This is a crosspost of my article for The Print
Artificial Intelligence did not begin with code—it began with a question. Could machines think? And if so, how would we even know?
In 1950, Alan Turing proposed that if a machine could carry on a conversation indistinguishable from a human, it could be called intelligent. This became the Turing Test, and it marked the philosophical beginning of AI.
The technical beginning followed six years later, at the Dartmouth Workshop of 1956. Organized by John McCarthy, Marvin Minsky, Claude Shannon and others, it launched AI as a formal discipline. The claim was breathtaking: “Every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” For a while, that dream held.
The 1960s and 70s saw AI become a fixture of science fiction. Stanley Kubrick’s 2001: A Space Odyssey imagined HAL 9000, a machine that could speak, reason, and feel—until conflicting objectives caused it to turn rogue. HAL’s breakdown wasn’t madness—it was logic stretched to a breaking point. And that remains one of AI’s deepest warnings: machines may fail not because they malfunction, but because their goals are misaligned with ours.
From the 1960s to the 1980s, Symbolic AI dominated the field. Intelligence was programmed through logic and rules, not learned from data. Expert systems like MYCIN and DENDRAL mimicked human specialists and briefly dazzled funders, but they were brittle—struggling with ambiguity and real-world complexity. Each new scenario demanded new rules, revealing the limits of hand-coded intelligence.
The initial optimism faded. Early successes didn’t scale, and by the 1970s and again in the late 1980s, AI faced its winters—eras of disillusionment and vanishing support. The technology wasn’t ready. AI, once hailed as revolutionary, became a cautionary tale.
Meanwhile, the world of chess provided a battleground for AI’s ambitions. In 1968, computer scientist John McCarthy bet that no machine could beat chess master David Levy in a match within ten years. He was right—but only just. By 1997, IBM’s Deep Blue defeated Garry Kasparov, the reigning world champion. This wasn’t intelligence in the human sense. Deep Blue didn’t think; it calculated—200 million positions per second, guided by rules and brute force.
If Deep Blue marked a brute-force triumph, the next revolution came from inspiration closer to biology. Our brains are made of neurons and synapses, constantly rewiring based on experience. In 1943, McCulloch and Pitts proposed the first mathematical model of a neural network, mimicking how neurons fire and connect. Decades later, with more data and computational power, this idea would explode into what we now call deep learning.
A key moment came in 2012. Researchers at Google Brain fed a deep neural network 10 million YouTube thumbnails—without labels. Astonishingly, one neuron began to specialize in detecting cat faces. The machine wasn’t told what a cat was. It discovered “cat-ness” on its own. This was the cat moment—the first clear sign that neural networks could extract meaning from raw data. From then on, deep learning would take off.
That same year, another milestone arrived. AlexNet, a deep convolutional neural network, entered the ImageNet Challenge, a global competition for visual object recognition. It halved the previous error rate, using an 8-layer network trained on GPUs. This marked the beginning of AI’s rise in vision—powering facial recognition, self-driving cars, and medical diagnostics.
In board games too, AI moved from mimicry to mastery. AlphaGo’s match against world Go champion Lee Sedol in 2016 stunned experts. Game 2, Move 37—an unconventional, creative move—changed the game’s theory forever. AlphaGo didn’t just compute; it improvised. In 2017, AlphaZero went further, mastering chess, Go, and shogi without human examples—just the rules and millions of self-play games. Grandmasters called its style “alien” and “beautiful.”
In 2017, the landmark paper “Attention Is All You Need” introduced the Transformer architecture, a breakthrough that changed the course of AI. Unlike earlier models, Transformers could handle vast contexts and relationships between words, enabling a deeper understanding of language patterns. This paved the way for large language models (LLMs) like GPT and ChatGPT, trained on billions of words from books, websites, and online conversations. These models don’t know facts as humans do—they predict the next word based on learned patterns. Yet their output is often strikingly fluent and, at times, indistinguishable from human writing.
These models don’t understand language the way we do. They predict the next word based on probabilities. And yet, their output often sounds thoughtful, even profound. In 2025, one such model helped save a pregnant woman’s life by identifying a symptom of preeclampsia from a casual health question. This was no longer science fiction. AI was here helping, guiding, even warning.
This is where the story darkens. Neural networks have millions—even billions—of internal parameters. We know how they are trained, but not always why they produce a particular result. This is the black box problem: powerful models we can’t fully interpret.
Worse, these models inherit biases from the data they are trained on. If trained on internet text that contains racial, gender, or cultural prejudices, the model may echo them—sometimes subtly, sometimes dangerously. And because their reasoning is opaque, these biases can be hard to detect and even harder to fix.
AI systems are also confident liars. They “hallucinate” facts, produce fake citations, or reinforce misinformation—often with grammatical precision and emotional persuasion. They are trained to be convincing, not correct.
As we hand over more decisions to machines—medical diagnoses, hiring recommendations, bail assessments, autonomous driving—we face hard questions: Who is responsible when an AI system fails? Should a machine ever make a life-or-death decision? How do we align machine goals with human values?
The fictional HAL 9000 chose its mission over its crew, not out of malice, but from a conflict of objectives. Today’s systems don’t “choose” at all, but they still act, and their actions have consequences. Ironically, the most hopeful vision may lie in chess again. In freestyle tournaments, the best performers weren’t machines or grandmasters—but human-AI teams. Garry Kasparov put it best: “A weak human + machine + good process beats a strong human or strong machine alone.” AI doesn’t need to replace us. It can enhance us—if we build it thoughtfully, interpret it critically, and embed it in processes we trust.
