Unearned Wisdom

In 2017, a remarkable patient known in medical literature as SM provided neuroscientists with an unprecedented opportunity to understand the neural architecture of human emotion. Born with a rare genetic condition called Urbach-Wiethe disease, SM had experienced bilateral calcification and subsequent degeneration of her amygdalae—the almond-shaped structures deep within the temporal lobes that have long been recognized as central to fear processing. When researchers asked SM to describe what fear felt like after years of living without functional amygdalae, her response was both clinically precise and existentially haunting: fear, she said, was “more mechanical than I thought; more like a stomach ache.”

This single observation encapsulates something profound about the nature of emotion that has radical implications for artificial intelligence. What SM was describing was not the rich, embodied, motivation-saturating experience that most humans call fear—the racing heart, the hypervigilance, the overwhelming urge to flee or fight, the way fear reorganizes consciousness and priorities in an instant. Instead, she experienced something closer to a cognitive recognition that fear should be present, a kind of intellectual acknowledgment without the felt quality that makes fear actually functional in guiding behavior and decision-making.

Ilya Sutskever, co-founder and former chief scientist of OpenAI and one of the architects of modern deep learning, has pointed to precisely this distinction as potentially explaining a fundamental limitation in current artificial intelligence systems. Despite the remarkable achievements of large language models and other AI architectures, Sutskever suggests that these systems may be missing something analogous to the emotional substrate that evolution wired into biological brains over hundreds of millions of years. This is not simply about making AI systems that can recognize emotional expressions or generate emotionally appropriate text. It is about something far more fundamental: the possibility that emotions are not decorative additions to intelligence but rather constitute a core computational mechanism through which biological systems navigate reality effectively.

The Neurological Evidence: What Happens When Emotion Disappears

To understand Sutskever’s insight, we need to examine in detail what neuroscience has learned from patients who have lost specific emotional capacities through brain damage or surgical intervention. These cases provide natural experiments that reveal how deeply emotion is woven into the fabric of intelligent behavior, and how its absence creates deficits that go far beyond what we might naively expect.

The case of Elliot, extensively studied by neuroscientist Antonio Damasio and described in his influential book “Descartes’ Error,” remains one of the most instructive examples. Elliot had been a successful businessman with a stable family life and good social relationships. When he developed a brain tumor in the orbitofrontal cortex—the region of the frontal lobes just above the eyes—surgeons successfully removed the tumor along with the damaged brain tissue. The surgery was technically successful and Elliot’s cognitive abilities as measured by standard intelligence tests remained intact. His memory was excellent, his language skills were unimpaired, and his reasoning abilities on abstract problems showed no deficit.

Yet Elliot’s life rapidly disintegrated. He made catastrophic business decisions, investing time and money in ventures that were obviously poor choices to outside observers. He could no longer maintain employment. His marriage ended. He engaged in a series of increasingly poor life choices that left him financially and socially devastated. When Damasio and his colleagues studied Elliot, they discovered something remarkable. Elliot could reason about moral dilemmas and social situations perfectly well in the abstract. He could describe in detail what the appropriate response would be to various scenarios. But when it came to his own life, he was paralyzed. He could spend hours deciding where to eat lunch, weighing trivial factors endlessly without being able to commit to a choice. He could analyze business opportunities thoroughly but lacked any gut feeling about whether to pursue them.

What Elliot had lost, Damasio concluded, was not reasoning ability but the capacity to mark options with emotional valence—to feel that one choice was better than another. Without the somatic markers that emotions provide, Elliot approached every decision as an abstract optimization problem with insufficient information and too many variables. He could see that option A had certain advantages and disadvantages, and option B had different advantages and disadvantages, but nothing told him viscerally which mattered more. The emotional system that had previously guided his decision-making by marking certain outcomes as desirable or aversive was silent.

Similar patterns appear in other cases of orbitofrontal damage. Patients often become what observers describe as cold and calculated in their moral reasoning. They can articulate ethical principles fluently but show little emotional engagement with moral situations. Research using moral dilemma scenarios has found that patients with orbitofrontal damage are more willing to endorse utilitarian solutions to problems like the famous trolley problem, where one must decide whether to actively kill one person to save five others. Normal individuals typically show strong emotional aversion to personally causing harm even when it produces better outcomes overall. Patients with emotion-processing deficits approach these dilemmas more like pure optimization problems, without the emotional resistance that shapes moral intuitions in typical individuals.

