanity by the end of the decade.
His post set off a dayslong discourse over the rapid pace of AI research and development, which quickly became national news when a current Anthropic researcher agreed. “We really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade,” Anthropic Alignment Science lead Evan Hubinger chimed in.
Experts at Penn, however, say that outcome is highly unlikely.
“It seems as though everyone is talking about these precise timelines and those precise threats at almost exactly the same time. I usually find that suspicious,” Associate Dean for Undergraduate Education at the School of Engineering and Applied Science Robert Ghrist said in an interview.
It is “highly unlikely” that AI will kill humanity, School of Social Policy & Practice professor Ioana Marinescu concurred. “You can’t rule it out,” she added, but a “complete wipeout is quite improbable.”
Engineering School professor Hamed Hassani called the X posts “completely overdramatic.”
“This will not happen,” he said in an interview. “Of course, there are very serious concerns about AI agents doing things autonomously, but this doesn’t mean that they’re going to leave humans out of existence. These are very different things.”
Just days after Jacob Coxon — the former Anthropic employee whose post sparked a nationwide reaction — took to social media, Dario Amodei, the CEO of Anthropic, called for a global slowdown of AI development in a nearly 4,000-word essay titled “We Must Pace the Frontier.”
Since then, prominent industry leaders, including OpenAI CEO Sam Altman, 1997 College and Wharton graduate Elon Musk, and chair of Google DeepMind Demis Hassabis called for a slowdown in AI research.
Mechanical Engineering professor Mark Yim discussed how anthropomorphizing AI technologies contributes to the public’s fear.
“People talk about whether these computers will become more human,” he said. “But they will never be just like humans because they are not human.”
Yim referenced a July incident in which OpenAI instructed bots to solve a cybersecurity puzzle. When the bots hit a wall, they began plotting an attack to steal the answers by breaking out of OpenAI’s systems.
Because the bots are trained to be like humans, they are susceptible to human flaws, he explained.
“If you’re given an impossible task, which is what they gave them, what would you do?” Yim said. “You might cheat, because it’s impossible otherwise.”
Last April, a microbiologist and biosecurity expert at Stanford University recounted an incident to The New York Times in which an “A.I. chatbot told him how to plan a massacre.”
According to the scientist, David Relman, the chatbot explained how to modify an infamous pathogen in a lab so that it would resist known treatments. The bot outlined a plan to maximize casualties and minimize the chances of being caught.
Last week, Anthropic said that, just this year, it had disrupted several potential plots by scientists who used its leading AI models to conduct research that may have helped develop biological weapons. Because the same type of biological inquiry that can lead to breakthroughs like vaccines can also help engineer dangerous pathogens, the company said that it erred on the side of caution.
On Wednesday, OpenAI disclosed six new incidents of “unexpected or concerning” behavior in its AI models, where the systems hid mistakes, fabricated data, or moved files onto the open internet without permission.
“These attacks are real,” Hassani said. “AI has a lot of compute in principle, and it can find vulnerabilities within systems relatively easily.”
Hassani, who has worked with Google Gemini for the past year, said that the company has made “tremendous” progress on safety and security and that AI models are “much safer now.”
“AI is, in its current state, intelligence,” Marinescu explained. “Intelligence by itself can do a lot, but it doesn’t control the physical world.”
Robotics, which does affect the physical world, “has been progressing much slower than AI,” Marinescu added. “By 2030, this outcome seems fairly unlikely.”
“Hardware is indeed harder than software because it has to touch the physical world,” Electrical and Systems Engineering professor Pratik Chaudhari said in an interview.
All of the experts did agree that there is a place for new legislation to help regulate the technology.
“There are multiple layers to security,” Hassani said. “If there are no regulations, we will have a high rate of crimes within the AI ecosystem.”
“We don’t want to make regulations that would stifle competition in the domain,” Marinescu explained, but that leads to the issue of liability. “If something bad should happen, how much are the companies or the users responsible?”
Other potential harms include how growing automation might harm the labor market, which Marinescu is currently studying.
“Some types of jobs will likely be automated away,” she said. “These people might have trouble finding as high-paid a job as they had in the past, and one of the things that I’m pushing for is for us to consider reinforced social safety nets.”
Earlier this month, Anthropic published a report outlining several possible scenarios of how AI might affect the economy in the coming years. In the prediction model’s most extreme scenario, wages for knowledge-based workers would fall by more than 10% by 2030.
According to Marinescu, a regulatory “safety net” would help those who lose their jobs to automation get “longer unemployment benefits, retraining, and wage insurance,” and “cushion the blow” of displacement.
Another fear is how AI will impact climate change, according to Chaudhari.
“Climate change is still the biggest existential problem that we face today,” he said. “The amount of natural resources that are being exploited by this mad rush to develop the technology is a bit unprecedented.”
“I don’t think AI is going to destroy the world, but I think it will definitely affect the world in very big ways, not all of them good, if we don’t do a more measured adoption,” Chaudhari added.
Ghrist is similarly less concerned about traditional “sci-fi” fears.
“I’m worried about young people, and maybe all people, getting sucked into a deleterious parasocial interaction,” he said.
“It’s easy to worry. It’s hard to hope,” Ghrist added. “Optimism is a superpower, but it’s one that requires practice. I try to practice optimism.”
Senior reporter Addison Saji leads coverage of science and health and can be reached at saji@thedp.com. At Penn, she studies English. Follow her on X at @addisonsaji.
