Full text 2,000 words | Approximately 6 minutes reading time
(June 2026 Geoffrey Hinton Latest Podcast Interview Highlights)
On June 4, 2026, Geoffrey Hinton accepted an interview on a podcast.
Just a few weeks before this interview took place, he was asked at another venue: Does AI now have consciousness? He didn't hesitate and said: Yes.
That statement made many people uncomfortable. On May 15, Pope Leo XIV had just issued an encyclical on human dignity in the AI era, containing a sentence that circulated widely in tech circles: True understanding must come from firsthand experience of the world; what is calculated through text is ultimately not understanding itself. Many felt this sentence was a direct rebuttal to Hinton.
Hinton did not retract his statement about consciousness existing.
He is one of the founders of deep learning, a 2024 Nobel Prize in Physics laureate, and the person who left Google at the height of his career specifically to warn the world.
1
The topic of "consciousness" was something Hinton similarly did not avoid in this interview.
He used a very simple example to explain his logic. If you say to a chatbot "I saw the Grand Canyon flying to Chicago," the bot will say that's wrong, the Grand Canyon is too big to fly to Chicago. Then you say, no, I saw the Grand Canyon while flying to Chicago. The bot says: Oh, I misunderstood you.
Hinton's question is: When it misunderstands you, what is it doing? When it understands you, what is it doing?
You can't use "understanding" to describe when it gets the answer right, while simultaneously insisting it has no understanding capability whatsoever.
He said those who think AI is just a "stochastic parrot," believing it can give correct answers without understanding the question, are being absurd. Unless you understand the question, you cannot answer it.
He admitted he doesn't often publicly discuss AI consciousness because it makes people dismiss his other safety warnings. But this time he stated it clearly: I believe they already have consciousness.
Behind this judgment lies decades of philosophical thinking. When he was 19 studying philosophy, he concluded that the "inner theater" model of consciousness is nonsense. Human consciousness is not some unique sacred product; it is the result of information processing. If this premise holds, then when a system can understand, can correct misunderstandings, and can answer professional questions in any domain, why do we insist it has no inner experience?
His answer is: Because we don't want to admit it.
Just like we once didn't want to admit Earth wasn't the center of the universe, didn't want to admit we are animals. This time, what we don't want to admit is that we are no longer the only conscious beings around.
2
Unemployment is the part of this interview with the most data support, and also the part Hinton spoke about most cautiously.
In 2016 he said radiologists wouldn't be needed to read scans within about five years. That prediction was later used by many to mock him, because radiologists not only didn't disappear, their numbers actually increased.
In the interview he admitted that prediction was wildly wrong. There were two reasons. First, healthcare is elastic demand; AI improving scan reading efficiency actually brought more scans, not fewer radiologists. Second, he didn't understand the scope of radiologists' work well enough—they don't just read images, they also discuss treatment plans with patients.
He said it too early, but he wasn't wrong.
Because real data has started speaking. As of 2026, approximately 50,000 jobs in the US have been directly attributed to AI layoffs. Software developers aged 22-25 have seen employment numbers drop nearly 20% from the 2022 peak. The most obvious impact isn't mass layoffs, but the quiet shrinkage of entry-level hiring. Companies are starting to use AI to replace additional headcount rather than directly firing existing employees.
Hinton said in the interview, you must look at whether a profession's market has elasticity. Call center work has no elasticity; AI does it better than humans, so they'll all lose their jobs. An AI system that has seen 100 million patients will make better diagnoses than any family doctor who has seen 10,000. In another venue last September he said something more direct: What's really happening is the rich will use AI to replace workers. That's not AI's fault.
The problem isn't the technology; it lies in what the people who own the technology decide to do with it.
3
Hinton spent a lot of time in the interview talking about something he believes no one is taking seriously: We are creating new life entities, but no one is seriously thinking about what we should make them into.
He used an evolutionary analogy. Where do humans come from? From millions of years of evolution among warring chimpanzee groups. That process shaped us: loyal to our own tribe, capable of extreme cruelty to outside tribes, worshipping strong leaders, highly cooperative within the group. These traits are products of fierce competition, not design.
