We usually think of fear as an emotion. Evolution may have understood it very differently. The more I thought about it, the less fear resembled a feeling and the more it resembled an answer.
That thought first occurred to me after reading a recent proposal arguing that one of AI’s greatest limitations is the absence of fear and not intelligence. The suggested remedy was Large Danger Models (LDMs), systems intended to capture the kind of danger sensitivity that biological organisms gradually evolved. AI’s most unsettling failures seem less like failures of reasoning than failures to recognise consequences. Models hallucinate with remarkable confidence, recommend actions whose risks they scarcely comprehend, and often treat ordinary uncertainty and irreversible catastrophe as though they belonged to the same category.
Yet the longer I sat with that thought, the more the original question began to feel secondary. It seemed to assume that fear was the thing needing explanation, when perhaps fear was itself an explanation for something older. The proposal ultimately points toward a much older question, one that reaches far beyond AI and into the history of life itself.
Why does fear exist at all?
We often speak of fear as though it were an unfortunate feature of being human, something courage overcomes, maturity tempers, and civilization gradually leaves behind. That language quietly assumes fear is little more than an emotional burden.
Every organism inhabited a world in which uncertainty could never be eliminated, and hesitation could not always be afforded. Most mistakes merely consumed energy, but a few eliminated every opportunity to learn from them. Fleeing a harmless shadow wasted effort, whereas ignoring a predator ended the possibility of every future decision, leaving natural selection with remarkably little interest in perfect reasoning and every interest in organisms that remained alive in a world where the consequences of being wrong were profoundly unequal.
Seen from that perspective, fear begins to look very different. It resembles less an emotion than a remarkably elegant solution to an ancient problem, one that evolution had been refining for millions of years before anyone understood probability, uncertainty, optimization, or risk management. Fear did not eliminate uncertainty, because uncertainty could never be eliminated; what it changed was the way living organisms responded to it, directing attention toward consequential danger before conscious reasoning ever had the chance to intervene.
Before asking whether artificial intelligence needs fear, we may first need to understand the problem that fear was solving all along.
More than a century ago, William James made an observation that has only grown more interesting with time.
“The progress from brute to man is characterized by nothing so much as by the decrease in frequency of proper occasions for fear.”
James wasn’t simply making a historical observation; he was quietly pointing toward something deeper. Civilization reduced the occasions for fear, but it never removed the machinery that made fear possible. We replaced exposed camps with homes, tribal conflict with laws, famine with agriculture, epidemics with medicine, and isolation with institutions. The everyday landscape of danger slowly receded.
James never suggested that fear itself disappeared, nor could he have, because evolution rarely abandons inventions that have survived millions of years of natural selection. Civilization changed the world around us, but it never changed the evolutionary logic that had made fear indispensable in the first place.
The more I reflected on James’s observation, the more difficult it became to think of fear merely as an emotion. It began to resemble something else entirely: an extraordinarily efficient solution to a problem that every living organism confronts long before it possesses anything we would call intelligence.
The deeper problem was never danger itself but asymmetry, and in nature, an unnecessary retreat wastes energy, whereas ignoring a predator may eliminate every future opportunity to learn from the mistake. Evolution therefore faced a world in which the consequences of being wrong were profoundly unequal, and fear emerged as one of evolution’s most elegant solutions.
Fear redirects attention before conscious reasoning has time to intervene, interrupts routine behaviour when familiar patterns become unreliable, raises vigilance precisely when uncertainty becomes consequential, and quietly assigns vastly different costs to different kinds of mistakes.
Looked at this way, fear is far more than an emotion that organisms happen to experience. It is a biological architecture for managing consequential uncertainty in a world where survival depends less on being perfectly right than on avoiding the kinds of mistakes from which there is no recovery.
How does any intelligence survive a world where some mistakes cannot be undone
Seen in this light, the proposal for Large Danger Models becomes more compelling. It recognizes that intelligence without a rich understanding of danger will remain incomplete.
Treating fear as the indispensable ingredient merits closer examination, because history repeatedly reminds us that evolution’s solutions, however brilliant, are rarely the only ones. Before asking whether artificial intelligence can inherit fear, we should perhaps ask whether fear is the only way to solve this problem, or simply the way evolution discovered.
The Wright brothers didn’t inherit feathers because feathers were never the principle. They inherited the problem of flight.
We do not inherit fear. We inherit the problem that fear solved.
Once that distinction became clear, aviation stopped imitating nature and began discovering possibilities that nature itself had never explored.
Engineering has repeated this pattern ever since. Eyes are not the only way to see. Cameras solved the same problem through an entirely different architecture. Every mature engineering discipline begins by admiring nature and eventually progresses by extracting the principle hidden beneath its implementations.
Humans navigate danger through fear because evolution works within biological constraints. It had neurons, hormones, bodies, instincts, and millions of generations over which successful adaptations could slowly accumulate. Fear emerged because it proved astonishingly effective within those biological constraints.
Engineers inherit a radically different design space. They are free to work with models rather than instincts, simulations rather than generations, and deliberate explanations instead of blind evolutionary search. Nothing requires them to reproduce biology if they can solve the same evolutionary challenge by entirely different means.
The proposal no longer appears as an attempt to teach machines fear. Instead, it begins to resemble something far more ambitious: the search for an architecture capable of performing the computational work that fear has carried out for biological intelligence throughout evolutionary history. Such an architecture would not reproduce fear. It would distinguish between mistakes that can be corrected and those that permanently reshape the future.
If fear is the only way intelligence can navigate danger, artificial intelligence will somehow have to learn to fear. But if fear is simply evolution’s solution to a much older problem, then engineering is free to search for a different solution. The real question is no longer whether machines can feel fear. It is how they will recognise irreversible consequences before they act.
The question is no longer whether artificial intelligence can inherit fear, but whether it can discover a different answer to the problem that fear solved.
We do not yet know what such an architecture will look like.
Perhaps it will resemble Large Danger Models. Perhaps it will emerge from advances we have not yet imagined. History offers little reason to believe that engineering’s final answer will look identical to evolution’s first.
Any artificial intelligence that operates in the real world must distinguish between mistakes that are inconvenient and mistakes that permanently reshape the future. It must recognize that uncertainty is not merely the absence of information but often a signal to slow down, investigate further, and preserve options that may prove valuable later. It must also understand that some decisions deserve extraordinary caution because their consequences cannot easily be undone.
Those requirements describe not an emotion but an architecture.
Whether that architecture relies on simulations of future worlds, explicit models of irreversibility, mechanisms we have not yet discovered, or principles that will appear obvious only in retrospect is impossible to say. The problem remains the same. Every sufficiently capable intelligence, whether biological or artificial, must somehow learn to live in a universe where the consequences of mistakes are profoundly unequal.
Evolution did not merely produce organisms. Over billions of years, it discovered practical solutions to an astonishing collection of computational problems, using fear to manage danger before conscious reasoning could, memory to preserve experience, attention to allocate scarce cognitive resources, curiosity to balance exploration against exploitation, and pain to transform damage into information.
Birds taught us that flight was possible without revealing every way to fly. Evolution may have taught us how one form of intelligence survives uncertainty without revealing every possible architecture capable of doing the same.
A proposal for Large Danger Models invited us to ask whether artificial intelligence needs fear. It is an important question, but perhaps it is not the deepest one.
The deeper question is what problem fear has been solving all along.
Once we understand the function fear evolved to perform, we may discover that evolution’s greatest contribution was never fear itself. It was revealing a principle that engineering has only just begun to recognize.
Evolution spent billions of years discovering one answer to the question of how intelligence survives uncertainty.
Artificial intelligence may discover another.