How Does AI Make Decisions?

Inference: The Decision-Making Phase

When an AI system makes a decision, it is performing inference. Inference is the opposite of training: instead of adjusting parameters to learn from data, the model uses its frozen parameters to process a new input and produce an output. The parameters do not change during inference. The model applies what it has already learned.

A trained image classifier receives a photo and passes it through its layers. Each layer extracts increasingly abstract features: edges in the first layers, textures in the middle layers, object parts in the later layers. The final layer outputs a probability for each possible class. If the model outputs 0.92 for "golden retriever," 0.05 for "labrador," and 0.03 for all other classes, the decision is "golden retriever" with 92% confidence.

For language models, inference is more complex because the output is not a single classification but a sequence of tokens. The model generates one token at a time, each time computing a probability distribution over the entire vocabulary (typically 50,000 to 100,000 tokens) and selecting the next token. This process repeats until the model produces a stop token or reaches a maximum length.

Probability Distributions and Softmax

Almost every AI decision involves a probability distribution. The raw output of a neural network's final layer is a set of numbers called logits, one per possible output class. These logits can be any real number, positive or negative, large or small. To convert them into probabilities, the model applies the softmax function, which exponentiates each logit and normalizes so they all sum to 1.

The softmax function has an important property: it amplifies differences. If the logits are close together, the probabilities will be relatively uniform (the model is uncertain). If one logit is much larger than the rest, softmax will assign it a probability near 1 and everything else near 0 (the model is confident). This is why AI systems can appear very confident even when they are wrong, a phenomenon that causes significant problems in safety-critical applications.

Decision Boundaries

For classification tasks, the model creates decision boundaries in the feature space, invisible surfaces that separate one class from another. A linear classifier creates flat boundaries (hyperplanes). A neural network creates curved, complex boundaries that can wrap around clusters of data points.

Imagine a two-dimensional feature space where one axis is "email length" and the other is "number of links." A spam classifier learns a boundary in this space. Emails on one side are classified as spam, emails on the other side as legitimate. The boundary is not a single line but a complex surface shaped by the training data. Emails near the boundary are classified with low confidence, while emails far from the boundary are classified with high confidence.

In reality, the feature space has thousands or millions of dimensions, not two. The decision boundaries exist in this high-dimensional space, and they are impossible to visualize directly. But the principle is the same: the model partitions the space into regions, and classification is a matter of determining which region the input falls into.

How Language Models Choose Words

Language models do not make a single decision per input. They make thousands of sequential decisions, one for each token in the output. Each decision depends on everything that came before it.

When a language model generates text, it computes a probability distribution over its vocabulary at each step. The word "the" might have probability 0.15, "a" might have 0.08, and thousands of other words have smaller probabilities. The model then selects the next token based on this distribution.

The selection method matters enormously. Greedy decoding always picks the highest-probability token. This produces coherent but repetitive, boring text. Temperature sampling scales the logits before applying softmax: a temperature below 1.0 sharpens the distribution (more deterministic), while a temperature above 1.0 flattens it (more random). Top-k sampling restricts the selection to the k most probable tokens. Top-p (nucleus) sampling restricts to the smallest set of tokens whose cumulative probability exceeds p.

This is why the same prompt can produce different outputs each time you use a language model. The model is not uncertain about the "right" answer, it is sampling from a distribution, and different random seeds produce different samples. The text sounds authoritative, but it is the output of a probabilistic process, not a deterministic retrieval of facts.

Confidence and Calibration

An AI model's stated confidence (its output probability) is not always reliable. A well-calibrated model's confidence matches reality: when it says 90% confident, it should be correct 90% of the time. Many models are poorly calibrated, especially large neural networks, which tend to be overconfident.

Research has shown that modern neural networks are often more overconfident than smaller, simpler models. A neural network that outputs 99% confidence might only be correct 85% of the time. This overconfidence is dangerous in medical diagnosis, autonomous driving, and other high-stakes applications where humans rely on the model's confidence to decide whether to trust its output.

Calibration techniques exist. Temperature scaling adjusts the sharpness of the output distribution after training to better match actual accuracy. Platt scaling fits a logistic regression to the model's outputs to produce calibrated probabilities. Ensemble methods combine predictions from multiple models, and the diversity of their outputs provides a more reliable uncertainty estimate.

AI Decisions vs Human Decisions

AI decision-making differs from human decision-making in several fundamental ways.

AI is consistent. Given the same input and the same random seed, an AI model produces the same output every time. It does not have good days and bad days, does not get tired, and is not influenced by mood, hunger, or personal bias (though it may reflect biases in its training data). This consistency is valuable in applications where fairness requires identical treatment of identical cases.

AI cannot explain its reasoning in the way humans can. A neural network that classifies a tumor as malignant cannot point to the specific features that led to its decision in a way that matches how a radiologist would explain the same diagnosis. Techniques like attention visualization, saliency maps, and SHAP values provide partial explanations, but they show correlations, not the causal reasoning process that human experts use.

AI does not understand consequences. A loan approval model does not understand that denying a loan affects someone's ability to buy a home. It computes a probability of default based on statistical patterns. The decision to deny the loan, and the ethical weight of that decision, exists only in the human system that surrounds the model.