Could Thinking in Graphs be the Next Step for AI?
AI sees, but does it understand?
Artificial intelligence has become astonishingly good at perception. It can describe a photo, summarize a document, and even generate a symphony of pixels from a sentence. The models behind these feats (convolutional networks, transformers, diffusion systems) have mastered how to recognize patterns. They see, they mimic, they predict. But they do not yet understand.
When a large language model writes a paragraph, it does not reason about the relationships between ideas; it maps one word to the next based on probability. A vision model can label a cat, a tree, or a human face, yet it has no awareness of how those objects interact. The world, however, is built on relationships. Cells communicate within tissues, traders influence markets, friends share information across social networks.
The question now facing the field is simple. What happens when machines begin to learn relationships instead of just patterns?
For decades, AI’s architecture has followed the shape of its data.
In the 2010s, convolutional networks conquered two-dimensional grids of images.
In the 2020s, transformers dominated sequential data like words, audio, code, and even video.
But reality is neither a grid nor a sequence. It is a graph.
That realization has started a new wave of research. It first emerged in niche academic corners such as chemistry labs, recommendation engines, and fraud-detection systems. It is now influencing how entire industries think about intelligence.
At its core, the premise is simple. If you want machines to reason about the world, they must first learn its structure.
The World Is a Graph
Think about almost any system you interact with, whether a social network, a payment system, a protein chain, a supply route, or even a classroom. Beneath the surface, they all share the same shape. They are nodes connected by relationships.
You are a node in a social graph linked by friendships and communication. Your bank account connects to others through transactions. A single protein interacts with thousands of others, forming the basis of life itself. Every meaningful process in nature and society can be described as a graph.
Graphs are not new. They have shaped fields from epidemiology to Google’s search algorithm. What is new is that neural networks can now learn from them. Traditional models flatten data into tables or sequences, losing the geometry of how things relate. Graph Neural Networks, or GNNs, preserve that structure. They do not just process data points; they process the connections between them.
A GNN treats information like a conversation among neighbors. Each node carries a small vector of knowledge and updates itself by listening to what the surrounding nodes are saying. Over many rounds, these exchanges spread context through the network until the model forms an understanding of the overall structure.
This seemingly simple mechanism is powerful. Where a transformer sees a list of tokens and a convolutional network sees a grid of pixels, a GNN sees structure. It can infer that two distant nodes behave similarly because they share a chain of intermediaries. It can detect clusters, anomalies, hierarchies, and dependencies, not because it was told to, but because those relationships emerge naturally from the graph’s topology.
That ability to reason (that is, probabilistically infer) through structure rather than surface features is what makes GNNs one of the most promising frontiers in artificial intelligence. They represent a shift from machines that only recognize patterns to machines that begin to interpret them.
Where This Starts to Matter
The power of graph learning becomes clear when it meets real data. In our own work, we trained a graph model on millions of financial transactions, where every account was treated as a node and every payment as a connection. The goal was not only to flag fraud or predict risk but to see whether a model could uncover the hidden structure of how money actually moves.
Once the data was arranged as a graph, familiar patterns began to change. What looked random in a spreadsheet turned into clear structures. Some accounts acted as quiet terminals while others became hubs of constant activity. Clusters of nodes moved in synchrony, small transfers rippling through dozens of accounts before converging somewhere unexpected. The network revealed a living system that a flat dataset could never show.
Graph Neural Networks are designed for this kind of discovery. They learn how a node’s behavior depends on its neighbors and how influence spreads through the system. A capable model can see that two accounts with no direct link still share a deeper connection several steps away, or that a burst of small transactions across many accounts may not be noise at all but a coordinated pattern. It learns the difference between coincidence and structure.
In practice, the signal varied by dataset. On IBM’s 24-million-transaction corpus, fraud was about 0.12 percent. A hand-built edge index reached roughly 8 GB, and training took hours, yet a simple multilayer perceptron on raw features matched or beat the graph model. We pivoted to DGraph-Fin with about 3 million nodes and 4 million edges. There, a relational GCN (Graph Convolutional Network) that models edge and node types outperformed a plain GCN by about 5 percent in accuracy and 4 percent in recall under a 30-minute budget with early stopping.
The difference was clear. Both models performed well, but the edge- and node-based approach consistently produced higher accuracy across every test. That improvement suggested that focusing on the relationships rather than the individual data points allowed the network to capture the true structure of the system.
That trade-off between local detail and global understanding sits at the center of artificial intelligence. Models that look too narrowly can predict the next action but often miss the larger story. Models that look too broadly can see structure but lose sensitivity to nuance. Graph learning offers a way to connect the two, giving machines the ability to understand both the parts and the whole, the pattern and the context that give them meaning.
The Industry Shift
What began as a research idea is now shaping some of the world’s largest systems. DeepMind used graph reasoning in AlphaFold to predict how proteins fold, one of biology’s hardest problems. Google applies similar ideas across search, recommendations, and mapping to model relationships between billions of entities. Visa and Mastercard rely on graph learning to trace transaction flows in real time, catching coordinated fraud that rule-based systems miss.
This quiet migration of graph reasoning into production marks a shift in how AI is built. Scaling performance is no longer only about bigger models or more data; it is about structure. Graph learning gives systems a sense of how things relate, allowing them to reason about influence and flow instead of treating every input as independent.
Across industries, the same pattern is emerging. In logistics, GNNs optimize routing and supply networks. In healthcare, they help identify drug targets by modeling molecular interactions. In social media, they reveal communities and information cascades. Each example points to the same idea. Models grow more powerful when they learn how their data connects.
The Future of Hybrid and Reasoning Models
Artificial intelligence is beginning to move beyond perception and prediction toward genuine reasoning. The aim is shifting from models that only recognize patterns to those that can understand relationships, draw inferences, and adapt to new situations. Graph learning represents a step in that evolution because it gives models an internal sense of how the pieces of a system fit together.
Researchers are now exploring how to combine that relational understanding with the expressive power of transformers. Early hybrid models such as Graphormer and GraphGPT suggest that attention mechanisms and graph reasoning can reinforce each other. These prototypes are still at the research stage, but they point toward architectures capable of processing text, images, and structured data within a single framework. A model that sees both the entities and their relationships could move from describing data to explaining it. It might read a financial report by mapping how firms and markets interact, or interpret a molecular structure by reasoning over bonds rather than pixels.
This direction highlights a broader truth about progress in machine learning. For much of the past decade, advances have come from scaling, with larger datasets, more parameters, and greater computational power. But real understanding may depend less on size and more on structure. Intelligence is not only about how much information a model can store but about how effectively it organizes and connects that information.
Conclusion
If deep learning taught machines to see and speak, graph learning may teach them to understand. It gives artificial intelligence the ability to connect context with content and meaning with structure. The next decade of progress will not be defined only by scale but by how well models can represent the relationships that shape the world they are trying to learn.
Understanding is not a matter of memorizing more data but of seeing how the pieces fit together. In that sense, the future of intelligence, both human and artificial, may depend on the same principle that defines every system we study. The world is a network of relationships, and to truly know it, a model must learn to think in graphs.