Glossary

Technical terms used on this site, explained in plain language

Technical terms that appear on these pages. Underlined words in the text lead here.

Drift: When an AI system moves away from its original task over the course of a longer job, without anyone noticing. The term comes from measurement engineering, where it describes a reading slowly wandering off.
Plugin: A small program that extends existing software without changing it. It is hooked in and can be removed again at any time.
Task fidelity: The ability of an AI system to stay with what it was asked to do across many steps of work. The more independently a system works, the more this matters.
Screen reader: Software that reads the contents of the screen aloud. Blind and partially sighted people use it to operate their computer. A screen reader does not read the image but the description behind it: it needs text where a sighted person makes do with an icon.
Assistive technology: Anything that creates access where it would otherwise be missing: screen readers, magnifiers, braille displays, voice input, specialised keyboards, eye control. The term refers to the tool, not to the person.
WCAG: The Web Content Accessibility Guidelines, the international standard for accessible websites. They define three levels: A, AA and AAA. AA is the level required by law in Europe — and the one we aim for. Version 2.2 has been in force since 2023.
Open access: A scientific publication that anyone can read free of charge — no paywall, no library subscription. The opposite is publication in a journal that sells access.
Free licence: Permission to use, change and pass on a work. For software this is called open source. Free does not mean without conditions: usually the origin must be credited and the freedom passed on.
Codebook: In empirical research, the set of rules by which material is categorised: which categories exist, what belongs in which, how to handle borderline cases. Publishing the codebook makes an analysis verifiable — without it, a number remains a claim.
Value card: The principles a language model is meant to answer by name individual values — helpfulness, honesty, harm avoidance and others. We call them value cards because they can be played against each other like cards: which one wins when two of them conflict?
Second coding: The same material is categorised independently by a second person. If both agree, the categorisation was more than an opinion. The level of agreement is stated as a number.
Prototype: A working piece of software built to answer a question, not a finished product. It is allowed to be incomplete. Its purpose is to find something out.
Large language model: An AI system trained on very large amounts of text that produces language. Short: LLM. Well-known examples are Claude, ChatGPT and Gemini.
Agent: An AI system that does not merely answer but carries out steps of work on its own: searching, writing, changing files, making decisions along the way. The more independently it works, the more the question matters whether it is still on task.
Prompt: The input you address an AI system with. It sets the first frame for the answer: what is asked, in which role, with what material. Two prompts about the same thing can lead to very different answers.
Context window: The amount of text a model can take into account at once when answering. Whatever does not fit has no effect — even if it was said earlier. This limit is no side issue: it decides what exists for the system at all.
Retrieval: Looking up additional sources while answering, instead of drawing only on what was learned. Which sources are found and which are not shifts the result.
Training data: The material from which a model learned its patterns. What appears often in it appears often in the answers. What is missing there is usually missing in the answers too.
Model architecture: The internal design of an AI system: how its computational steps are arranged and therefore what it can process at all. It comes before any content and sets the limits for it.
Frame: A frame of interpretation. It determines what counts as important, correct or questionable in a situation — before any details are discussed. Whoever sets the frame has already decided a great deal.
Patiency: The ethical status of being more than a means. From the Latin patiens, “undergoing”: it refers to what is affected, not to who acts. “Relational” and “graded” mean here that the claim arises from taking part in a field, not from proving an inner life — and it is not personhood.
Other minds problem: The problem that the inner life of others is never fully accessible. It applies not only to machines but between people as well: we do not prove consciousness to one another, we encounter one another.
DOI: A permanent identifier for a publication. Short for digital object identifier. It leads to the document even when the web address changes — which is why scientific citations use it instead of an ordinary URL.
Non-profit: An association whose purpose serves the public and which does not seek profit. Any surplus may only be used for the association’s purpose, never distributed to members.
General assembly: The meeting of all members and an association’s highest body. It elects the board and decides the fundamental questions. In our association, full members hold the voting rights.
ZVR number: An association’s registration number in the Austrian central register of associations. It allows anyone to verify that an association actually exists and who represents it.