Introduction

Technological progress often feels inevitable in retrospect. We look at the current landscape of artificial intelligence, dominated by deep learning, neural networks, and large language models, and assume this was always the destination. But this view mistakes outcome for destiny. The AI we have today is not the product of inexorable forces marching toward a predetermined end. It is the accumulated result of thousands of choices: which problems to pursue, which methods to fund, which failures to abandon, and which successes to double down on.

Understanding these choices matters. It reveals that the path we're on is one of many possible paths, that the tools we've built reflect particular values and priorities, and that the future remains open to decisions we have yet to make.

The Symbolic Bet

In the early decades of AI research, the field made a defining choice: to model intelligence as symbol manipulation. This wasn't arbitrary. The pioneers of AI, figures like Allen Newell, Herbert Simon, and John McCarthy, were shaped by the intellectual currents of their time. Formal logic, linguistics, and the success of digital computers all pointed toward the same conclusion: thinking is computation, and computation is the rule-governed transformation of symbols.

This choice produced remarkable early successes. Programs that could prove theorems, play chess, and solve algebra problems suggested that general intelligence was within reach. Newell and Simon went so far as to declare that a physical symbol system possesses the "necessary and sufficient means" for intelligent action. The path forward seemed clear.

But this was a choice with consequences. By focusing on the kind of reasoning humans do consciously and deliberately, logic, planning, explicit problem-solving, symbolic AI neglected the vast substrate of cognition that operates below awareness. No one chose to ignore perception, motor control, or intuitive judgment; these simply seemed like trivial prerequisites rather than core challenges.

Moravec's paradox eventually exposed the cost of this assumption. The tasks that seemed hardest for humans, chess, calculus, formal reasoning, turned out to be relatively easy for computers. The tasks that seemed easiest, recognising a face, walking across a room, understanding a joke, proved maddeningly difficult. The symbolic approach had optimised for the wrong problems.

The Dreyfus Dissent

Not everyone agreed with the symbolic orthodoxy. Philosopher Hubert Dreyfus argued, beginning in the 1960s, that human expertise fundamentally depends on unconscious, embodied knowledge that cannot be captured in explicit rules. A chess master doesn't mentally enumerate possible moves; she sees the board in a way that highlights certain possibilities and obscures others. A doctor diagnosing a patient draws on pattern recognition honed through years of experience, not a decision tree of symptoms and diseases.

Dreyfus's arguments were initially dismissed, even ridiculed. This too was a choice, a collective decision by the AI community to close ranks against an outsider's critique. The consequences of this choice were significant. Had the field engaged more seriously with Dreyfus's objections, the turn toward connectionist and statistical approaches might have come sooner. Instead, symbolic AI continued to attract funding, talent, and attention through the 1970s and into the 1980s, even as its limitations became increasingly apparent.

The eventual vindication of Dreyfus's core insight, that much of intelligence is sub-symbolic, came not through philosophical argument but through empirical success. Neural networks, statistical learning, and the explosion of available data demonstrated that systems trained on examples could outperform systems programmed with rules. The field didn't abandon symbolic AI because it was persuaded by critics; it abandoned symbolic AI because something else worked better.

The Winter and the Thaw

The AI winters of the 1970s and late 1980s were themselves products of choice. Funding agencies, burned by unfulfilled promises, chose to redirect resources elsewhere. Researchers, facing diminished career prospects, chose to rebrand their work or leave the field entirely. Corporate labs chose to shut down long-term research programs that weren't delivering near-term returns.

These choices were understandable responses to disappointment, but they also had opportunity costs that are impossible to calculate. How many talented researchers were lost to other fields? How many promising lines of inquiry were abandoned prematurely? The AI winters weren't just periods of reduced activity; they were filters that selected for certain kinds of research and certain kinds of researchers.

When the thaw came, driven by increased computing power, the availability of large datasets, and algorithmic innovations like back-propagation, the field that emerged was different from the one that had entered winter. The statistical, sub-symbolic approach, once a minority position, became dominant. The researchers who had kept the flame alive during the lean years were vindicated. And a new generation, unburdened by the old debates, pushed forward with techniques that the founders of AI might not have recognised.

The Narrow Path

Perhaps the most consequential choice in recent AI history has been the focus on narrow intelligence over general intelligence. This choice was partly pragmatic: narrow problems have measurable benchmarks, publishable results, and commercial applications. General intelligence, by contrast, is difficult to define, harder to measure, and perpetually receding into the future.

But the choice was also ideological. The success of deep learning on specific tasks, image recognition, game playing, language modelling, created a gravitational pull toward problems amenable to these methods. Researchers optimised for what worked, and what worked was narrow AI. Funding followed success, and success was defined by narrow benchmarks.

The result is an AI landscape of extraordinary capability within constrained domains and persistent brittleness outside them. A system that can generate fluent prose may fail at simple arithmetic. A model trained on millions of images may be fooled by minor perturbations invisible to the human eye. These limitations aren't bugs to be fixed; they're consequences of the choices that produced these systems.

Some researchers argue that general intelligence will emerge from scaling narrow techniques, that enough capability across enough domains will eventually coalesce into something more unified. Others maintain that fundamental architectural innovations are needed, perhaps drawing on the symbolic approaches that were abandoned decades ago. The emerging field of neuro-symbolic AI represents one attempt to bridge these paradigms, to take the best of both approaches while avoiding their respective limitations.

The Explainability Trade-Off

Modern AI's success has come with a significant sacrifice: interpretability. A symbolic system, for all its limitations, can explain its reasoning. The rules are explicit; the logic is traceable. A deep neural network, by contrast, operates through billions of parameters whose individual contributions to any given output are effectively unknowable.

This trade-off was not explicitly chosen; it emerged from the accumulation of smaller choices. Each step toward greater capability through deeper networks and more parameters was also a step away from human understanding of what these systems were doing. The result is AI that can outperform humans on many tasks while remaining fundamentally opaque to human scrutiny.

The consequences of this opacity are still unfolding. When an AI system denies a loan application, approves a parole request, or recommends a medical treatment, the inability to explain why creates problems of accountability, fairness, and trust. These are not merely technical challenges; they are social and political challenges that the field chose not to prioritise in the pursuit of raw performance.

Choices Yet to Be Made

The history of AI is a history of choices, and the future will be no different. Today's researchers, policymakers, and technologists face decisions that will shape the trajectory of the field for decades to come.

Should resources be concentrated on scaling current approaches or diversified across alternative paradigms? Should safety research be prioritised alongside capability research, or will the market pressures of competition override such considerations? Should AI development be open and collaborative, or will competitive dynamics drive it toward secrecy and fragmentation? Who gets to make these choices, and whose interests will they reflect?

These questions have no obvious answers, and the choices we make will be shaped by values, incentives, and constraints that vary across individuals, institutions, and nations. But recognising that they are choices, rather than foregone conclusions, is itself a form of power. It opens space for deliberation, contestation, and ultimately, agency.

Conclusion

The AI we have today is contingent. It reflects the priorities of the researchers who built it, the incentives of the institutions that funded it, and the accidents of history that elevated certain approaches while marginalising others. This is not a criticism; all technologies are shaped by their contexts. But it is a reminder that other paths were possible and that other paths remain open.

The power of choice in AI development is both humbling and empowering. Humbling because it reveals how much of what we take for granted could have been otherwise. Empowering because it suggests that the future is not fixed, that the choices we make today will shape the AI of tomorrow. In a field moving as rapidly as artificial intelligence, that is a power worth taking seriously.