September 10, 2025 Technical Insight

Hallucinations are a Feature

Published by the Multiaxis Team

It's easy to be dazzled by neural networks. They generate text that sounds eerily human, draw images that seem imaginative, and even mimic our conversational style. But don't be fooled – beneath the surface, these systems are not "intelligent" in the way we are. At their core, neural networks are just glorified statistical pattern matchers.

Humans are natural storytellers. When we see technology give human-like responses, we instinctively project human-like behavior and reasoning onto it. That projection is misleading, and potentially dangerous.

These ideas build on insights from researchers and journalists exploring AI's limitations.

Neural Networks: Just Pattern Matchers

Neural networks excel at finding patterns in massive datasets. Sometimes they hone in with uncanny precision; other times, they latch onto entirely wrong correlations. The problem is, the patterns they compute are not legible to humans.

Unlike traditional algorithms, deep learning is a black box – a daisy chain of highly abstracted numbers. Even the researchers who build these systems cannot reliably explain why a model behaves a certain way, especially in edge cases (those unusual situations that fall outside normal training examples). This inscrutability is more than an academic problem; it's also a potential security vulnerability – attackers can exploit these blind spots to make AI systems behave unpredictably.

A Real Example: When Pattern Matching Goes Wrong

Here's a perfect example of how these "hallucinations" work. Ask an AI about a fictional academic paper – say, "Smith et al. (2019) on quantum effects in photosynthesis" – and it might confidently describe the paper's methodology, findings, and conclusions in convincing detail. The AI isn't lying; it's doing exactly what it was designed to do: matching patterns from millions of real academic papers to generate plausible-sounding text.

The model has learned that academic papers follow certain patterns: they have authors, dates, methodologies, and findings. When asked about a non-existent paper, it fills in these patterns with statistically likely content. It might tell you Smith's team used spectroscopy techniques, found evidence of quantum coherence at room temperature, and published in Nature Physics. All perfectly reasonable. All completely fabricated.

This isn't a malfunction – it's the system working as intended, finding and reproducing patterns without any understanding of truth or fiction.

Bias, Discrimination, and Scale

Because neural networks amplify statistical patterns, they also amplify bias. If there's even a trace of imbalance in the training data, the model will learn it and reproduce it at scale. This is why deep learning models so often exhibit discriminatory impacts: they are reflecting and magnifying flaws in their inputs.

The scale of these models only makes the issue worse. For perspective:

The human brain has roughly 100 trillion synapses.
GPT-3 has about 175 billion parameters.

That makes GPT-3 about a thousand times smaller than a human brain – and yet, even with this staggering size, the model can't do what humans do best: causal reasoning (understanding that A causes B, not just that they appear together).

The Limits of Deep Learning

Humans can generalize across contexts. Learn to drive in one city, and you can adapt to another. Neural networks can't. They don't understand causality – only correlations. That's why many researchers believe the future lies in neural-symbolic AI – systems that combine pattern recognition with explicit logical rules, like teaching a computer both to recognize chess pieces (neural) AND the rules of chess (symbolic).

Generative AI – models that turn old text into new text, or old images into new images – represents the maximalist form of deep learning. It is pattern matching at its most ambitious scale. In 2022, OpenAI demonstrated this by pouring unprecedented resources into scaling models: consuming massive datasets, leveraging enormous compute, and pushing energy use to new extremes.

But bigger isn't always better. Edge cases still break the illusion. The more these systems train on internet data, the more they absorb – including conspiracy theories, hate speech, and the peculiar vocabularies of niche online communities. A model trained on Reddit might suddenly start using obscure gaming slang in a business report, because it can't distinguish context the way humans do.

Anthropomorphism and the AGI Mirage

Human psychology betrays us. The moment something talks back, we're tempted to call it intelligent – even conscious. This anthropomorphic projection (attributing human characteristics to non-human things) is hardwired into our brains. OpenAI's advancements in deep learning have led many to believe we're on the cusp of artificial general intelligence (AGI).

The reality is far less dramatic. These models don't "understand." They mimic patterns. The analogy to intelligence is an anthropomorphic projection – not a scientific truth. Many deep learning proponents once believed that scaling would erase the gap between humans and machines. In fact, the gap has in some ways widened, with unreliability and unpredictability becoming more obvious at scale.