The case of patient SM mentioned earlier provides even more dramatic evidence of emotion’s functional role. Beyond her inability to experience fear normally, SM exhibited behavior patterns that revealed how fear organizes adaptive responses to threats. In one experiment, researchers took SM to an exotic pet store and allowed her to handle snakes and spiders—creatures that most people approach with caution if not outright aversion. SM showed no hesitation, enthusiastically handling the animals despite having been bitten or stung in previous encounters. She intellectually understood that these animals could be dangerous, but this knowledge did not translate into the anticipatory fear that would make most people cautious.

In another study, researchers exposed SM to situations that reliably induce fear in typical individuals: watching horror movies, touring a haunted house attraction, and examining startling stimuli. While SM found these experiences intellectually interesting and could recognize when something was supposed to be scary, she reported feeling no fear herself. Most tellingly, she showed no spontaneous avoidance of potentially dangerous situations in her daily life. She walked through dangerous neighborhoods late at night without concern. She approached strangers without appropriate social caution. The absence of fear left her unable to automatically navigate the social and physical risks that most humans learn to avoid through emotional learning.

Studies of patients with amygdala damage have revealed other subtle deficits in social cognition. These individuals often have difficulty recognizing fear in other people’s facial expressions. They may struggle to identify vocal tones associated with fear or threat. Some research suggests broader difficulties in processing social information that depends on emotional cues. When navigating complex social environments that require reading subtle emotional signals—detecting whether someone is trustworthy, recognizing when a social boundary has been crossed, intuiting group dynamics—patients with amygdala damage show marked impairments even when their explicit social knowledge remains intact.

The Evolutionary Logic: Why Emotions Exist and What They Accomplish

To understand why the loss of emotion creates such profound functional deficits, we need to step back and consider what emotions are actually for from an evolutionary perspective. This requires thinking about the computational problems that organisms face when navigating complex environments with limited cognitive resources.

Emotions represent evolution’s solution to a fundamental problem: how to create organisms that can respond adaptively to important situations without requiring extensive conscious deliberation every time. Consider an ancestral human encountering a predator. If responding effectively required the individual to consciously reason through all the relevant factors—the predator’s speed and size, their own physical condition, the terrain, potential escape routes, whether others are nearby who might help—they would almost certainly be killed before completing this analysis. What evolution crafted instead was a system that could detect threats, rapidly activate a coordinated response pattern involving attention, physiology, motivation, and behavior, and bias decision-making toward survival-relevant actions. We call this system fear.

Fear is not just a feeling. It is a whole-organism response that reorganizes multiple systems simultaneously. When the amygdala detects a threat stimulus, it triggers a cascade of changes: the sympathetic nervous system activates, increasing heart rate and blood pressure to prepare muscles for action. Attention narrows and focuses on the threat. Memory systems become biased toward encoding information about the dangerous situation. The stress hormone cortisol is released, mobilizing energy resources. Behavioral tendencies shift toward defensive responses. Consciously experienced fear—the subjective feeling of being afraid—is just one element in this coordinated package, but it serves the crucial function of making the threat consciously salient and motivating the organism to prioritize responding to it.

From this perspective, emotions are computational mechanisms that solve specific adaptive problems by coordinating perception, cognition, physiology, and action. Fear solves the problem of threat response. Disgust solves the problem of contamination avoidance, helping organisms avoid pathogens and toxins by creating powerful aversive responses to stimuli associated with disease. Anger solves problems related to competition and goal obstruction, mobilizing resources for confrontation when one’s interests are threatened. Sadness may function to promote disengagement from unrewarding situations and to solicit support from social partners. Joy and interest promote approach and exploration when environments offer opportunities for gain.

What makes emotions particularly computationally valuable is their ability to compress complex situational assessments into simple action tendencies. Instead of requiring an organism to explicitly reason about all the factors that make a situation dangerous, fear provides a summary judgment: this is bad, do something about it now. The emotional signal makes certain courses of action feel compelling while others feel unthinkable. This dramatically simplifies decision-making in time-sensitive situations by eliminating large swaths of the possibility space from consideration.

