Neuro-symbolic AI

In AI research there have been two very different basic ideas for decades. One idea: you give the computer fixed rules, such as “Every human is a living being.” Such systems compute strictly logically and justify every step in a comprehensible way. The other idea: you show the computer millions of examples, and it recognizes patterns in them on its own. These learning systems are astonishingly flexible, but they don’t explain why they arrive at an answer. Neuro-symbolic AI combines both approaches in one system and thereby aims to harness the strengths of both.

The problem with guessing

Today’s language models are at their core probability machines. They predict which word is most plausible to come next. This works excellently for text production, but it provides no guarantee of correctness. That’s why such models occasionally invent facts that sound convincing yet are false. In technical terms this is called hallucination.

In many areas this is unacceptable. A bank must be able to prove why a loan was denied. A drug development program must not violate basic chemical rules. A tax program must apply the law exactly, not approximately. Rules have a decisive advantage here: they are verifiable and traceable.

Pure rule-based systems, however, fail in the face of messy reality. No one can write down all the rules that would be needed to recognize a cat in a photo. Learning systems solve exactly this task effortlessly. The combination of both approaches is therefore not a passing trend, but an attempt to offset two well-known weaknesses against each other.

Division of labor between network and rule set

Usually the learning part takes over perception. It turns images, sounds, or sentences into clear building blocks. From a photo, for example, comes the statement: “a red cube on the left, a blue sphere on the right.” These building blocks are symbols, i.e. unambiguous labels for things and their relationships.

The second part, often a logic engine or a program, then works with these symbols. It checks rules, draws conclusions, and computes exactly. To the question “Is there something red on the left?” it does not answer by guessing, but by looking it up. The result can be traced back step by step.

A vivid image is a team made up of an eyewitness and a judge. The eyewitness quickly recognizes what is visible, but can be mistaken. The judge knows the rules and checks whether the conclusion is permissible. What matters is that this division only works if the handover goes cleanly. If the learning part describes the scene incorrectly, the logic afterward works flawlessly with incorrect information.

Where hybrid systems run today

A well-known example is mathematics systems from Google DeepMind. The program AlphaGeometry solved problems from the International Mathematical Olympiad in 2024. A language model suggests possible auxiliary lines, while a classical proof program rigorously checks the steps. Neither the creative nor the verifying part could have managed this alone.

You also encounter the principle in everyday life without hearing the term. When a chatbot passes a calculation to a calculator or queries a database, that is a simple form of it. Voice assistants translate your question into a formal search query. Calendar and booking systems then check hard rules such as opening hours or available slots.

In the news, the term often comes up in connection with regulation. Authorities are increasingly demanding that automated decisions be explainable. A common misconception here is that neuro-symbolic AI is a step backward to the time before language models. In fact, it is about complementing them: the learning model stays, but is given a rule set alongside it.