Annex III: Glossary of terms
- Application programming interface (API) – An API is a set of rules or protocols that enables software applications to communicate with each other to exchange data, functions and features.
- Artificial general intelligence (AGI) – AGI is a hypothetical stage in the development of artificial intelligence in which a system can match or exceed the cognitive abilities of human beings across any task. To an extent, it means the replication of human intelligence in an artificial form.
- Chain of thought (CoT) – CoT is a method of engineering prompts of LLMs, particularly for complex tasks in which the AI will need to complete multiple sub-tasks. This technique aims to facilitate problem-solving by guiding the model through a step-by-step reasoning process by using a sequence of coherent steps.
- Contexts – Context computing is the term for the ability of a system or application to understand and engage with the situational context of its user. Contexts could include a user’s location, local time, environmental conditions or preferences.
- Control mechanisms – Control mechanisms are any means of ensuring that an AI system performs in line with expected functionality. It sometimes refers to controls on AI systems that are not based on model training alone.
- Data poisoning – Data poisoning refers to manipulation of the data that a model is trained on or provided with in order to introduce vulnerabilities, compromises or biases. This can lead to reduced security or performance, or behavioural issues, resulting in harms.
- Edge cases – An edge case is a problem or situation that falls outside normal operation in software development. Edge case testing can expose unusual behaviour or flaws.
- Generative AI – Generative AI can synthesise new content (music, images, text, audio and video) from training data. Often trained on extensive datasets, these tools can exhibit a broad range of general-purpose capabilities.
- Guardrails – Guardrails constrain AI systems to ensure or prevent certain outputs. They might include tools that monitor the processing of personal information, or identification of banned content, for example.
- Interfaces – An interface is a point of interaction between two systems or components that allows them to communicate.
- Multimodal – Multimodal interfaces are systems that process combined user input modes (eg voice, gaze, gestures and movements) in a coordinated manner.
- Privacy-enhancing technologies (PETs) – PETs are technologies that embody fundamental data protection principles by minimising personal information use, maximising information security or empowering people.
- Retrieval-augmented generation (RAG) – RAG is the process of optimising the output of an LLM so that it references an authoritative source outside its training data, such as a database, before it generates a response.