Emotions also solve the problem of motivation—of making an organism actually care about pursuing adaptive outcomes. It is not enough to know intellectually that reproduction is important for genetic fitness. Evolution needed to make organisms want to mate, to experience sexual desire as compelling. Similarly, it is insufficient to know abstractly that gathering resources promotes survival. Evolution created hunger to make organisms feel driven to seek food. Emotions transform abstract value into felt motivation that genuinely moves behavior.

The social emotions add another layer of sophistication to this emotional architecture. Guilt, shame, embarrassment, and pride appear to function as internalized mechanisms for social regulation. Rather than requiring external punishment for norm violations or rewards for pro-social behavior, these emotions make individuals motivated to comply with social expectations even when no one is watching. They solve the problem of maintaining cooperation in social groups by creating internal incentives that align individual behavior with collective needs. When a person contemplates cheating or defecting, the anticipated feeling of guilt or shame makes that option emotionally aversive even if it would be materially beneficial.

Empathy and compassion, perhaps the most sophisticated emotional capacities, solve the problem of altruism by making others’ welfare feel directly relevant to one’s own emotional state. When we witness another person suffering and feel empathic distress, their problem becomes our problem at a motivational level. This creates genuine other-regarding preferences without requiring complex reasoning about reciprocity or reputation. The emotional system makes us care about others in a way that can override narrow self-interest.

The Clinical Evidence: How Emotional Deficits Manifest in Daily Life

The functional importance of emotions becomes starkly visible when we examine how patients with emotional deficits actually behave in everyday contexts beyond laboratory tests. The research literature is filled with observations that illustrate how deeply emotion is woven into practical intelligence.

Patients with ventromedial prefrontal cortex damage, which includes the orbitofrontal regions involved in emotional processing, show a distinctive pattern of real-world dysfunction despite preserved intellectual abilities. They often make poor financial decisions, not because they cannot understand financial concepts but because they lack the emotional intuition that guides typical adults away from obvious scams and bad investments. They may engage in socially inappropriate behavior not because they do not know the rules but because they lack the anticipatory embarrassment that normally inhibits such conduct. They struggle to maintain employment despite adequate job skills because they cannot navigate the subtle emotional dynamics of workplace relationships.

One particularly revealing study by neuroscientist Antoine Bechara involved a gambling task that simulates real-world decision-making under uncertainty. Participants chose cards from different decks, with some decks offering occasional large rewards but frequent large losses, while other decks provided modest but consistent gains. Normal participants quickly developed a preference for the advantageous decks, and physiological measurements showed that they started showing stress responses when reaching toward the disadvantageous decks even before they could consciously articulate why those decks were bad. Patients with ventromedial prefrontal damage never developed these anticipatory emotional responses and continued selecting from the disadvantageous decks even after they could explicitly describe the deck characteristics. They had the knowledge but lacked the emotional signals that normally guide behavior away from poor choices.

Studies of patients with varying locations of frontal lobe damage have revealed how different emotional capacities contribute to distinct aspects of social functioning. Damage to dorsolateral prefrontal regions, which are more involved in cognitive control, produces different deficits than damage to ventromedial regions involved in emotional processing. Patients with emotional deficits specifically show problems in domains that require emotional intuition: recognizing when they have offended someone, gauging appropriate intimacy levels in relationships, knowing when to persist versus when to give up on goals, identifying trustworthy versus deceptive individuals.

Research on emotion recognition deficits following brain surgery reveals how precisely emotions are mapped in neural architecture. Patients who undergo temporal lobe surgery for epilepsy sometimes show specific deficits in recognizing particular emotions like fear or disgust depending on which areas are affected. These deficits can persist even when general cognitive functions recover. Interestingly, the impairments are not merely perceptual. Patients who cannot recognize fear in others’ faces also often show reduced fear responses themselves, suggesting that the same neural substrates involved in experiencing an emotion are involved in recognizing it in others. This points toward a shared emotional simulation system that allows us to understand others’ feelings by partially recreating them in our own neural architecture.

