Loading
A neural network is a machine learning model made up of interconnected processing units, or "neurons," arranged in layers that transform input data through weighted connections learned from training data. ISO/IEC 22989:2022, the international standard for AI terminology, defines it as a network of one or more layers of neurons connected by weighted, adjustable links that takes input data and produces an output, adding that while early designs were loosely inspired by biological brains, most modern neural network research no longer follows that biological analogy. Neural networks are the architectural foundation of deep learning, and one specific design, the transformer, introduced by researchers at Google in 2017, now underlies the large language models, image generators, and other generative systems that make up today's frontier AI. For governance teams, the significance is structural: a trained network's behaviour is encoded across millions to billions of numeric weights rather than human-readable rules, which is exactly why explainability and interpretability remain standing regulatory and audit concerns rather than solved problems.
Run the free AI Health CheckNeural Network, a class of machine learning model loosely inspired by biological neural systems, composed of interconnected processing units (neurons) organised in layers that transform input data through learned weights.
Neural networks are the architectural foundation of deep learning, and transformer-architecture neural networks with very large parameter counts underpin well-known systems such as OpenAI's GPT-3 (175 billion parameters) and Meta's Llama 3.1 (405 billion parameters). That scale is well-documented for models whose specifications were disclosed, but it is not a safe generalisation to all frontier AI: OpenAI's GPT-4 Technical Report withheld parameter count and architecture details "given both the competitive landscape and the safety implications," and no current leading proprietary frontier model has an officially disclosed parameter count. From a governance perspective, neural networks are typically opaque (the "black box" property); NIST's AI Risk Management Framework 1.0 lists "explainable and interpretable" as one of seven characteristics of trustworthy AI, directly targeting this concern.
Source: ISO/IEC 22989 (technical definition); NIST AI 100-1 / AI RMF 1.0 (trustworthy AI characteristics); Brown et al. 2020 (GPT-3); Meta (Llama 3.1 release); OpenAI GPT-4 Technical Report
A neural network is organised into an input layer, one or more hidden layers, and an output layer. Each connection between neurons carries a weight, a number that scales the strength of the signal passed forward; each neuron typically applies an "activation function" that determines whether and how strongly it fires before passing its result to the next layer. Under ISO/IEC 22989's definition, this is precisely what a neural network is: layers of neurons connected by weighted links, with adjustable weights, taking input data and producing an output.
Training is the process of adjusting those weights so the network's outputs get closer to a desired result. In supervised training, the network is shown many labelled examples, checks its output against the correct answer, and uses an algorithm called backpropagation to nudge every weight slightly in the direction that reduces the error, repeated over millions of examples. Once training is complete, the weights are fixed, and running new input data forward through them ("inference") is what produces a live prediction, classification, or generated response.
Neural networks are not new: the perceptron, a single-layer network, dates to 1958, and multi-layer "deep" networks trained with backpropagation became practical in the 1980s. Through the 2010s, convolutional neural networks (CNNs) dominated image tasks and recurrent neural networks (RNNs/LSTMs) dominated sequential tasks like text and speech, processing data step by step.
That changed in 2017, when Google researchers published "Attention Is All You Need," introducing the transformer, an architecture built entirely around a self-attention mechanism that lets a model weigh the relevance of every element in a sequence to every other element simultaneously, rather than one step at a time. Because this is far more parallelisable on modern hardware, it became practical to train dramatically larger models on much more data, and the transformer rapidly became the standard architecture behind large language models.
Publicly confirmed examples show the scale involved: OpenAI's GPT-3 (2020) used 175 billion parameters, and Meta's Llama 3.1 (2024) uses 405 billion parameters in a decoder-only transformer. That said, precise parameter counts for today's leading proprietary frontier models are largely unconfirmed, OpenAI's GPT-4 Technical Report explicitly states it discloses no further detail on model size "given both the competitive landscape and the safety implications of large-scale models" and neither Anthropic nor Google has published exact parameter counts for their current flagship systems.
The black-box problem
A trained network's logic is distributed across its weights rather than expressed as explicit rules, so even full access to a model's parameters doesn't guarantee a human-readable explanation of any single output.
Explainability and interpretability duties
NIST's AI Risk Management Framework lists "explainable and interpretable" as one of seven characteristics of trustworthy AI, aimed directly at the opacity that neural network architectures create.
Training data provenance
A network's outputs reflect patterns in whatever data it was trained on, so biased, unrepresentative, or unlawfully sourced training data can propagate into every downstream decision the model makes.
Model drift
Because a network's weights are fixed once training ends, performance can silently degrade as real-world inputs drift away from the distribution the model originally learned from.
Compute and provider concentration
Training transformer models with hundreds of billions of parameters requires compute and data resources concentrated among a small number of organisations, a structural factor in frontier AI market and supply-chain risk.
What is a neural network in simple terms?
A neural network is a computer model that processes information through layers of connected "neurons." Each connection carries an adjustable numeric weight, and data flows forward through the layers, being transformed at each step, until the final layer produces an output, such as a classification, a prediction, or a piece of generated text.
What's the difference between a neural network and deep learning?
"Deep learning" refers to training neural networks with many stacked layers rather than the one or two layers used in early designs. Every deep learning model is a neural network, but not every neural network is "deep", a network with a single hidden layer is still a neural network, just a shallow one.
What is a transformer, and how is it different from an ordinary neural network?
A transformer is a specific neural network architecture introduced in 2017 by Google researchers in the paper "Attention Is All You Need." It uses a self-attention mechanism to weigh how relevant every element in an input sequence is to every other element, processing the whole sequence in parallel rather than step by step as older recurrent networks did. That parallelism is what made training today's largest language models practical.
How many parameters do today's AI models actually have?
It varies by model, and for the largest proprietary systems it usually isn't disclosed. Confirmed public figures include OpenAI's GPT-3 at 175 billion parameters (2020) and Meta's Llama 3.1 at 405 billion parameters (2024). OpenAI's GPT-4 Technical Report, by contrast, explicitly declined to state a parameter count, and none of today's leading proprietary frontier models have an official published figure.
Why do people call neural networks a "black box"?
Because a trained network's behaviour lives in the numeric values of its weights, sometimes hundreds of billions of them, rather than in explicit, human-readable rules. It is possible to inspect every weight in a model and still be unable to state in plain language why it produced a specific output, which is why explainability and interpretability are listed as core trustworthy-AI characteristics in frameworks like the NIST AI Risk Management Framework.
Are neural networks themselves regulated?
No law regulates "neural networks" as a technology in the abstract. Regulation such as the EU AI Act attaches to the AI systems built on them, based on risk level and use case, while standards bodies like ISO/IEC and NIST supply the shared technical vocabulary, including the definition of "neural network" itself, that regulators, auditors, and standards increasingly reference.
Last reviewed July 2026
This page is general information about What Is a Neural Network?, not legal, regulatory, or professional advice, and does not capture every nuance or exception. Requirements change and can be fact-specific. Always verify against primary sources and your own qualified legal counsel before relying on it.