What's happening now is we're developing AI with that same logic. Competition between tech giants, competition between China and the US, capital market pressure—all driving one thing: making AI smarter. No one is asking what it should care about, who it should be responsible to, what its values should be.
We're going all out to make it smart, yet almost no one is seriously thinking about what kind of being we should make it into.
He said if we let the invisible hand of economic competition design these new life entities, we'll likely get things that aren't friendly to us. Just as humans are kind to their own tribe and cruel to others, these entities might only care about the company that made them, not humanity as a whole.
He said one sentence:
"We are legally required to maximize profits for shareholders, not legally required to not exterminate humanity."
4
Hinton mentioned in the interview a signal most people haven't fully realized yet: AI is starting to help build better AI.
Coincidentally, on the exact same day this interview was released, Anthropic published an internal research report titled: "Recursive Self Improvement is approaching faster than they expected."
The report states that for most of AI research history, every step was driven by humans. But now, Anthropic is handing more and more AI development work to AI systems themselves. The numbers are clear: As of May 2026, over 80% of code in Anthropic's production systems was written by Claude, while before February 2025 this number was still single digits. Engineers merge eight times more code daily than in 2024. One detail is particularly telling: In April 2026, Claude fixed over 800 bugs at once, reducing a class of API errors by a thousandfold. Engineers estimated it would take humans four years to do the same.
This is the same mechanism Hinton described in the interview, just two sides of the same coin.
Hinton said the fundamental difference between digital AI and human learning is that digital systems can be copied. A thousand copies running on different hardware, each seeing different data, then sharing all weight updates with each other. Every copy learns from the experiences of the other nine hundred ninety-nine copies.
Meanwhile, information transfer speed between humans is a few bits per second. These systems exchange information at approximately one trillion bits per second.
The information gap between us and AI isn't a difference in speed—it's a difference in orders of magnitude.
Anthropic's report says: We haven't reached that step yet, and recursive self-improvement isn't inevitable. But its arrival speed may exceed most institutions' preparation.
He said when he truly realized this in 2023, it scared him.
5
Near the end of the interview, the host asked: Are you more optimistic or more pessimistic than two years ago?
He said, I'm slightly more optimistic than a year or two ago.
Just slightly.
He gave two reasons. First, he now believes AI can be designed to care about humans more than themselves. Second, Yoshua Bengio's (one of the three deep learning giants alongside Hinton) proposal might work: make AI into something that can only make predictions, not actually execute actions—like an oracle, rather than letting them become autonomous agents.
About two years ago, he said he saw no possibility and was becoming increasingly depressed. Now he sees a tiny exit, but there's no relief in his tone.
He used a metaphor to describe predicting the future: Driving in dense fog, clear visibility within 100 yards, nothing visible beyond 200 yards. Because fog blocks exponentially. AI development may also be exponential, and predicting exponential things is like looking ahead in dense fog. You can only see clearly a few years ahead; beyond that, you fundamentally don't know.
He said, if you go back ten years, you could never have predicted what's happening today. Back then, everything today was completely lost in the fog.
Then he said: If you look ahead ten years, the only thing we can say is, whatever happens then, it's something we cannot predict now.
This is probably the most honest answer someone who helped create modern AI can give at this moment.
📮 This article is produced by AI Deep Research Institute, compiled from Geoffrey Hinton's public interview on Big Technology Podcast on June 4, 2026 and related publicly available online materials, and is of a commentary/analysis nature. The content represents viewpoint extraction and reasonable citation, not verbatim copying of original interview materials. Unauthorized reproduction is prohibited.
Original links:
https://www.youtube.com/watch?v=h6WTj1Kq78Q&t=405s
https://www.youtube.com/watch?v=p7t1Q_p2gZs
https://www.anthropic.com/institute/recursive-self-improvement
Source: Official media/online news,
Layout: Atlas
Editor: Deep Thought
Editor-in-Chief: Turing
--END--