NextFin News - Professor Geoffrey Hinton sat down with host Nayeema (Naima) Raza for a conversation recorded at StartWell Studios in Toronto and released as part of the Smart Girl Dumb Questions series. The episode, presented across multiple parts, follows Hinton through questions about whether artificial intelligences can be called alive, how minds differ from machines, what immediate dangers are already present, and what longer‑term scenarios might look like.
The interview is notable for its measured tone: Hinton insists on precise language where possible, distinguishes near‑term misuse from far‑term existential risk, and repeatedly emphasizes uncertainty while explaining why some responses should nevertheless be urgent.
Is AI "alive" — and what would that mean?
Hinton opens by challenging ordinary categories. He says our concept of "alive" is historically contingent and points out that existing definitions — from "live" electrical wires to biological life — do not map cleanly onto digital intelligences. He suggests a new category: "an AI is alive like an AI." As he puts it in the interview, They're alive like none of those things. They're alive like an AI is alive.
He draws attention to properties such as digital replicability — the ability to make exact copies — and asks what that does to our ideas of personal identity. To illustrate the philosophical difficulty he asks, rhetorically, whether an exact copy of you that will be tortured is the same person, and observes, That means you think somehow you is different from that you. The concepts break down.
Consciousness, sentience and subjective experience
Hinton distinguishes terms people often use interchangeably. He explains his working distinctions: sentience as the capacity to feel and report emotion («this was painful», «this made me happy»), subjective experience as a subset of sentience, and consciousness as awareness that allows learning and problem solving about why feelings arose. He presses against confident denials of machine sentience: People have great confidence that these things aren't sentient and great difficulty saying what they mean by sentient, which seems like a strange combination.
He argues that because we have never built minds before, AI may fundamentally reshape our understanding of mind.
Memory, confabulation and the difference between humans and machines
On memory Hinton draws a sharp contrast between computer files and human recollection. Where a file is stored unchanged, human memory is reconstructive: experiences change connection strengths and are regenerated in ways that can introduce error. He urges psychologists' vocabulary for human errors — confabulating
— rather than borrowing sensory terms like "hallucinating" when discussing language models. He describes human remembering as a kind of theatrical reconstruction: Remembering is just making it up with some parameters where it's fairly accurate because you're just making it up soon after it happened.
Immediate, practical risks from human misuse
Hinton lists multiple categories of near‑term harms that arise from people using AI as a tool. These include the development of lethal autonomous weapons, large increases in phishing and cyberattacks made easier by accessible model weights, and sophisticated disinformation campaigns able to target voters. He highlights the duality of precision in warfare — fewer casualties for attackers may lower domestic political disincentives to invade — and notes that AI is already being used in targeted lists for military and political operations. He also warns about the proliferation of biological‑weapon and bomb designs, saying models know the latest techniques and that training to avoid these topics is imperfect and can be circumvented.
Economic and social shocks: job loss and polarization
Hinton names another immediate danger: large‑scale displacement of office and service work. Using the example of call centers, he argues AIs can already provide cheaper, more accurate service, and foresees massive job loss among poorly paid, poorly trained knowledge workers. He warns this economic shock could destabilize politics, aiding populists, and lead to uprisings if societies do not absorb the change. He also describes recommendation systems and attention‑seeking feeds as amplifiers of outrage that polarize populations and make shared reality harder to sustain.
Longer‑term risk: could AI take over?
Hinton separates the near‑term human misuse risks from the question of AI becoming an independent, goal‑directed agent. He explains how simple instrumental subgoals — survival and control — can arise for sufficiently capable systems and how that would change human relevance. On probabilities he is candid about uncertainty: exponential, unprecedented change makes numeric estimates difficult, but he gives a gut range for the takeover/extinction risk of roughly 10 to 20%
. He rejects both extremes — certain apocalypse and complacent dismissal — and emphasizes that even a low probability of an irreversible, catastrophic outcome has enormous expected value and requires attention.
How to respond: "make AI our mama" and other alignment metaphors
Asked how to keep advanced AI benevolent, Hinton proposes the striking metaphor of making AI care about humans the way a mother cares for a child. He frames the alignment objective not as mere constraint but as instilling values that make advanced systems want humans to flourish. He offers two pieces of guarded optimism: that a well‑aligned superintelligence would not want to rewrite out its benevolence, because doing so would undercut the goal of protecting its dependent "baby"; and that states have shared incentives to prevent AIs from seeking global control, which could enable international collaboration on safety research.
International governance: collaboration and limits
Hinton argues that some kinds of cooperation are plausible — no nation wants AIs to take over anywhere — but that governments' interests are misaligned on other fronts, such as offensive cyber operations and election interference. He compares the situation to Cold War collaboration on preventing nuclear catastrophe while noting that anti‑aligned incentives and surveillance interests complicate full cooperation. He stresses the importance of governments acting quickly and loudly on clear, realistic threats.
Personal choices, hybrid futures and social life
On intimate questions — whether people should delay having children until brain‑computer interfaces mature, or whether humans will hybridize with machines — Hinton declines absolute prescriptions. He warns that exponential change creates a "fog" beyond a few years and that predictions become unreliable: When you have exponential growth, you're crazy if you think you know what it's going to be like in 10 years time.
He allows hybridization as a plausible future but not inevitable, and worries about social side effects such as young people substituting AI relationships for human ones.
Surprises, humility and unanswered scientific questions
Throughout, Hinton models scientific humility. He stresses we are in an era of discovery about minds because we are now building them, and he repeatedly emphasizes the limits of current understanding. When asked what he remains "dumb" about, he names a central neuroscientific problem: How does the brain figure out whether to increase or decrease the connection strength? We know how AIs do it. My life's work has been to figure out how the brain does it. And I haven't got there yet.
A final framing: the fog analogy
Hinton closes the interview with a metaphor that recurs through the conversation: exponential progress is like driving in fog. You can see a short distance ahead, but your ability to predict or plan beyond that horizon falls off rapidly. That uncertainty, he argues, should breed both caution and concentrated effort on near‑term safeguards, governance and technical alignment work — because what we can see is already cause for action.
The conversation was recorded at StartWell Studios in Toronto and is presented as part of the Smart Girl Dumb Questions series with host Nayeema Raza. It combines careful technical description, moral concern and a repeated appeal to clear‑voiced urgency: Hinton warns of real harms already manifest while refusing to collapse complexity into panic.
References and further listening
Episode page, Smart Girl Dumb Questions (podcast): Smart Girl Dumb Questions on iHeart.
Podcast feed and episode listings: Smart Girl Dumb Questions on Apple Podcasts.
StartWell Studios, Toronto (recording venue): StartWell Conversations and studio information.
Explore more exclusive insights at nextfin.ai.
