Essays/Risk
Risk
Sangeetha Bharath
For many years, I lived life terrified of risk. Like billions of others, abuse and financial insecurity in childhood left me circumspect. My father worked three jobs at a time, and as a child alone with my mother, I quickly became an emotional broker. Caution ruled my decision-making for years, and meaningful personal risks were a luxury I could not afford.
I became religiously interested in mathematical sciences, neural networks, programming, cryptography, physics, as a crutch to shore myself against fear and to give myself tools to define risk as a variable. I built a world where I achieved my best outcomes in fear of uncertainty. I saw decisions as optimization problems to help me avoid instead of embrace risk.
I have changed a lot in the last few years of my life and have come to believe measuring risk isn’t about eliminating randomness but about being brave enough to act within it.
The mathematics & history of risk
I triple majored in computer science, mathematics, and comparative literature, working across the algorithmic, the quantitative, and the interpretive.
In my first year of college I took a class on linear algebra. Linear algebra is a foundational language across data, computers, algorithms, and even physical life. Studying it was a milestone in my multidisciplinary journey. Linear algebra breaks down complex nonlinear problems (life) into linear approximations. We use these approximations to model, simulate, and solve the more complex, non-linear phenomena.
Additionally, linear algebra is deeply related to probability theory. Linear algebra provides the machinery for representing and manipulating relationships. Probability theory tells us what those relationships mean in a world full of noise and uncertainty. The former provides the language; the latter, the story. They are essential to both AI and a mathematical understanding of natural life.
I was superficially obsessed with such classes, including one on economic principles, but I didn’t actually understand anything about probability or risk theory. I was excitedly working through profound mathematical concepts, learning about machine learning and neural networks, but still naively attached to a “perfect-conditions,” ECON 101 style of thinking that assumes rational human behavior and linearizes phenomena which in reality are much more jagged.
Risk in its mathematical form is simple:
Elegant and applicable to everything, the equation wraps our messy reality of uncertainty into a neat mathematical package and in doing so obscures the truth that real life requires us to discover probabilities the hard way.
Facing uncertainty, often when we don’t know the odds, is difficult and critical.
Risk assessment across history generally reveals humanity’s evolving relationship with uncertainty. Humans have dealt, reasoned, reckoned with uncertainty in varying ways since the dawn of our evolution. Ancient Egyptians faced the unpredictable flooding of the Nile, which could yield bountiful harvests for decades before delivering devastating droughts. Mathematical risk and decision theory expert Roger Cooke in his “A Brief History of Quantitative Risk Assessment” observed that “if the ancient Egyptians knew in advance exactly when the Nile would fail to flood, they would not have needed scribes, taxation, writing, calculations, surveying, geometry, or astronomy. Civilization owes much to risk. Without uncertainty there is no risk, only adversity.”1
The marriage of intuition and measurement finally found a more formal expression via Renaissance mathematicians (read: gamblers) in 1654. Blaise Pascal and Pierre de Fermat corresponded about a gambling problem (now referred to as The Problem of Points2) that had stumped prior mathematicians for two centuries. Their solution to an ostensibly trivial puzzle (how to fairly divide prize money when a game of chance must end early and one player is ahead) gave birth to probability theory.
Pascal engaged mathematically with probability but recognized its limits. Intensely religious, he saw quantitative reasoning as just one approach to truth and developed his famous “wager”3 not as a calculation of probability but as a framework for decision-making under uncertainty (his wager argued it is more rational to bet on God’s existence than against it). Modernity has largely forgotten this humility, assuming we can calculate what others knew we could only contemplate.
Decisive breakthroughs emerged from there in rapid succession. Jakob Bernoulli proved the law of large numbers, demonstrating that with sufficient observations, empirical probabilities converge to theoretical ones.4 This discovery illustrated that beneath apparent randomness lies mathematical certainty, and for the first time, humanity had mathematical proof that collecting enough data could reveal underlying truths about uncertain events.
From there, the discovery of the normal distribution tells us how observations scatter around their average, with small deviations being common and large ones becoming exponentially rare. If random events naturally cluster around an average in a predictable way, then risk isn’t just something that happens to us; it’s something we can measure and model.
Daniel Bernoulli (Jakob’s nephew) introduced the concept of expected utility, recognizing that identical outcomes impact people differently: a thousand dollars means something vastly different to a billionaire than the average person.
This idea is essential to how we navigate risk. There are many other advances that have led humans to our present-day understanding of it, and computational power has dramatically accelerated this evolution. What Pascal and Fermat might have spent weeks or months calculating can now be processed in milliseconds.
Nevertheless, the reality is that only a few hundred years separate today’s risk assessment from decisions guided by superstition, blind faith, and instinct, all of which are influenced by our circumstances and beliefs.
Anticipation machines
Our neural architecture offers a biological template for understanding not just human decision-making but also the artificial systems we’ve created in our image. The leap from neurological risk processing to artificial intelligence is not metaphorical; it’s genealogical. When we build machines that “think,” we inevitably encode our own relationship with uncertainty.
At the core of mainstream modern AI is the phrase “attention is all you need,” the title of the foundational 2017 paper5 that introduced the transformer architecture and precipitated the rise of the large language model. To me, the title is a philosophical statement about how intelligent decision-making works. The core insight is that selective focus on relevant information is fundamental to understanding. Rather than processing everything with equal weight, attention mechanisms calculate interdependencies between elements and decide what matters most. This mirrors how humans navigate risk: through selective attention to relevant signals amid overwhelming noise.
