Understanding the Language of AI: A Glossary of 30 Essential Terms

AI terminology can seem technical at first, but understanding it is becoming increasingly important as these tools find their way into humanitarian work. Knowing what a large language model is, what a hallucination is, or what human-in-the-loop means can help practitioners ask better questions, recognize potential risks, and make more informed decisions about the tools they use.

For humanitarian organizations, this understanding goes beyond keeping up with technological developments. It is about ensuring that AI is used responsibly, that people remain accountable for decisions, and that the rights, safety, and dignity of affected communities remain central.

AI will continue to evolve, and so will the language used to describe it. But the starting point is simple: understand the technology, question its outputs, and keep people at the center of its use.

In this glossary, we break down some of the key AI terms and concepts that humanitarian practitioners should know.

1. Artificial Intelligence (AI): Technology that enables computers to perform tasks that typically require human intelligence, such as recognizing patterns, understanding language, analyzing information, or generating content.

• Humanitarian example: An AI system helping an organization analyze large volumes of situation reports to identify emerging humanitarian needs.

2. Machine Learning (ML): A branch of AI that allows computer systems to identify patterns in data and improve their performance on particular tasks without being explicitly programmed for every possible situation.

• Humanitarian example: A system trained on historical disaster data to identify patterns that may help inform future risk assessments.

3. Generative AI: A type of AI that creates new content, including text, images, audio, video, or computer code, based on patterns learned from existing data.

• Humanitarian example: A practitioner using an AI tool to draft a situation update, translate a message, or summarize a lengthy report.

4. Large Language Model (LLM): An AI model trained on large amounts of text to understand and generate human language. LLMs power many conversational AI tools and can perform tasks such as answering questions, summarizing documents, and translating text.

5. Prompt: The instruction, question, or information a user provides to an AI system to guide its response.

• Humanitarian example: Asking an AI tool to summarize the key protection concerns identified in a collection of humanitarian reports.

6. Training Data: The information used to teach an AI model to recognize patterns and perform tasks. The quality, diversity, and representation of this data can influence how well the model works.

• Humanitarian example: If the data used to train a model contains limited information about certain regions or languages, its performance may be weaker when working with those contexts.

7. Algorithm: A set of instructions or rules that a computer follows to perform a task or solve a problem. Algorithms are used in many AI systems to process information and generate outputs.

8. Hallucination: When an AI system generates information that sounds convincing but is inaccurate, fabricated, or unsupported by the available evidence.

• Humanitarian example: An AI assistant incorrectly identifying an organization as operating in a particular area or providing a statistic that cannot be verified.

9. Algorithmic Bias: When an AI system produces systematically unfair or skewed results, often because of limitations or biases in its training data, design, or use.

• Humanitarian example: An AI system that performs less accurately for certain languages or communities because they are underrepresented in its training data.

10. Retrieval-Augmented Generation (RAG): An approach that allows an AI model to retrieve information from an external collection of documents or sources before generating an answer. This can help ground responses in specific information rather than relying only on what the model learned during training.

• Humanitarian example: A virtual assistant retrieving information from verified humanitarian reports before answering a question about a crisis.

11. Human-in-the-Loop (HITL): An approach in which people remain actively involved in reviewing, guiding, or validating AI outputs before they are used or acted upon.

• Humanitarian example: A humanitarian analyst reviewing AI-generated findings against original reports before including them in a situation analysis.

12. Explainability: The extent to which people can understand how an AI system produces its outputs or why it makes particular recommendations.

13. Transparency: Being open about how an AI system is developed, what information it uses, what it can and cannot do, and how its outputs should be interpreted.

14. AI Governance: The policies, processes, responsibilities, and safeguards that guide how an organization develops, adopts, and uses AI.

• Humanitarian example: An organization establishing rules about which AI tools staff can use, what information they can enter, and who is responsible for reviewing AI-generated outputs.

15. AI Localization: Adapting AI systems to the languages, cultures, needs, and operating conditions of particular communities or regions. This can include improving language support, contextual relevance, and representation.

• Humanitarian example: Ensuring an AI tool can understand local languages and expressions rather than relying only on widely used international languages.

16. Shadow AI: The use of AI tools by employees without formal organizational approval, awareness, or oversight. It can happen when staff turn to publicly available tools to meet their work needs without clear guidance or approved alternatives.

• Humanitarian example: A humanitarian worker entering sensitive information from a field report into a public AI chatbot without knowing whether the information is stored or how it may be used.

17. AI model: A computer system trained on data to recognize patterns and perform specific tasks, such as generating text, identifying images, translating languages or making predictions. Different models are designed for different purposes, and their capabilities depend on their training data, design and development.

• Open-Source Model: An open-source model is an AI model made available for others to access, use, modify or build upon under its license. The level of openness varies: some models provide access to their weights and code, while others may restrict access to training data or other components.
• Humanitarian example: An organization might use an open-source AI model to develop a tool that translates local-language reports or helps analyze humanitarian information.

18. Fine-tuning: Fine-tuning is the process of taking an already trained AI model and providing additional training on a more specific dataset to adapt its performance for a particular task, subject or context. Rather than training a model from scratch, fine-tuning builds on what the model has already learned.

