The "neural" in neural network was a marketing decision from the 1940s. As Gary Marcus argues in his critical appraisal, the resemblance to biological brains is superficial at best (Marcus, 2018). The analogy has been productive in some ways — it inspired architectures that genuinely work — but it misleads in consequential ways when taken literally. Understanding where the metaphor breaks down helps us reason more clearly about what these systems can and cannot do.
The Surface Similarity
Neural networks have "neurons" that "activate" and pass signals to each other through weighted "connections." Brains also have neurons connected by synapses. Both systems learn by adjusting connection strengths.
This is roughly where the similarity ends — and the differences are more instructive than the resemblance.
What Artificial Neurons Actually Do
An artificial neuron computes a weighted sum of its inputs, adds a bias term, and passes the result through a nonlinear function. It's a mathematical operation: multiply, add, squash.
output = activation(sum(weights * inputs) + bias)
This happens in a single forward pass. Input goes in, output comes out. No state, no memory, no temporal dynamics beyond what's explicitly encoded in the architecture.
A biological neuron is a cell with its own metabolism, capable of producing dozens of different neurotransmitters, with thousands of connections that vary in type and strength, operating with spike timing that carries information, and constantly rewiring based on activity. It's a complex adaptive system in itself.
What Learning Actually Is
In machine learning, "learning" means adjusting numerical parameters to minimize a loss function via gradient descent. The system receives training examples, computes how wrong its predictions are, and nudges parameters in directions that reduce wrongness.
This requires:
- An explicit objective (the loss function)
- Labeled examples
- The ability to compute gradients
- Many, many iterations
Biological learning involves:
- Multiple interacting systems (hippocampus, cortex, cerebellum...)
- Unsupervised pattern recognition
- Learning from single examples (sometimes)
- Sleep consolidation
- Emotional modulation
- Development over years
The word "learning" covers radically different processes. Using the same word obscures more than it reveals.
Why This Matters
It sets wrong expectations. If AI "learns like we do," then maybe it understands like we do. But next-word prediction is not understanding, even when the predictions are good — a point Bender and Koller make forcefully in their analysis of meaning and form (Bender & Koller, 2020). The reasoning by analogy is faulty because the underlying processes are different.
It misguides research. Neuroscience-inspired approaches have generated useful architectures (convolutional networks from visual cortex, attention mechanisms from cognitive attention). The successes came from borrowing loose inspiration, not from neural fidelity. Emerging work on World Models vs. Large Language Models explores whether fundamentally different architectures might capture aspects of cognition that current approaches miss — a productive research direction precisely because it takes the brain metaphor's limits seriously.
It confuses accountability. If AI "learns," did the developers "teach" it? Are they responsible for what it learned, the way teachers are responsible for what they teach? The metaphor imports assumptions that don't fit.
It enables anthropomorphism. Once AI "learns," the next step is that it "knows," then "decides," then "wants." Each step in the anthropomorphism ladder rests on the brain analogy. Pull the foundation, and the tower wobbles. The Autonomy Illusion traces where this ladder leads — and why it matters for accountability. The Two-Species Fallacy examines the most extreme version of this mistake: treating AI as a new kind of being rather than a new kind of tool.
What To Say Instead
Neural networks are pattern-matching systems trained on statistical regularities in data. They're powerful — remarkably so — but their power comes from optimization at scale, not from brain-like cognition. Recognizing this is not diminishing the technology. It's understanding it clearly enough to use it well and govern it wisely.
When you need to explain what AI does:
- "It finds patterns in data"
- "It's trained to predict likely continuations"
- "It optimizes for an objective"
- "It encodes statistical associations"
These descriptions are less evocative than "it learns" but more accurate. Accuracy matters when decisions follow — whether those decisions involve clinical deployment, creative work, or regulatory policy.
Related
- How Language Models Actually Work — What these systems actually do, step by step
- The Autonomy Illusion — Where the brain metaphor leads when applied to accountability
- The Training Pipeline — The actual process behind "learning"
- The Two-Species Fallacy — The most extreme form of the anthropomorphism problem
- World Models vs. Large Language Models — Alternative architectures that take different inspiration
- Stochastic Parrots — The debate over what language models actually understand
- What Fine-Tuning Really Does — How "learning" works in practice after pretraining
Sources
- Marcus, Gary. "Deep Learning: A Critical Appraisal." 2018. arXiv
- Lake, Brenden et al. "Building Machines That Learn and Think Like People." Behavioral and Brain Sciences, 2017. arXiv
- Mitchell, Melanie. "Why AI is Harder Than We Think." 2021. arXiv
- Bender, Emily & Koller, Alexander. "Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data." ACL, 2020. ACL