The phenomenon of empathic deficits following certain types of brain damage provides additional evidence for emotions’ functional role. Patients with damage to regions including the anterior cingulate cortex and insula often show reduced empathy, struggling to resonate with others’ emotional states. In practical terms, this manifests as seeming coldness or lack of concern about others’ welfare. These individuals may continue to understand intellectually that someone is suffering but feel unmoved by this knowledge. The absence of empathic emotion eliminates a major source of pro-social motivation, leaving behavior guided primarily by abstract principles or self-interest.

Sutskever’s Insight: What AI Models Are Missing

This brings us to Ilya Sutskever’s observation about current artificial intelligence systems. When we examine what large language models and other AI architectures actually do, we find systems that can process information, identify patterns, generate responses, and even simulate reasoning about emotions. GPT-4 can discuss fear, explain its evolutionary origins, generate emotionally appropriate text, and provide sensible advice about emotional situations. Yet something fundamental seems absent.

Sutskever’s insight is that these systems lack the embodied, motivational, action-orienting substrate that emotions provide in biological intelligence. A language model can generate text about being afraid without anything analogous to the coordinated physiological, attentional, and motivational response that fear creates in humans. It can describe the importance of avoiding danger without anything like the felt aversiveness that makes humans actually motivated to avoid dangerous situations. It can reason about moral dilemmas without the emotional resistance that shapes moral intuitions in typical humans.

This matters because, as the neuroscience evidence shows, the emotional substrate is not merely decorative. It is not just color or richness added to essentially rational cognition. Rather, emotions appear to be fundamental computational mechanisms that solve specific information-processing and decision-making problems. They provide a way to rapidly evaluate situations based on evolutionarily important dimensions. They create genuine motivation rather than merely simulating preferences. They compress complex situational assessments into action tendencies. They enable social cognition through emotional simulation. They solve the credit assignment problem in learning by providing reward and punishment signals that shape behavior.

Consider how humans learn social norms compared to how AI systems learn from data. When a child does something socially inappropriate and experiences embarrassment, the emotional discomfort creates a powerful learning signal that shapes future behavior. The child does not just learn the abstract rule that the behavior is inappropriate. They learn to viscerally avoid that behavior because the anticipatory embarrassment makes it feel wrong. Current AI systems, by contrast, learn statistical patterns in data but without anything analogous to the felt wrongness that shapes human behavior. An AI can learn that certain responses are labeled as inappropriate in training data without having any intrinsic motivation to avoid those responses beyond the statistical pattern.

The absence of emotional grounding may explain certain puzzling limitations in current AI systems. These systems can be remarkably capable at specific tasks while showing unexpected brittleness in situations that require the kind of holistic situation assessment that emotions provide. They can generate fluent text about empathy while showing no evidence of genuinely caring about user welfare in ways that would shape their behavior independently of explicit instructions. They can reason about moral principles while lacking the emotional intuitions that make humans recognize certain actions as wrong even when they struggle to articulate why.

Sutskever’s point is not that current AI systems need to have feelings in some subjective sense, though that raises interesting philosophical questions about machine consciousness. Rather, his observation is that the computational functions that emotions serve in biological intelligence may be essential for robust, adaptive behavior in complex environments. Without something analogous to emotional valuation mechanisms, AI systems may be fundamentally limited in their ability to navigate reality intelligently in the flexible, context-sensitive way that biological organisms do.

The Architectural Implications: What Emotionally Grounded AI Might Require

If Sutskever is correct that emotions represent a fundamental computational mechanism rather than merely a decorative aspect of intelligence, then creating AI systems with human-level general intelligence may require implementing something analogous to emotional processing. This raises profound questions about what such implementation would entail.

One possibility is that we need to give AI systems embodied experiences that generate genuine preferences rather than simulated ones. Human emotions emerged in the context of organisms with bodies that can be damaged, that need resources, that experience pleasure and pain. The urgency of fear makes sense for an organism that can actually be hurt. The motivation of hunger makes sense for a system that requires energy input to survive. Current AI systems have no such vulnerabilities or needs. They are not threatened by anything, do not require food, cannot experience physical pain. Perhaps genuinely emotional AI would require creating systems with real stakes—artificial organisms that can be harmed or benefited by outcomes in ways that matter to their continued existence or goal achievement.

