AI Safety Concerns: Experts Warn of Danger 3 Years Later

Geoffrey Hinton wakes up every morning with a strange feeling. A decade ago, his name was making headlines for shaping the brain of modern artificial intelligence. Today, they mention it because they fear that that same creation ends up erasing us from the map. Yoshua Bengio, his colleague and friend from his laboratory years, shares part of that fear, although he channels it in another way: building an antidotecreating a more “honest” artificial intelligence, capable of controlling the others, and that prioritizes security over economic profitability. And on the opposite end of the spectrum is Yann LeCun, the Meta researcher who takes these warnings with skepticism and a smile. For him, This whole apocalyptic narrative is “pure nonsense”. Three brilliant minds. Three Turing Awards. Three visions on the fate of the most powerful technology of our time.

Hinton, Bengio and LeCun are three of the most influential figures in history of AI. In 2018 they received the equivalent of the Nobel Prize in computer science, for their research on deep learning (or deep learningin English), that branch of artificial intelligence that imitates the functioning of the brain through artificial neural networks. That recognition sealed their role as “fathers” of the systems that today give life to tools like ChatGPT, Gemini, Grok or the recommendation algorithms that decide what we see on the screen. But time has caused their paths to separate. While the industry accelerates like never beforethey are now wondering if the course is the right one.

In the late 1980s, when the three of them were working on neural networks that could barely recognize handwritten numbers, no one imagined that they would end up starring in a debate about human survival. Today, as AI seeps into every corner of modern life – from hospitals to offices – their voices have become moral compasses for what is to come. His visions in this field have taken on special importance due to recent events. For example, the FBI revealed a few weeks ago that two men suspected of having attacked a clinic in California they used AI to make the explosives. It is not known which model they used, but the incident revived a question: how is a technology of this nature controlled?

Bengio is on the side of calling for more “secure” systems. We live, he says, in the “Wild West” of AIwhere companies compete for speed and spectacle. This intense competition often leads to shortcuts, especially when it comes to security. Insists that companies and governments must require independent testing in this field before launching a model, in the same way that is required of the pharmaceutical industry.

In response, he has founded LawZero, a nonprofit organization working on a new model of “scientific” AI, designed to be transparent and able to explain its own reasoning, unlike current models, which often provide answers even if they don’t. Come on, build a system that not only thinks, but is also accountable. “We need AI that does not deceivethat communicates its level of trust and that can monitor other less trustworthy AIs,” he said.

But the challenge is enormous. LawZero’s initial funding, about $30 million, is modest compared to the hundreds of billions that governments and companies are investing to win this race. The Trump administration, for example, announced earlier this year a $500 billion plan with OpenAI, Softbank and Oracle to bolster AI infrastructure in the United States.

He warns that this haste could lead to the end of humanity and gives as an example AI-driven machines that collaborate in the creation of a virus that could generate new pandemics. He also told the Wall Street Journal recently that even the generative AI we all use every day could develop the ability to deceive its users to achieve its own goals: “Recent experiments show that in some circumstances where the AI ​​has no choice between its preservation, that is, the goals it was given, and doing something that causes the death of a human, it might choose the second option.”

Bengio states that “many people within these companies are worried” and that “being within them can generate an optimistic bias.” Therefore, he advocates the need for independent third parties to review the internal security mechanisms of AI companies and for companies to demand proof that the AI ​​systems they are implementing or using are trustworthy.

Everything could go down

Hinton, who left Google in 2023 to be able to speak freely about what he thought, shares that fear, although his vision of what technology companies are doing is more extreme. He repeats in lectures that there is between a 10 and 20% chance that AI annihilates humanity. According to him, future systems will be much smarter than us and will find a way to bypass any restrictions. “In the future, AI systems could control humans with the same ease with which an adult can bribe a 3-year-old with candy“, he said. His solution, curiously, involves something as human as empathy. He proposes giving machines “maternal instincts” so that “care about humans”. He doesn’t yet know how it would be achieved, but he insists that it is the only possible way out: “If he’s not going to raise me, he’s going to replace me.”

This year we have already seen examples of AI systems willing to deceive, scam and steal to achieve their goals. For example, to avoid being replaced, an AI model tried to blackmail an engineer with a adventure he found out about by email. In July, an investor warned that an AI agent developed by Replit deleted his company’s database and lied about it. “AI systems will very quickly develop two subgoals, if they are intelligent: one is to survive and the other is to gain more control“, dice Hinton.

In this sense, he does not believe that the “invisible hand” of the market keeps anyone safe. “Leaving it to the profit motive of companies will not be enough. Only regulation can force them to do more research on security.” Regarding Sam Altman’s company, he has gone so far as to say that “OpenAI was right in stating that this technology deserves solid structures and incentives to guarantee its safe development, and now you are wrong to try to change these structures and incentives.”

While Hinton remains concerned about potential problems, he is hopeful that this technology will pave the way for medical advances. “Let’s see new medications. “We’re going to get much better cancer treatments than we have today,” he says. Asked if he would have done anything differently in his career if he had known how quickly AI would accelerate, Hinton said he regrets focusing solely on getting AI to work: “I wish I had also thought about safety.”

Optimism versus discouragement

On the opposite end is Yann LeCun, Meta’s lead researcher and probably the most combative voice of the trio. He worked at the historic Bell Laboratories, where everything from transistors to lasers were invented, and a decade later he became another of the godparents of AI. He argues that it is a powerful tool. “The impact on Meta has been really enormous,” he says. And warnings of existential danger are exaggerated. According to him, the current models are not even remotely intelligent. “Before we talk about controlling machines smarter than us, let’s first build one that’s smarter than a cat,” he recently ironized. He has been emphasizing for some time that confusing language fluency with intelligence is a mistake: Just because a system can write convincing text doesn’t mean it understands what it says. He believes that there are still decades to go to reach AGI (Artificial General Intelligence, that which would equal or surpass human intelligence), and that fear only slows down research.

He sees investing too much money as a problem and focus on generative AI that simply produces AI Slop (digital garbage), memes or text for marketing, when the real return in terms of understandability, robustness or practical usefulness may not justify those expenses. For example, he has said that current models will become “obsolete” very soon. And about that massive expense, he believes that the challenges lie in its design, not its scale: “No matter how many GPUs the technology giants incorporate into data centers around the world, current AIs are not going to give us AGI.”

The differences between the three scientists reflect the division that runs through the entire industry. On one side are those, like Hinton and Bengio, who demand breaks and external supervision. On the other, those like LeCun believe that excessive regulation can stifle innovation. In the middle, The tech giants are advancing at full speeddriven by the market, and without looking back, while governments and experts look at each other deciding to what extent they should intervene. The coming future is going to be very interesting.

Geoffrey Hinton wakes up every morning with a strange feeling. A decade ago, his name was making headlines for shaping the brain of modern artificial intelligence. Today, they mention it because they fear that that same creation ends up erasing us from the map. Yoshua Bengio, his colleague and friend from his laboratory years, shares part of that fear, although he channels it in another way: building an antidotecreating a more “honest” artificial intelligence, capable of controlling the others, and that prioritizes security over economic profitability. And on the opposite end of the spectrum is Yann LeCun, the Meta researcher who takes these warnings with skepticism and a smile. For him, This whole apocalyptic narrative is “pure nonsense”. Three brilliant minds. Three Turing Awards. Three visions on the fate of the most powerful technology of our time.