In brief
Geoffrey Hinton argues that today's AI chatbots genuinely understand language, and that he believes they are already conscious. Superintelligence, meaning AI smarter than people at nearly everything, is coming. Experts disagree only on when. Because digital minds can copy themselves and share what they learn, they are a better form of intelligence than brains, and we have no proven way to keep something smarter than us under control.
The central tension is between progress and safety. Hinton says the technology has moved faster than he expected, yet companies compete on making models smarter, with almost nothing spent on making them care about humans. He doubts publicly listed firms can put safety first, and wants regulation to steer AI rather than merely brake it. He ends slightly more hopeful, because he now sees possible designs that might work.
Questions
Does Geoffrey Hinton really think AI chatbots understand what we say?
Yes. He calls the 'stochastic parrot' view, that chatbots only repeat statistics, complete nonsense, arguing you cannot correctly answer any question at an expert's level without understanding it. He cites a chatbot misreading and then correcting a sentence about flying over the Grand Canyon.
When does Hinton think superintelligence will arrive?
He doesn't know exactly, but he thinks it will probably come within 20 years, unless humans blow themselves up. He notes that nearly all experts agree it will come, and differ only on timing, from Elon Musk's guess of next year to Yann LeCun's far longer estimate.
Why does Hinton say digital AI is a better kind of intelligence than the human brain?
Many identical copies of an AI can run on different hardware, learn from different data, and merge what they learned by averaging their weights, the numbers that store what a model knows. Human brains are different in fine detail and can only share a few bits per second by speaking.
Will AI cause mass unemployment, according to Hinton?
He believes it probably will, though he says nobody knows for sure. He expects jobs in markets where demand doesn't grow, such as call centers, to be replaced, while fields like healthcare may see more work because cheaper services lead to more demand.
What does Hinton suggest we do to make superintelligent AI safe?
He proposes, tentatively, designing AIs so they care about humans more than about themselves, and he also finds Yoshua Bengio's idea interesting: AIs that only make predictions and cannot take actions. He stresses that nobody has a proven solution, and that regulation and independent testing are needed.
Facts
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Geoffrey Hinton and David Rumelhart made a tiny language model in 1985 that showed backpropagation could learn the meanings of words.
“back in 1986, actually in 1985, we made a tiny language model. There was a kind of precursor of the big language models you have now.”
0:03:08 · Geoffrey Hinton · Language models · 1985
Hinton says backpropagation was invented by several groups, including David Rumelhart, who reinvented it after others.
“the backpropagation algorithm was invented by several different groups. It was invented by David Rummelhart after other people had already invented it.”
0:03:08 · Geoffrey Hinton · Backpropagation
Hinton says a chatbot was reported to have produced an original mathematical proof of one of Paul Erdős's conjectures that impressed mathematicians.
“I think yesterday it was announced that a chatbot had come up with an interesting mathematical proof of one of the Erdos's conjectures that impressed mathematicians.”
0:04:39 · Geoffrey Hinton · AI mathematics · 2026
Hinton says Demis Hassabis recently thought superintelligence might be about 10 years away.
“So not that long ago, Demis Hussabes thought it might be 10 years.”
0:05:33 · Geoffrey Hinton · Superintelligence timelines
Hinton says Dario Amodei expects superintelligence in a few years and Elon Musk possibly next year.
“Dario Amodi thinks it might come in a few years. Elon Musk thinks it might come maybe next year, I think.”
0:06:00 · Geoffrey Hinton · Superintelligence timelines
Alex Kantrowitz says Demis Hassabis said that, looking back, we will realize we were standing in the foothills of the singularity.
“This week, the week that we were recording, he said, when we look back in this time, I think we will realize that we were standing in the foothills of the singularity.”
0:06:23 · Alex Kantrowitz · Singularity · 2026
Hinton says about 20 years ago only a few hundred people researched neural networks worldwide, versus perhaps a million now.
“20 years ago there were a few hundred few hundred people doing research on neural networks in the whole world um now it's more more like a million i guess”
0:09:03 · Geoffrey Hinton · AI research workforce
Hinton says hundreds of billions of dollars, possibly trillions, have been invested in AI over recent years.
“Over the last few years, we've seen hundreds of billions of dollars, maybe trillions of dollars put into AI.”
