Updated: July 24, 2024

Version 1.126

AI Design Glossary

Intro

To provide you with some quick help on AI terminology used in or referred to within our guidelines for ‘Designing Intelligent Systems’, you can find a set of basic terminology below.

Terms are sorted in alphabetical order and grouped into five categories:

Artificial Intelligence
AI Ethics and AI User-Centered Design
Generative AI
Machine Learning (ML)
Natural Language Processing (NLP)

Information

In writing this glossary, we extensively researched SAP and global definitions, including those by UNESCO, OECD.AI, and other institutes.

In the process, we leveraged ChatGPT to simplify these explanations so that anyone, regardless of their previous knowledge, can build their understanding of AI.

We thoroughly reviewed, validated and edited those outputs with other human AI experts at SAP, and will continue updating and refining them.

Your SAP AI Design team

List of Terms

Artificial Intelligence

Artificial intelligence (AI)

AI is typically defined as a machine’s ability to perceive, reason, learn, talk, and solve problems. It’s a big field, so keep on reading!

Artificial general intelligence (AGI)

Not to be confused with generative AI. The concept behind AGI is that AI can learn to accomplish tasks that would usually require general human cognitive abilities.

Artificial superintelligence (ASI)

The stuff of science fiction: self-aware AI, which has surpassed the abilities of the human brain.

Human-in-the-loop (HITL)

An actual human being is involved in training, testing, and optimizing an AI system. Think: A child learning about apples might mistake pears for apples. An adult would correct this, thus teaching the child the right labels.

Learning-based AI

Systems in which humans define the problem and goal, and systems “learn” how to accomplish this goal and get better at it.

Narrow AI (ANI)

AI solutions designed to perform specific tasks within a limited domain. Also called weak AI.

Rule-based AI (business rule AI)

Systems that use rules made by human experts, also known as “symbolic” or “expert” systems.

Strong AI

Artificial general intelligence (AGI) and artificial superintelligence (ASI).

AI Ethics and AI User-Centered Design

Accountability

Taking responsibility for the actions and outcomes of AI systems.

Bias

Systematic errors or unfairness in AI outputs that can result from biased training data or algorithmic decision-making.

Bias mitigation

The process of identifying and addressing biases in AI systems to promote fairness, avoid discrimination, and ensure equitable outcomes.

Data ethics

Ethical considerations and responsible practices related to data collection, storage, and use in generative AI systems.

Ethical AI

AI ethics is a set of values, principles, and techniques that employ widely accepted standards to guide moral conduct in the development, deployment, use, and sale of AI systems.

Explainability

Enabling users to understand and interpret the decisions and outputs of AI systems, ensuring transparency and user trust.

Explicit feedback

Clear and direct feedback provided to generative AI models to guide and influence their future output.

Fairness

The process embedded into the algorithm to prevent bias. Because a biased approach to building something with biased data leads to biased output.

Implicit feedback

Indirect feedback that is collected from user interactions, preferences, or behaviors, and used to refine generative AI models.

Interpretability

Ability to interpret and understand the inner workings and decision-making processes of AI models.

Privacy

The right to have your personal and business data protected.

Progressive disclosure

The technique of gradually revealing information or functionality to enhance user understanding.

Responsible AI (RAI)

Ethical and responsible development, deployment, and use of AI systems to make sure they are fair, transparent, and respectful of people’s rights.

Robustness

The ability of generative AI systems to perform consistently and reliably in various scenarios and handle unexpected inputs or conditions.

Transparency

Providing information about how the AI system makes decisions so users can understand and challenge them.

User empowerment

Designing AI experiences that empower users by giving them control, customization options, and opportunities for meaningful engagement.

User feedback loops

Mechanisms for users to provide feedback on generative AI outputs, enabling iterative improvements and personalized experiences.

Generative AI

Augmenting

Enhancing or enriching generative AI outputs by adding supplementary information, context, or data.

Base prompt

A core set of instructions given to the large language model (LLM) that serves as a foundation for generating responses or completing tasks.

