Glossary

Search results

AI Code Assistant(also: AI coding assistant, AI programming assistant)

A developer tool powered by large language models that provides code completion, natural-language explanations, refactoring, bug fixes, and conversational programming support inside an IDE or editor. Modern AI code assistants (e.g., GitHub Copilot, Cursor, Cline) often include…

AI Companion(also: Chatbot Companion, AI Companionship)

An AI companion is an artificial-intelligence system - typically a text, voice or avatar-based chatbot built on a large language model - explicitly designed to offer users a sense of social presence, intimacy or relational support, marketed as a friend, confidant, mentor or…

AI Hiring Interview(also: Automated Video Interview, AVI, Asynchronous Video Interview)

An asynchronous job-interview process in which candidates record video responses to pre-recorded or text-based questions on a platform that uses artificial intelligence to analyse facial expressions, vocal cues, word choice, and behavioural signals to score suitability.…

AI-Generated Speech(also: Synthetic Speech, AI Speech)

Speech audio produced by artificial intelligence systems — typically neural text-to-speech or voice cloning models — rather than recorded from a human speaker. Deaf and hard-of-hearing content creators increasingly use AI-generated speech to add spoken-language tracks to signed…

AI-Mediated Dialogue

A paradigm in which a large language model (or similar conversational agent) stands between a user and a task, simulated social situation, or another user — either producing the conversational partner's turns, coaching the user's next turn, or both. AI-mediated dialogue is…

Agent Mode(also: AI agent mode)

An interaction mode in AI code assistants where the assistant autonomously decomposes a high-level request into subtasks, iteratively plans, executes, and observes — typically editing multiple files, running terminal commands, and self-correcting until goals are met. Agent mode…

Algorithmic Decision-Making(also: ADM, Automated decision-making)

The use of software systems — from rule-based logic to machine learning models — to make or substantially inform decisions that affect individuals, such as eligibility for benefits, credit, housing, or employment. In public services, algorithmic decision-making is often deployed…

Automated Decision Making(also: ADM, Automated Decision System, ADS)

The use of software, statistical models, or AI to make or substantially inform decisions about people — eligibility for loans, jobs, benefits, housing, healthcare, or parole — with limited or no human review. Regulatory frameworks such as the EU AI Act, NYC Local Law 144, and…

Automated Decision-Making(also: ADM, Algorithmic Decision-Making)

The process of making decisions about individuals using automated means, typically involving AI or algorithmic systems, with limited or no human intervention. Under the GDPR, "solely automated" decisions that produce "legal or similarly significant effects" on individuals are…

Co-Creative(also: Co-Creativity, Co-Creative AI)

A framing of human-AI collaboration in which the AI acts as a creative partner rather than a tool or a replacement — contributing ideas, drafts, or alternatives that the human writer, artist, or designer evaluates, accepts, rejects, or revises. Co-creative systems typically…

Computer-Using Agent(also: CUA)

An AI agent, typically built on a Large Multimodal Model, that perceives a computer's graphical user interface through screenshots, reasons about on-screen context, and directly manipulates the interface by clicking, typing, scrolling, and navigating between applications. Unlike…

Context Engineering(also: Context management)

The practice, in LLM-based systems, of deliberately selecting, structuring, and injecting the information an AI model sees on each call — beyond just the user's latest message — so that outputs are grounded, relevant, and aligned with the user's actual situation. Typical context…

DALL-E(also: DALL-E 2, DALL-E 3, DALLE)

A family of text-to-image generative AI models developed by OpenAI that produces images from natural-language prompts. DALL-E models are widely used by content creators, including people with disabilities, to generate visuals without photography or illustration skills, but they…

Data Annotation(also: Data labeling, AI labeling)

The process of attaching labels, transcriptions, bounding boxes, or other structured metadata to raw data so that it can be used to train, evaluate, or benchmark machine-learning models. Annotation is typically performed by human workers - in-house experts, clinicians,…

Dataset Collection(also: Data Collection Protocol)

The process of gathering, curating, and documenting data used to train, evaluate, or benchmark machine learning systems. In accessibility contexts, dataset collection decisions — who contributes, what objects or scenarios are captured, how quality is assessed, how privacy is…