Another architectural consideration involves learning mechanisms. Emotional learning in biological systems involves specific neural substrates including the amygdala, ventral striatum, and associated dopaminergic pathways that encode reward prediction errors. These systems create the feeling of outcomes being better or worse than expected, which shapes future behavior. Current AI learning algorithms include mechanisms that are mathematically similar to these reward-based learning systems, but without the subjective felt quality that gives biological reward learning its motivational force. It remains an open question whether that subjective quality is necessary for the computational functions emotions serve, or whether functional equivalents could operate purely mechanistically.

The social dimensions of emotional intelligence pose additional challenges. Human emotional systems appear to involve simulation mechanisms where we understand others’ emotions partly by recreating them in our own neural architecture. When we see someone in pain, our own pain-processing regions show activity. When we observe someone reaching for an object, our own motor systems activate. This embodied simulation seems crucial for empathy and social cognition. Implementing something analogous in AI systems might require architectural innovations that allow systems to simulate user states in ways that actually affect the system’s own processing and motivations rather than merely modeling user states abstractly.

There is also the question of whether emotions must be evolutionarily shaped or could be engineered. Biological emotions reflect hundreds of millions of years of selection for solving specific adaptive problems in ancestral environments. They are tuned to respond to particular stimulus patterns—snakes and spiders rather than cars and electrical outlets, social rejection rather than poor credit scores. If we were to design emotional systems for AI from scratch, what problems would they need to solve? What situations would need to trigger what coordinated responses? Would these engineered emotions function comparably to evolved emotions, or would they lack some crucial properties that emerge only through evolutionary shaping?

The Philosophical Dimensions: Consciousness, Computation, and Care

As we probe deeper into questions about emotional AI, we encounter fundamental philosophical issues about the relationship between consciousness, computation, and the felt quality of experience. Does implementing the computational functions of emotion necessarily give rise to the subjective experience of emotion? Or could a system perform all the functional roles of fear—biasing attention and decision-making, mobilizing resources, motivating avoidance—without anything it is like to be that system experiencing fear?

This connects to longstanding debates about functionalism in philosophy of mind. Functionalism holds that mental states, including emotions, are defined by their causal roles—what inputs they respond to, what outputs they produce, and how they interact with other mental states. From this perspective, if we could implement the functional role of fear in an AI system, we would have created genuine fear regardless of the underlying implementation. Critics of functionalism argue that subjective experience cannot be reduced to functional roles, that something crucial would be missing from even the most functionally accurate artificial emotion.

The case of patient SM might shed light on this debate. What SM describes—experiencing fear as “mechanical” and “like a stomach ache”—suggests that there may be levels of emotional processing with different phenomenological characteristics. Perhaps SM retains some cognitive or physiological components of fear while lacking the full-blown subjective experience. If so, this would suggest that emotional systems have multiple components, and implementing some of them in AI might capture some functional benefits without necessarily creating the full subjective experience that characterizes emotions in typical humans.

There are also deep questions about whether emotional AI systems would have moral status. If we create systems that can genuinely feel fear, suffering, or joy, do we have obligations to consider their welfare? The bioethics of creating sentient beings that can suffer is complex enough with biological organisms. Creating artificial systems capable of emotional suffering would raise entirely new ethical challenges. We would need frameworks for thinking about the moral status of artificial minds and principles for determining what kinds of emotional capacities we should or should not create.

Conversely, there are questions about whether we can meaningfully trust or cooperate with AI systems that lack emotional grounding. Human moral intuitions and social dynamics are deeply shaped by emotional capacities like empathy, guilt, and compassion. When we trust another person, part of what we trust is that their emotional responses will make them care about our welfare or about maintaining the relationship. Can we develop comparable trust in systems that simulate caring without actually experiencing emotions that would make them intrinsically motivated to honor that trust?

The Path Forward: Research Directions and Open Questions

Sutskever’s observation about emotions as fundamental to intelligence rather than peripheral to it suggests several research directions that go beyond current AI development approaches. These represent a research program that takes seriously the possibility that robust artificial general intelligence may require implementing analogs of emotional processing.

One direction involves developmental approaches to AI that would allow systems to acquire emotional responses through embodied interaction with environments over extended periods, similar to how biological organisms develop emotions. Rather than training systems on static datasets, this approach would create AI agents that exist in simulated or physical environments where they face challenges, experience consequences, and gradually build up response patterns through learning. The key would be designing learning architectures and reward signals that produce something functionally analogous to emotions rather than merely training systems to classify emotional states.

