Letter H

An AI answer that sounds plausible but is false, unsupported or invented.

Language models generate likely continuations, not guaranteed facts. They may invent names, quotations, links or events, especially when asked about uncertain or missing information.

#

A layer inside a neural network between the original input and the final output.

It is called hidden because people do not directly supply or read its values. Hidden layers build intermediate representations: in an image system, early layers might notice edges while later ones combine those clues into shapes and objects.

#

People judging the quality, usefulness or safety of an AI system's output.

Human evaluation is useful when there is no single automatic test for qualities such as clarity, humour or helpfulness. Results depend on clear instructions and suitable reviewers; people can disagree or bring their own biases.

#

A design where a person reviews, guides or approves part of an AI process.

Human involvement is especially important when errors could affect money, rights, safety or reputation. It only helps when the reviewer has enough time, information and authority to intervene.

#

A setting chosen for the training process rather than learned by the model itself.

Parameters are the model's learned notes; hyperparameters are the teacher's choices about how learning should happen. Examples include learning speed, batch size and number of training passes.

#

Trying different training settings to find a combination that works well.

It is like adjusting oven temperature and cooking time before settling on the best recipe. The settings are not learned as ordinary model parameters, so developers test alternatives and compare their performance using separate validation data.

#