AI Glossary

A variant of the attack in which malicious instructions are not entered directly by the user but are hidden in content that the model processes (a website, document, or email), and the model incorrectly interprets them as a command.

A specific type of model tuning based on training responses to users' prompts and instructions.

An attempt to bypass a model's safety restrictions using specially crafted prompts so that it produces content it would otherwise refuse.

An instruction given by a user to an artificial intelligence system in order to produce an output.

The storing of repeatedly used parts of an input in order to speed up and reduce the cost of subsequent queries to the model.

A term for the legal questions surrounding the copyright protection of prompts.

Methods of creating and optimizing instructions that lead to higher-quality AI outputs.

A security attack in which an attacker inserts hidden instructions into the input in order to bypass the model's intended behavior or its restrictions.

A hidden instruction defining the rules and behavior of an AI system.

A lawsuit filed by The New York Times in December 2023 in a federal court in Manhattan, in which the newspaper claims that OpenAI and Microsoft unlawfully used millions of its articles to train the ChatGPT and Copilot models and that the systems reproduce protected content verbatim. OpenAI, by contrast, argues that the plaintiff deliberately steered the models with deceptive prompts in order to elicit the reproduction of text.

Models for creating video from a text prompt; Sora comes from OpenAI, Veo from Google. They are associated with licensing agreements as well as concerns about deepfakes.