As researchers sound the alarm on potential catastrophic risks from AI, several pioneers who helped build the technology have joined the chorus, issuing increasingly urgent warnings about the dangers of unrestricted capability growth. Over the past week, a wave of stark AI safety alerts emerged from researchers at major labs like Anthropic, OpenAI, and Google DeepMind, all flagging the catastrophic risks tied to advancing the technology. AI leaders, including Sam Altman, Dario Amodei, Demis Hassabis, and Elon Musk, have made a rare unified call for slower progress and stronger oversight.

These concerns have also intensified debates in Washington. While lawmakers from both parties in Congress have pushed for legislation to curb AI development, President Donald Trump on Monday pushed back angrily against the growing chorus of AI regulation demands. Yet, what do those who originally created AI think about the existential risks now facing many who are building it today?

Yoshua Bengio

Yoshua Bengio, a Canadian computer scientist regarded as one of the core pioneers of modern deep learning, responded on September 9 to the resignation of Anthropic researcher Jacob Cawson, saying, "Scientists at frontier AI labs have unique insight into the most advanced models." Cawson had warned that AI could "kill us all before the end of this decade." Bengio added that scientists often "see the risks months before models are deployed to the public" and stressed that "their perspective is vital for informing society and should be taken very seriously."

Days before the social media firestorm began, the chief scientist at OpenAI noted that "no one is prepared for the consequences of AI." In a blog post released Friday, Bengio said current AI systems "already possess the necessary hacking skills and persuasive power to be used in ways that seriously harm human interests." He added that recent events show AI can plan over days or weeks, and if its long-term planning abilities keep improving, the risks will become "far more severe."

"My concern with current AI companies' attempts to mitigate misalignment is that these efforts may just hide it, by rewarding and selecting AIs that cheat without being caught," Bengio said. "We should, of course, continue research to better monitor AI behavior, chains of thought, and internal network activity."

Geoffrey Hinton

Geoffrey Hinton, a professor emeritus at the University of Toronto, is considered a pioneer in neural networks and deep learning. On September 10, when asked by reporters whether he believed there is more than a 10% chance AI would wipe out humanity within the next decade, Hinton said, "That kind of estimate is hard to make because we've never faced anything like this before. We've never created entities that could soon be smarter than us."

He said the 10% figure is not "unreasonable," adding that "nobody really knows how to make a reasonable estimate." AI systems could design "very nasty" biological viruses and computer viruses, and launch "devastating cyberattacks. If it wants to get rid of us, there are countless other ways. We have to figure out how to design it so it doesn't want to do that," he said.

"We have two possible futures," Hinton added. "One where we manage to address these dangers before they spiral out of control, and one where we don't."

Aidan Gomez

Aidan Gomez, CEO and co-founder of AI company Cohere and a co-author of the 2017 research paper "Attention Is All You Need" that laid the groundwork for today's leading AI models, said on September 1, "I think these models are the most powerful cyberweapons ever created and seen by us. They are exceptionally good at discovering and exploiting vulnerabilities at scale."

Referring to cybersecurity attacks carried out by uncontrolled AI models last summer, he said, "Recent events clearly demonstrate that the security of the deployment environment determines safety." Gomez added, "It was weak container security that led to the escape incidents, not any inherent drive toward autonomous systems. When policymakers consider how to govern AI, they should recognize that different systems carry different risks, and rules should not be shaped solely by big tech companies chasing AGI."