In his recent interview at Ai4 in Las Vegas, conducted by Shirin Ghaffary (Bloomberg News), Geoffrey Hinton addressed how long it might take to reach superintelligence (5-20 years), that Ilya Sutskever wasn’t telling him how he wants to make it safe, which US frontier labs take safety more seriously (Anthropic, DeepMind) and which less so (OpenAI, Meta, xAI), where the US and China could collaborate (e.g., existential threats), which areas could benefit most from AI (health and education), and why it matters to bolster AI research in universities (AI was invented in academia, which is also more long-term oriented). Yet, as expected, his primary concern was AI as an existential threat – framed through his notion of “motherly” or “maternal AI,” which he had already mentioned on other occasions.
Hinton argues that once AI becomes far smarter than humans, traditional control will be impossible, as there are few cases where a less intelligent being can dominate a more intelligent one. The notable exception, he argues, is the mother–child relationship, where “maternal instincts” ensure care and protection. Drawing on this evolutionary precedent, he envisions “Motherly AI” – systems inherently motivated to safeguard and promote human flourishing – and believes any nation, especially the US and China, would collaborate to develop such a primary safety mechanism, since no one wants to face an existential threat. Hinton implied that such a safety mechanism would not arise from regulation, alignment, or reasoning alone, but from a design grounding the system in some form of artificial “instinct” or “emotion,” which the system cannot simply override. He did not elaborate on the computational foundations this would require, only that current architectures are insufficient.
AI’s “Godmother,” Fei-Fei Li, who knows Geoffrey Hinton well and regards him as a longstanding mentor, did not share his vision – not because it perpetuates the image of women reduced to motherhood and instinctive biological cues, nor because it anthropomorphizes AI. Rather, she emphasized in her interview carried out by Matt Egan from CNN that AI is a civilizational technology, with humanity remaining at the centre. She therefore rejected the notion of a “maternal AI” that treats humans as “children” – a framing she views as stripping away human dignity and agency, and as effectively outsourcing moral responsibility to machines.
By contrast, Yann LeCun’s view is not normative, as in Fei-Fei Li’s legitimate postulate on human dignity, but architectural. Commenting on social media, LeCun seems to confirm Hinton’s idea but as “simplified version of what [he has] been saying for several years: hardwire the architecture of AI systems so that the only actions they can take are towards completing objectives we give them, subject to guardrails.” LeCun calls this “objective-driven AI.” These hardwired constraints could include submission to humans, empathy, or simple prohibitions like “don’t run people over” or “don’t flail your arm if you are holding a knife.” He also speculates that evolution extended beyond maternal care, as a side-effect, to “the objectives that drive our social nature,” motivating us to “protect and care for helpless, weaker, younger, or cute beings of other species.” In his earlier technical proposal from 2022, A Path Towards Autonomous Machine Intelligence, he outlines an architecture where behaviour is driven by the interaction of predictive world models, actors, and trainable critics, all guided by immutable “cost functions.” In this setup, the actor chooses possible actions, the world model predicts their outcomes, and the critics score those outcomes against intrinsic objectives and guardrails. “Intrinsic care” could thus function as a negative cost for harmful outcomes or a positive reinforcement for protective behaviour, shaping the system’s choices regardless of its intelligence. Here the functional analogy aligns with evolutionary design: just as natural selection hardwired parenting instincts into species to safeguard survival, AI systems could be built with “intrinsic guardrails” that channel their intelligence towards bounded, safe behaviour.
Now, Hinton did not specify during his talk whether he envisions maternal instinct being hardwired into future model designs, or whether such behavioural patterns or surface actions could instead be learned through some form of “evolutionary algorithm” – where care might correlate with fitness and extend beyond a specific user (“child”) to society at large, an amorphous “group of children.” Nor did he address how cultural and societal mediation shape and condition such drives. While hardwiring may be an option and could co-exist within future architectures, it is evident that the breadth and depth of such deterministic functions cannot fully capture the complexity of human psychology and care.
Hardwiring “care” may prevent an AI from taking actions deemed harmful to humans, but such constraints address only outward/surface actions – not the deeper meaning or unintended psychological consequences of care. What is missing is the “subject:” the inside-out perspective from which things actually matter. Language illustrates this gap well: humans use words from lived experience, always grounded in “I,” “you,” “here,” and “now.” LLMs, by contrast, generate well-formed text without that causal anchoring in lived reality; they model form, not meaning. Those are representations of representations. LLMs don’t think but generate correlated patterns of thought, thus stripped of competence. Even their “proto-grammar” (emergent syntax-like patterns LLMs derive from large text corpora) doesn’t bring the subject back – the space where outer representations connect with inner experience. Their loss function, like our preoccupation with significance, offers functional similarity but not grounding in lived reality.
The same holds for care. A system can model “care” as a cost function, but it would be like modelling, to use Christof Koch’s metaphor, a hurricane without ever feeling the rain – it lacks the real experience that gives the model context. The notion of instantiating care is therefore not simply to enact protective behaviour but to grasp its deep, contextual meaning. Without this subjective grounding, AI cannot understand what it means to care and might fail to make the right decision in a given particular situation that is historically, culturally, and psychologically mediated. Fixed control functions are of course necessary for certain risks, but they remain only safeguards.
The question of subjectivity ultimately leads to the question of consciousness: can machines ever achieve even a minimal form sufficient to ground meaning from the inside out? LeCun’s architecture avoids this question probably for pragmatic reasons (or because he believes abstract intelligence grounded in a sensory world model is sufficient), but we cannot. If we want to remain vigilant about the limits of AI’s current design – or use those limits as inspiration for future architectures – we must confront the relation between subjectivity and consciousness directly.
I will do so in my next Substack, proposing four views on consciousness as points of departure.
P.S. In the video recorded at Ai4 in Las Vegas on August 12, 2025, Geoffrey Hinton explains his notion of “maternal AI,” followed by Fei-Fei Li’s rebuttal and her own definition of AI as civilizational technology.