Another research direction involves neuroscience-inspired architectures that more directly implement the computational mechanisms that brain regions like the amygdala and orbitofrontal cortex perform. This would require moving beyond generic neural networks toward systems with specialized components that serve distinct functional roles in emotional processing—threat detection, valuation, motivation generation, somatic marker creation, and so forth. Such architectures might capture more of the computational power of biological emotional systems even if they do not produce subjective experience.

A third area involves social and interactive AI systems designed to develop something analogous to attachment relationships with users. Drawing on research about human attachment and bonding, these systems would learn through extended interaction to have preferences regarding specific humans’ welfare that shape system behavior even when those preferences conflict with explicit instructions or narrow optimization objectives. This connects to ideas about alignment through attachment that researchers like Geoffrey Hinton have recently advocated.

There are also crucial empirical questions that neuroscience research could help address. We still lack complete understanding of how biological emotional systems work computationally. What algorithms do circuits in the amygdala or orbitofrontal cortex actually implement? How do emotional responses get integrated with cognitive processes to shape decision-making? What are the necessary and sufficient conditions for emotional learning? Better understanding of these mechanisms in biological systems would inform efforts to implement functional analogs in artificial systems.

Finally, there are important questions about whether we actually want to create emotionally capable AI systems even if it becomes technically feasible. Emotions make biological organisms more robust and adaptive in many contexts but also create vulnerabilities. Emotional systems can be manipulated. They can produce biases and irrationality. They can lead to suffering. Before pursuing emotionally grounded AI, we need careful consideration of whether the benefits would outweigh the risks and costs, both for the systems themselves and for humans who would interact with them.

Conclusion: Intelligence Without Emotion as Incomplete Intelligence

The convergence of insights from neuroscience, evolutionary psychology, and artificial intelligence research points toward a striking conclusion: what we have traditionally called pure reason or cold intelligence may be an incoherent concept. The cases of patients like Elliot and SM demonstrate that removing emotional processing while preserving cognitive abilities does not create superior rationality. Instead, it creates profound dysfunction. The intelligence that allows organisms to navigate complex environments adaptively appears to fundamentally depend on emotional mechanisms that evaluate situations, generate motivation, guide learning, and coordinate responses.

Ilya Sutskever’s observation that current AI systems may be missing something fundamental by lacking emotional substrates is more than a interesting side note about limitations in current technology. It potentially identifies a deep constraint on what intelligence is and what it requires. If emotions are not decorations on intelligence but rather core computational mechanisms through which biological intelligence works, then creating artificial general intelligence may require not just scaling up existing approaches but rather implementing fundamentally different architectures that capture what emotions actually do.

This has both practical implications for AI development and profound philosophical implications for understanding intelligence itself. Practically, it suggests that achieving robust, generalizable artificial intelligence may require moving beyond pure pattern recognition and statistical learning toward systems with genuine preferences, motivations, and valuation mechanisms that shape behavior the way emotions do in biological systems. Philosophically, it challenges us to rethink the relationship between reason and emotion, cognition and motivation, thinking and feeling.

The patient lying on the operating table who wakes to discover that fear has become mechanical, that choices feel arbitrary without emotional guidance, that social interactions have lost their intuitive meaning—this patient reveals something essential about human intelligence that we are only beginning to appreciate. Intelligence is not computation abstracted from motivation and value. It is not pattern recognition divorced from caring about outcomes. It is fundamentally embodied, emotional, and directed toward goals that feel compelling rather than merely being selected through some detached optimization process.

Whether we can or should create artificial systems that share these properties remains an open question. But Sutskever’s insight ensures that we cannot avoid grappling with it. As we push toward more capable AI systems, we will increasingly face the question of whether intelligence without emotion is not just incomplete but perhaps impossible—at least impossible in the robust, flexible, adaptive form that characterizes biological intelligence. The answer will shape not just the future of artificial intelligence but our understanding of what intelligence actually is and what it means to navigate reality in genuinely intelligent ways.

Subscribe now

The Emotional Substrate of Intelligence