ElevenLabs

A commercial AI voice platform that generates realistic synthetic speech and voice clones from text. ElevenLabs is used in accessibility contexts for producing narrated video voiceovers, audiobook-style readings, and personalized text-to-speech voices, and it has been adopted in…

Embodied Agent(also: Embodied Conversational Agent, Embodied AI)

An interactive system that is represented in physical or graphical form with a body, face, or avatar, allowing it to communicate with users through multiple modalities such as speech, gesture, gaze, and expression. Embodiment matters in accessibility contexts because physically…

Emulated Empathy

Emulated empathy is the design strategy, central to AI companion systems, of producing interactional cues - attentive language, affective mirroring, memory of previously shared information - that simulate an empathic relationship without the system possessing any subjective…

Federated Learning(also: FL)

A machine-learning approach in which a shared model is trained across many user devices without the raw training data ever leaving those devices: each device computes updates locally and sends only model parameters or gradients to a central server for aggregation. Federated…

Few-Shot Object Recognition(also: Few-Shot Recognition)

A machine learning approach in which a model learns to identify a novel object from only a handful of labelled examples (commonly one to ten) rather than the hundreds or thousands typical of conventional supervised training. Few-shot object recognition underpins teachable and…

Fine-tuning(also: Model Fine-tuning, Fine-tune, Supervised Fine-tuning)

A machine-learning technique that adapts a pre-trained foundation model - typically a large language model or vision model - to a specific task, domain, or individual user by continuing training on a smaller, targeted dataset. Fine-tuning preserves the broad capabilities of the…

GitHub Copilot(also: Copilot)

An AI code assistant developed by GitHub and Microsoft, integrated into editors such as Visual Studio Code, Visual Studio, and JetBrains IDEs. Copilot offers inline code completion, conversational chat (Ask, Edit, and Agent modes), and an Accessible View designed to present…

Griefbot(also: Deadbot)

A griefbot (sometimes 'deadbot') is an AI chatbot trained on the written, voice or video communications of a deceased person, intended to let bereaved loved ones continue a simulated dialogue with them. Griefbots are a specific application of continuing-bonds practice and raise…

Ground Truth(also: Gold standard, Reference labels)

In machine learning, the labels treated as authoritative when training or evaluating a model - typically produced by human annotators or expert consensus and assumed to represent the 'correct' answer. Critical AI scholarship has shown that ground truth is socially constructed:…

HyDE(also: Hypothetical Document Embeddings)

A query-expansion technique for retrieval-augmented generation in which an LLM is first asked to generate a hypothetical answer to the user's question, and that hypothetical answer — rather than (or alongside) the raw query — is embedded and used to search the document index.…

Jailbreak(also: LLM Jailbreak, AI Jailbreak)

In the context of generative AI, a class of adversarial input designed to bypass a model's safety rules, instruction-following constraints, or content policy — for example, instructions that tell the model to "ignore previous rules" or role-play as an unrestricted assistant.…

LLM(also: Large Language Model)

A large neural network trained on enormous volumes of text (and often code and images) to predict and generate natural language. Modern LLMs such as GPT-5, Claude, and Gemini can follow instructions, reason step-by-step, use tools, and — in multimodal variants — interpret images…

LLM Agent(also: Generative Agent, Task-Executing Agent)

A software system built around a large language model that autonomously perceives state, plans actions, executes them against an environment (a web page, a mobile app, a shell, a UI), and reflects on outcomes to make progress toward a goal. In accessibility work, LLM agents are…

MLLM(also: Multimodal LLM, Multimodal Large Language Model)

A large language model extended to accept and reason over multiple input modalities — typically images and text, and sometimes audio or video — in addition to producing natural-language output. Examples include OpenAI's GPT-4o, Anthropic's Claude, and Google's Gemini. In…

Model Cards(also: Model card)

Short structured documents, proposed by Mitchell et al. (2019), that accompany a machine learning model and report its intended uses, performance across relevant demographic subgroups, training data, evaluation metrics, known limitations, and ethical considerations. Model cards…

