Will computers ever reach, or even exceed, human-level intelligence? And if they do, what form will that breakthrough take? Earlier this year, the Association for the Advancement of Artificial Intelligence (AAAI) posed this question directly to its members. Although neural networks dominate today’s AI landscape, most respondents were skeptical that neural networks alone could achieve true human-like cognition. Instead, they pointed toward a revival of an older tradition in the AI community — symbolic reasoning — as a necessary partner on the path to advanced intelligence.
Symbolic AI, sometimes referred to as “good old-fashioned AI,” operates on explicit rules and logical structures. Mathematics, if-then statements, programming languages such as Python, and visual tools like flowcharts and Venn diagrams all belong to this symbolic tradition. In the early decades of AI research, symbolic systems were viewed as the primary path toward machine intelligence. But around the early 2010s, they were overshadowed by the explosive progress of neural networks, which thrive on massive datasets and now power large-scale models such as ChatGPT.
Over the past few years, however, researchers have increasingly argued that neither approach can reach artificial general intelligence (AGI) alone. The field is now gravitating toward a fusion model — “neurosymbolic AI” — that aims to combine the strengths of both. Computer scientist Brandon Colelough of the University of Maryland has tracked this surge of interest, noting a dramatic rise in academic publications discussing neurosymbolic techniques beginning around 2021.
A growing number of scientists believe such hybrid architectures could help AI reason, generalize and make decisions more like humans. They also see strong potential for improving the reliability of AI used in high-stakes environments such as medicine, law, defense and critical infrastructure. Unlike neural networks, whose internal processes are often opaque, symbolic systems are interpretable and transparent, making them easier for humans to audit.
Some early examples of neurosymbolic systems are already demonstrating impressive capabilities. Google DeepMind’s AlphaGeometry, unveiled last year, reliably solves geometry problems at the level of global mathematics competitions. Still, designing a system that integrates symbolic reasoning with neural learning at scale remains a profound challenge. As William Regli, a computer scientist at the University of Maryland, puts it: “You’re really architecting a two-headed machine.”
A Longstanding Debate Resurfaces
The discussion around how much structure should be built into AI systems dates back decades. In 2019, Richard Sutton — a pioneer of reinforcement learning — reignited this controversy with a brief but influential essay titled “The bitter lesson.” Sutton argued that from the 1950s onward, one trend has repeated itself: approaches based on handcrafted symbolic rules consistently lose to methods founded on computation, data and search. Early chess engines crafted around human strategy, for example, ultimately performed worse than models that simply learned from vast gameplay records.
Advocates of large neural networks often cite Sutton’s essay to argue that scaling up data and model size is the most reliable path to AGI. But critics say this view glosses over the crucial role symbolic reasoning already plays in advanced systems. The current world champion chess program, Stockfish, combines a neural evaluator with a symbolic search tree, demonstrating that hybrid solutions can outperform either method alone.
Neural networks excel at recognizing patterns, generating language, and learning from examples, but they remain prone to factual errors, hallucinations and brittle reasoning outside the bounds of their training data. Symbolic systems, by contrast, are explicitly structured, good at logic and explanation, but struggle with ambiguity and the vast irregularities of natural language and everyday life.
These gaps become strikingly visible in real-world scenarios. Image-generation models sometimes draw people with extra fingers because they have not internalized the concept of how many fingers a human hand should have. Video models have difficulty simulating objects obeying physical laws such as gravity because those principles are not explicitly embedded in their training.
Some researchers see these errors as evidence that neural networks fundamentally lack the ability to generalize concepts or reason. They argue that symbolic scaffolding is the only practical way to give machines logical structure. Technology companies such as IBM have publicly embraced neurosymbolic AI as a strategic direction. Others, like Meta’s chief AI scientist Yann LeCun, remain unconvinced and argue that symbolic rules are incompatible with neural learning dynamics.
Hybrid Approaches Take Shape
Although “neurosymbolic AI” is a broad umbrella term, several methods have emerged that combine symbolic and neural ideas.
One strategy uses symbolic tools to improve neural networks. AlphaGeometry, for instance, trains its neural components on synthetic data generated from symbolic mathematical engines. This enables the system to learn consistent geometric relations and verify solutions more reliably. Other examples include “logic tensor networks,” which assign degrees of truth to symbolic statements, allowing neural networks to follow graded rules.
A second approach does the opposite: it uses neural networks to enhance symbolic engines. Classic symbolic search systems, such as those used in board-game algorithms, often face enormous search spaces. Neural models can narrow down the most promising pathways, greatly accelerating symbolic planning. This method underpinned AlphaGo’s landmark victory over human Go champions.
A third model embeds symbolic reasoning tools inside large language models (LLMs). Systems such as Program-Aided Language (PAL) convert natural-language prompts into Python code, use the code to compute an answer, and then translate the result back into text. This allows complex tasks such as mathematical reasoning to be handled more reliably than by language models alone.
Researchers are also applying neurosymbolic principles to robotics. Jiayuan Mao, an AI researcher trained at MIT, has shown that integrating symbolic reasoning into robotic perception allows robots to learn with far fewer examples. Instead of requiring hundreds of thousands of training samples, her approach can achieve high accuracy with a fraction of the data. In practical tests, neurosymbolic robots outperformed purely neural systems in tasks involving unfamiliar objects, such as sorting items or navigating household chores.
The Limits of Representing Human Knowledge
Despite the renewed enthusiasm, symbolic AI faces a major challenge: representing messy, inconsistent human knowledge in rigid logical rules. One of the most famous attempts was the Cyc project, which began in 1984. Its goal was to encode common-sense knowledge — everything from family relationships to emotional reactions — into a structured logical database. Cyc eventually accumulated more than 25 million rules and inspired modern knowledge graphs, including Google’s.
But human experience is filled with exceptions. Symbolic rules such as “parents love their children” or “seeing someone you care about makes you smile” hold true often, but not universally. Encoding such inconsistent truths into logic can lead to errors or produce brittle systems.
When researchers asked GPT-3 to generate Cyc-style logic statements, the model initially produced text that looked plausible. But upon inspection, the outputs contained major logical mistakes, reinforcing the difficulty of simply merging symbolic and neural systems without deeper integration.
Experts such as Colelough argue that the next step is to develop AI “metacognition” — systems capable of monitoring and coordinating their own reasoning processes. A robust orchestrator that seamlessly integrates symbolic logic with neural pattern recognition could, in his view, become the foundation of general intelligence.
What Comes Next
Achieving neurosymbolic AGI will not rely solely on new algorithms. Researchers anticipate the need for new chip designs optimized for hybrid operations, as current hardware is tailored mainly for neural networks. Future breakthroughs may also arise from entirely different paradigms, including quantum AI, which leverages quantum mechanical properties to accelerate computation.
For many researchers, the ultimate ambition is for AI systems to generate their own concepts, rules and symbolic structures — ones that humans have never formulated. Such systems might discover new scientific principles or mathematical constructs, expanding human knowledge rather than simply imitating it.
As Jiayuan Mao puts it, the long-term goal is for machines not just to learn from humans, but eventually to teach us: “We need to understand how computers can uncover new ideas that humans haven’t yet imagined.”
Leave a Reply