• Humanitarian example: A general language model could be fine-tuned using humanitarian terminology and relevant documents to improve its ability to handle specialized humanitarian information. The quality and representativeness of the additional data remain important.

19. Reinforcement Learning from Human Feedback (RLHF): A method used to improve an AI model's responses using feedback from people. Human reviewers assess or rank different model outputs, and that feedback is used to guide the model toward responses that better match the desired behavior.

• Humanitarian example: Human reviewers could assess AI-generated responses to humanitarian questions, indicating which answers are clearer, more useful, and appropriately cautious. Their feedback can help improve the model's responses, although it does not guarantee accuracy.

20. Chatbot: A software application designed to communicate with users through text or voice. It responds to questions, follows conversational instructions, and can help users access information or complete simple tasks. Some chatbots use predefined rules, while others are powered by AI models that can generate responses.

• Humanitarian example: A humanitarian organization might use a chatbot to answer frequently asked questions about available services, eligibility requirements, or assistance programs.

21. Prompt Engineering: The practice of designing, structuring, and refining instructions given to an AI system to produce more useful and relevant responses. It involves providing clear questions, context, constraints and examples to guide the system.

• Humanitarian example: Instead of asking an AI tool to "summarize this report", a practitioner could specify the report's location, the intended audience, the key themes to extract and the preferred summary format.

22. Token: A unit of text that an AI model processes. A token can represent a whole word, part of a word, punctuation mark, or other piece of text. AI models break text into tokens before processing it, and the number of tokens can affect how much information a model can handle at once.

• Humanitarian example: When analyzing a lengthy humanitarian assessment report, the amount of text the AI system can process may be limited by its token capacity.

23. Context Window: The amount of information an AI model can consider at one time when generating a response. It is measured in tokens and may include the user's prompt, previous messages, and documents or other information supplied to the model.

• Humanitarian example: When reviewing a lengthy situation report, an AI tool with a limited context window may not be able to consider the entire document at once. Important information may need to be provided in sections or processed using a suitable document-handling approach.

24. Natural Language Processing (NLP): A branch of AI that enables computers to process, interpret, and generate human language. It supports tasks such as translation, speech recognition, text classification, information extraction, and sentiment analysis.

• Humanitarian example: NLP can help humanitarian organizations analyze large volumes of community feedback, identify recurring concerns and process reports written in different languages.

25. Deepfake: Synthetic media, such as an image, video, or audio recording, generated or manipulated using AI to make someone appear to say or do something they did not actually say or do. Deepfakes can be used for entertainment and creative purposes, but they can also facilitate misinformation, impersonation, and fraud.

• Humanitarian example: A fabricated video showing a humanitarian organization making a false announcement could undermine public trust, create confusion and affect the credibility of an emergency response.

26. Alignment: AI alignment refers to the process of ensuring that an AI system's behavior, objectives and outputs are consistent with the intentions, values and expectations of the people responsible for developing or using it. It involves reducing the likelihood that a system will produce harmful, misleading, or unintended outcomes.

• Humanitarian example: An AI system used to support humanitarian analysis should be designed and used in ways that respect humanitarian principles, avoid reinforcing discrimination, and support rather than replace human judgment.

27. Guardrails: Guardrails are safeguards, rules, and technical measures designed to guide how an AI system operates and reduce the risk of harmful, inappropriate, or unintended outputs. They may include restrictions on certain requests, privacy protections, content filters, and requirements for human review.

• Humanitarian example: A humanitarian AI tool could have guardrails that prevent it from exposing personally identifiable information, generating responses to certain high-risk requests, or presenting unverified information as established fact.

28. AI Agent: An AI agent is an AI system designed to work towards a goal by carrying out a sequence of tasks, making decisions based on available information and, in some cases, using external tools or systems. Unlike a basic conversational tool that primarily responds to individual prompts, an AI agent may be able to plan and take multiple steps to complete a task, subject to its permissions and level of human oversight.

• Humanitarian example: An AI agent could help a humanitarian team monitor selected information sources, identify relevant updates, and organize findings for an analyst to review. Human oversight remains important, particularly when the information may influence operational decisions.

29. Multimodal AI: AI systems that can process and work with more than one type of information, such as text, images, audio, and video. These systems can combine different forms of data to understand information and generate responses.

• Humanitarian example: A multimodal AI tool could help humanitarian practitioners analyze satellite imagery alongside written situation reports to better understand the impact of a disaster.

30. Synthetic Data: Artificially generated information designed to resemble real-world data. It can be used to train, test, or evaluate AI systems, particularly when real data is limited, difficult to access, or sensitive. However, synthetic data may still reproduce biases or inaccuracies present in the data or methods used to generate it.

• Humanitarian example: Synthetic data could help organizations test an AI system for analyzing population trends without directly using identifiable personal information from affected communities.

Still curious about humanitarian AI or AI in general?

Encountered an AI term in your humanitarian work that you don't quite understand, or is there a concept you'd like us to explain?

We'd love to hear from you! Reach out to us at hello@datafriendlyspace.org and let us know which AI terms or concepts you'd like us to explore in future explainers.