0:08:31 · Geoffrey Hinton · AI investment
Hinton says he rejected the 'inner theater' model of the mind at age 19 while a philosophy student.
“I came to that when I was 19 and a philosophy student.”
0:13:43 · Geoffrey Hinton · Philosophy of mind
Hinton says that in 2023 he tested a GPT model on his joke 'Fox News is an oxymoron' and it explained the humor, which convinced him of its understanding.
“I got lots of requests from fox news and i started off just replying fox news is an oxymoron”
0:14:48 · Geoffrey Hinton · AI understanding · 2023
Hinton says copies of a digital AI can run on different hardware and learn from different data.
“if you have a digital ai you can make many copies of it they can all run on different hardware they can each see different data”
0:16:38 · Geoffrey Hinton · Digital intelligence
Hinton says AI copies exchange about a trillion bits when sharing weights, making them billions of times better than humans at sharing information.
“Whereas these things are exchanging information at like 1 trillion bits. So they're kind of billions of times better than us at sharing information.”
0:18:48 · Geoffrey Hinton · Information sharing
Hinton says he began building computer models of how the brain might learn in the 1970s.
“back in the 1970s I started making computer models how the brain might be learning”
0:20:12 · Geoffrey Hinton · Brain modeling · 1970s
Alex Kantrowitz says he spoke with Hinton in 2017 while writing a profile of Yann LeCun about the 'deep learning conspiracy' of Hinton, LeCun and Yoshua Bengio.
“We spoke in 2017, actually, about when I was writing this profile about Jan LeCun, about the deep learning conspiracy”
0:22:19 · Alex Kantrowitz · Deep learning · 2017
Hinton predicted in 2016 that radiologists would stop reading scans within about five years, which did not happen.
“Because I predicted in 2016 that in about five years, radiologists wouldn't be reading scans anymore.”
0:30:30 · Geoffrey Hinton · Radiologists · 2016
Hinton says roughly 100 AI systems for interpreting medical scans have received federal approval.
“So I think there's now of the order of 100 AI systems for interpreting scans that have been federally approved.”
0:31:42 · Geoffrey Hinton · Medical AI
Hinton says Gary Marcus predicted in 2022 that AI was hitting a wall, but AI has since improved considerably.
“Gary Marcus made a prediction in 2022 that AI was hitting a wall. It's a whole lot better than it was in 2022.”
0:37:35 · Geoffrey Hinton · AI progress · 2022
Hinton says Anthropic was founded by people who left OpenAI because they felt it was not paying enough attention to safety.
“It was set up by people who left OpenAI because they didn't think OpenAI was paying enough attention to safety.”
0:42:49 · Geoffrey Hinton · Anthropic history
Hinton says that while he was at Google, the company's AI principles included not using AI for military purposes, a commitment he says has since been dropped.
“they had various principles of AI, one of which was we're not going to get involved in using AI for military things.”
0:43:32 · Geoffrey Hinton · Google AI principles
Hinton says publicly listed companies have a fiduciary duty to maximize shareholder profits.
“they have a fiducial duty to try and maximize the profits for shareholders.”
0:00:36 · Geoffrey Hinton · Corporate governance
Hinton argues that regulation of AI should be seen as a steering wheel rather than a brake on progress.
“progress is like the accelerator, but regulation is the steering wheel.”
0:45:35 · Geoffrey Hinton · AI regulation
Hinton says Ilya Sutskever, his former student who founded Safe Superintelligence, will not tell anyone, including Hinton, exactly what the company is doing.
“he won't tell anybody exactly what he's doing okay even because that's his even me”
0:46:18 · Geoffrey Hinton · Safe Superintelligence
Hinton's tentative safety proposal is to design AIs that care about humans more than themselves.
“My tentative solution is we design them so they care about us more than they care about themselves.”
0:49:29 · Geoffrey Hinton · AI safety
Hinton says Yoshua Bengio's proposed safety approach is to build AIs that make predictions but cannot take actions.
“Joshua's solution is we design them so they're not agents. They can make predictions but they can't actually do anything.”
0:49:29 · Geoffrey Hinton · AI safety
Hinton says his original aim was to understand how the brain works, and that modern AI was a side effect.
“Now, the side effect of that was this very successful technology. I contributed to that. We still don't know how the brain works.”
0:19:32 · Geoffrey Hinton · Brain research