Philosopher Daniel Dennett wrote of the human brain as an “anticipation machine.” Our most potent cognitive ability isn’t reasoning about but instead projecting what might be, playing the counterfactual game of “if this, then that” that allows us to let our theories die in our stead. We have now built artificial systems that mimic this capacity, encoding our anticipatory nature into algorithms that predict everything from the next word in a sentence to the next move in the market.
As an engineer and strategist in AI, I help create systems that algorithmically determine what deserves focus. As a model (not AI), I work in an industry explicitly designed to capture and direct attention. We are witnessing the emergence of an “attention society,” where the ability to direct and capture mindshare is a primary form of capital. Risk and reward increasingly revolve around what we attend to and what we ignore.
LLMs notoriously struggle with uncertainty, and their ability to engage with risk exhibits striking parallels to that ability in humans. Consider hallucinations in LLM outputs, where models generate plausible but factually inaccurate information. These errors occur when the model makes connections that appear statistically valid but lack grounding in reality. This is not unlike the human tendency toward confabulation. We have a remarkable ability to create coherent narratives from incomplete or contradictory information and often create elaborate explanations for events we barely understand.
In an ideal world, when we as humans are uncertain, we’d hedge and say “maybe,” accommodating multiple possibilities and admitting what we don’t know. AI training processes discourage this instinct, often penalizing models for spreading probability mass across many possible answers. High confidence is rewarded. Models don’t pick up the epistemic state of “I don’t know,” because they are optimized to commit.
Recent research on mechanistic interpretability reveals these systems don’t just mechanically predict next words; they develop internal frameworks for reasoning that often remain black boxes to us. This implies failure modes in these systems often mirror human cognitive biases rather than traditional software bugs. Unlike conventional programming errors that manifest consistently under specific inputs, AI failures can be contextual, probabilistic, learned—appearing only under particular circumstances that may be difficult to anticipate. The systems we’ve built to help us navigate uncertainty have imported our own blind spots.
But fear of the implications of powerful AI is an insufficient motivator: we need hope as well. The same dynamics that shape our personal risk decisions (fear versus hope, avoidance versus approach, calculation versus courage) now operate at a civilizational scale as we navigate the possibilities and perils of increasingly powerful anticipation machines. The question isn’t whether we can calculate all possible outcomes but whether we can maintain the courageous optimism necessary to guide these tools toward positive-sum outcomes. AI risk represents a paradigmatic shift in our relationship with uncertainty.
If industrialization has transformed us into a “risk society”6 where the distribution of hazards is as central as the distribution of goods, then the algorithmic revolution represents an equally profound transformation.
This requires a fundamental shift in how we think about risk, not as something to be eliminated by increasingly sophisticated calculations, but as the inevitable companion to progress. Just as personal growth requires leaving our comfort zones, technological advancement requires embracing possibilities we cannot predict. We can only work to ensure these systems are guided by values that transcend prediction, like compassion, justice, and the expansion of human potential.
Modeling courage
In continuous, hopeful service of these long-term goals, I’ve pursued careers in entertainment and technology. My work in both has taught me firsthand that attention is the foundation of not only social and material capital but also meaningful risk assessment.
The irony of working in industries dedicated to capturing and directing attention while writing about navigating risk to achieve mutually beneficial outcomes is not lost on me. As a model, I perform for the camera, knowing each image will be scrutinized in ways I cannot control. The fashion industry is infamously unpredictable; careers rise and fall based on changing tastes and trends no one can fully anticipate. Technology evolves at an equally dizzying pace. In this field, I navigate the technical and ethical complexities of systems designed to calculate risks and act on behalf of humans. I create systems that pick and choose what information matters.
No equation can tell me how to present myself publicly, what modeling contracts to accept, whether to work on an AI system with uncertain social impacts. Thriving in both worlds requires developing not just resilience against shocks but the capacity to grow from them.
When making decisions under uncertainty, I’ve learned to focus less on calculating unknowable probabilities and more on understanding potential consequences, asking: “If this goes wrong, can I live with the results? If this goes right, will it take me where I want to go?”
This perspective liberates us. Originally, I started this essay to explore why we limit ourselves, through negative self-talk, limiting beliefs, and self-inflicted isolation from others. I’ve come to believe that self-limitation often stems not from fear itself but from a misunderstanding of risk—treating uncertainty as something to be eliminated rather than navigated. The same risk assessment systems that evolved to protect and serve us become prisons when they prevent us from taking the leaps necessary for growth.
There is simply no path to excellence that does not pass through the territory of failure. The greatest risks I’ve taken—leaving a traditional path to enter modeling, working on emerging AI technologies, combining these ostensibly disparate worlds—have yielded the greatest rewards. Navigating uncharted territory, in career, in creativity, in life, requires acceptance of the unknown.
How do we move through uncertainty with purpose? Understanding risk mathematically is valuable; understanding the limits of that understanding is wisdom. In that space between calculation and courage, we find not just growth but the freedom to venture beyond what is known into what is possible. In this age of algorithms and anticipation machines, perhaps the most radical act is embracing the limits of calculation, to complement reason and logic and math with the courage to traverse the unknown. To model, in all senses of the verb, courage.