If you look at the history of AI research over the past 70 years, you’ll find a lot of flip-flopping between two major approaches: symbolic AI and neural AI.

Symbolic AI tries to formulate explicit rules. It excels at mathematical and logical tasks, and it’s at the heart of most traditional computer software - any software application can be regarded as a kind of machine intelligence, and most programming languages are symbolic by nature.

Neural AI learns implicit rules from data on its own. It excels at pattern recognition, which makes it good for tasks like image classification and text generation. This is why it’s at the heart of generative AI tools like large language models.

Where does language fit in AI?

The key battleground in AI development today is essentially the question: where does language understanding fit in this picture? Should we think of it as being more like maths and logic, or more like pattern recognition? And what does that mean for the tools we should use?

To understand the difference, think about two fundamentally different kinds of data.

Two kinds of data

First, imagine data that follows a regular pattern - a smooth curve where every point sits exactly on the line.

If I gave you a new data point and told you only its position on one axis, you could place it perfectly on the chart. You’d be able to do this despite having incomplete information because the data follows a rule. And with that rule, whenever a new data point comes in, you can always be confident about where it’s going to land.

Now imagine data that follows a noisy pattern - a broad S-shape, but with scatter.

You can still formulate a rule that draws an approximate line through the data, but when you make a prediction, you’re going to have a margin of error.

The error is just the difference between where your line is and where the data points actually are - the difference between expectation and reality.

Two kinds of AI

With regular patterns, symbolic AI excels. The whole approach is about formulating rules to describe simple regularities which can combine together to create complex ones - patterns within patterns.

With noisy patterns in high dimensions, neural AI excels. Language models plot words in a space with thousands of dimensions, where distance represents similarity of meaning. There’s no practical way to formulate explicit rules for these surfaces, so neural models learn them implicitly through feedback.

So we’ve got two kinds of data and two kinds of AI suited to each. The question is what happens when we use the wrong tool for the job.

The wrong tool for the job

A model can score well on the wrong kind of data and still be fundamentally lost. Apply a neural model to data with a clean underlying rule and it will approximate the rule well enough to seem like it understands it, until the edge cases arrive and the margin of error starts to matter. Apply symbolic AI to genuinely noisy data and it will either fail outright or overfit to noise it mistakes for signal.

This is the real argument for neurosymbolic AI: not as a compromise between two historical camps, but as a recognition that real-world problems rarely arrive pre-sorted. Language in particular straddles both kinds of data, it has deep logical structure and irreducible statistical noise, and a system built for only one will always be missing half the picture. Getting the balance right isn’t just an engineering preference. It’s the difference between a model that performs well under controlled conditions and one that can be trusted when conditions change.