In a feedforward neural network, information moves in one direction—from the input through any hidden layers to the output—with no cycles or loops. This single-pass flow means each input is processed through a fixed sequence of computations to produce a result, which is ideal for tasks like image classification where you map image features to a class probability in one forward pass. For image classification, this structure is powerful because the network can learn hierarchical representations: early layers detect simple features, later layers combine them into more complex concepts, and the final layer assigns a class. While modern image classifiers often use convolutional structures to handle image data efficiently, those networks remain feedforward in that data flows forward through layers without feedback. The other terms don’t describe this forward-only architecture. The input layer is just the starting point, not the entire forward-pass design. A hidden layer is one component within many networks, not the overall architectural property. Association rules are a data-mining concept unrelated to neural network architectures.

Data poisoning is the risk where the data used to train or validate a model is tampered with to embed backdoors or biases. By injecting carefully crafted examples into the training set, an attacker can make the model learn hidden behaviors that only activate under specific triggers, effectively creating a backdoor. Poisoned validation data can also skew evaluation results and encourage biased or unsafe model behavior once deployed, because the model has learned from misleading signals. Other terms describe different issues: improper output handling concerns how outputs are managed or sanitized, unbounded consumption relates to resource or data ingestion limits, and system prompt leakage involves hidden instructions or prompts leaking into the model’s behavior.

When evaluating AI risk classes, a system that manipulates human behavior in a way that could cause severe harm is considered unacceptable risk. This category covers scenarios where the harm to people’s safety or autonomy is so serious that deployment is not justified, even with safeguards. Unacceptable risk means the system should not be allowed to operate. This differs from high-risk systems, which may be deployed only under strict controls and risk management; limited risk and low/no risk describe progressively less severe or easier-to-manage threats with lighter or no required safeguards. In the described case, the potential for severe harm from behavioral manipulation clearly places it in the unacceptable risk category.

This question is about modeling the probability of a binary event using a method that yields probability estimates based on several input features. The best fit is logistic regression. It takes a linear combination of the predictors, z = β0 + β1X1 + β2X2 + ... , and maps it through a sigmoid function to produce a probability: P(Y=1|X) = 1 / (1 + exp(-z)). The coefficients are typically estimated by maximum likelihood, which aligns the model’s predicted probabilities with the observed outcomes. Logistic regression is ideal for binary classification because its output is always between 0 and 1, making it easy to interpret as a probability. It works with multiple predictors (continuous or categorical, once encoded), and the coefficients tell you how each predictor changes the odds of the event. You can then choose a decision threshold (commonly 0.5, but adjustable) to assign class labels. In contrast, linear regression predicts a continuous outcome and can give values outside the 0–1 range, which isn’t meaningful for probability. K-means clustering is an unsupervised method for grouping data, not for predicting event probabilities. Support vector machines focus on creating a decision boundary; while probabilistic outputs can be obtained with additional calibration, SVMs aren’t inherently modeling probabilities in the straightforward way logistic regression does.