Blocklist

A list of specific words, phrases, or types of content that are filtered out to prevent the AI from producing inappropriate outputs.

Custom prompt

An instruction for the generative AI model that the user writes from scratch in natural language.

Embeddings

Embeddings enhance prompts by searching a knowledge base for context, providing a semantic representation of relevant documents, and improving the ability of a large language model (LLM) to find semantically similar information.

Emergent abilities

Unintended or unexpected patterns, behaviors, or capabilities that arise from the interaction of complex AI systems like generative AI models.

Fine-tuning

Fine-tuning large language models (LLMs) is the resource-intensive process of customizing a pre-trained language model on specific tasks or datasets to make it more proficient and accurate in generating relevant text.

Foundation models

Deep learning models trained on large volumes of unlabelled data using self-supervised learning. Applicable to a wide range of tasks.

Generative adversarial networks (GANs)

Machine learning model with two neural networks that compete with each other to make better predictions.

Generative AI

A form of artificial intelligence that, when instructed by the user, can create novel content based on its training data, including text, images, sound, or video.

GPT (generative pre-trained transformer)

A type of generative AI model that utilizes transformer architecture for tasks like language generation.

Grounding

Limiting the scope of generative AI models by connecting the generated content with specific real-world data or references to ensure the generated outputs align with the intended purpose.

Guided prompt

Settings and options that the user selects to create precise instructions for the generative AI model.

Hallucinations

Generated AI outputs that sound plausible but are either false or unrelated to the given context, making them difficult for humans to detect as errors.

Hidden prompt

A hidden prompt is like a secret instruction guiding a language model—like the puppeteer behind the scenes.

Metaprompt

Instructions given to a generative AI model to guide its behavior or output in a desired direction.

Parameters

The internal settings or variables of a generative AI model that control its behavior, output, and learning process.

Probabilistic

Situations with multiple possible outcomes, each having varying degrees of certainty.

Prompt

An instruction that users give to generative AI models to guide their output.

Prompt engineering

The process of designing and refining instructions to guide the behavior and output of generative AI models.

Quick prompt

Presets provided by the system and expertly crafted by prompt engineers, eliminating the need for users to write their own prompts.

Response filtering

Selecting or filtering generated responses from a generative AI model based on specific criteria or quality measures.

Retrieval augmented generation (RAG)

A model retrieves relevant information from a pre-existing dataset or knowledge source to generate more accurate and contextually appropriate outputs.

Style transfer

A technique that applies the visual style of one image to another, combining content with specific aesthetic characteristics.

Transformer model

A neural network architecture that learns context and relationships in sequential data. It enables generation of new content, such as text or images, based on patterns and examples provided during training.

Machine Learning

Deep learning

Machine learning with three or more layers of the neural network. Most suited to processing massive quantities of data.

Machine learning (ML)

Subset of AI that enables systems to learn from data and improve without explicit programming.

Machine learning model

A computer program made of algorithms and mathematical equations. It can learn independently by recognizing patterns in data.

Neural networks

Computing systems inspired by the structure of the human brain. They are made of layers of algebraic equations called artificial neurons (or nodes); the first layer receives the data, and the last outputs the results.

Reinforcement learning

The system learns by being placed into an environment where it figures out what is possible and what’s not through experience and reward, without human involvement.

Supervised learning

The machine finds similarities and differences in the untrained dataset by learning from pairs of data: input and output labeled by humans.

Unsupervised learning

Data isn’t labeled in unsupervised learning. The system studies the dataset, looks for patterns, and suggests how to group things.

Natural Language Processing (NLP)

ChatGPT

An AI language model developed by OpenAI that focuses on generating conversational responses.

Large language models (LLMs)

A subcategory of foundational models that can learn to predict the next word in a text through analysing vast amounts of text available on the internet.

Natural language processing (NLP)

The field of AI concerned with understanding and processing human language, including tasks like speech recognition, text analysis, sentiment analysis, and natural language generation.

Sentiment analysis

Determining the sentiment in a text and classifying it as positive, negative, or neutral.