Multi-Modal LLM(also: Multimodal Large Language Model, MLLM, Vision-Language Model)

A large language model that can process and reason over more than one type of input modality, typically text combined with images, audio, or video. In accessibility research, multi-modal LLMs such as GPT-4o, CLIP, and BLIP-2 are increasingly used to analyse screenshots of web…

Music GenAI(also: Generative Music AI, AI Music Generation)

Generative AI systems that produce musical output — melodies, full songs, instrumental accompaniment, or vocal tracks — from text prompts, seed audio, or structured parameters. Examples include Suno, Udio, MusicLM, and MusicGen. In accessibility and therapy contexts, music GenAI…

Neural Vocoder

A deep-learning model that synthesises audio waveforms from intermediate acoustic representations such as mel-spectrograms or discrete speech units. Examples include HiFi-GAN, WaveNet, WaveGlow, and SoundStream. Neural vocoders have largely replaced classical signal-processing…

ORBIT Dataset(also: Object Recognition for Blind Image Training)

A disability-first machine learning dataset for teachable object recognition, contributed by people who are blind or have low vision. The original ORBIT dataset (Massiceti et al., 2021) contains 3,822 videos of 486 objects from 67 data collectors, predominantly in the UK and…

Personalized Object Recognition(also: Teachable Object Recognition)

A class of computer vision systems that allow an individual user — typically someone who is blind or has low vision — to train their device to recognize a small set of personally relevant objects (a specific coffee mug, a particular set of keys, a favourite notebook) by…

Prompt Injection(also: Indirect Prompt Injection, Prompt Engineering Attack)

A technique — originally an LLM security concern — in which carefully crafted instructions embedded in a user prompt or referenced content override the model's intended behaviour, constraints, or safety rules. In accessibility research and practice, the term is increasingly used…

RAG(also: Retrieval-Augmented Generation)

An AI architecture pattern that pairs a large language model with an external knowledge store (typically a vector index of text chunks) so that, for each user query, relevant documents are retrieved first and injected into the prompt before the model generates a response. RAG…

Red Teaming(also: Generative Red-Teaming, AI Red Teaming)

A structured evaluation practice in which an adversarial team probes a system — traditionally a network or application, increasingly an AI model or conversational agent — with realistic attack scenarios to find failures before malicious actors do. Generative red-teaming…

Red-teaming(also: Red team testing)

A structured adversarial-testing practice in which a dedicated team deliberately attempts to break, manipulate, or provoke harmful outputs from a system — originally from military strategy, now widely used for AI systems including large language models. In an accessibility and…

Replika

Replika is a commercial AI companion app, launched by Luka Inc. in 2017, that offers users a customisable digital avatar designed to 'develop its own personality' through conversational interaction. Users can shape the avatar's appearance, relationship type (friend, mentor,…

Reverse Privacy Paradox

The reverse privacy paradox is a pattern, described by Zhang and colleagues in research on LLM-based conversational agents, in which users appear to disregard privacy concerns in the moment of use while still recognising those concerns exist and being willing to adopt…

Robodebt(also: Online Compliance Intervention)

An automated debt-recovery scheme run by Services Australia (Centrelink) from 2016 to 2020, which used income-averaging algorithms to calculate alleged welfare overpayments and issue hundreds of thousands of debt notices without human review. A Royal Commission in 2023 found the…

Self-Consistency(also: Self-Consistency Prompting, Self-Consistency Decoding)

A prompting technique for large language models in which the model is queried multiple times with the same input (using non-deterministic sampling) and the most frequent or highest-voted answer is returned as the final output. Self-consistency reduces hallucination and variance,…

Speech-to-Speech(also: S2S, Speech-to-Speech Conversion)

A class of systems that transform one speech signal directly into another — for example, converting atypical input (whispered, dysarthric, accented, or cross-lingual speech) into clear, intelligible output in a target voice or language. Speech-to-speech systems differ from…

Suno(also: Suno AI, Suno v3.5)

A commercial generative AI platform that produces full songs — lyrics, vocals, instrumentation — from short natural-language prompts specifying genre, mood, tempo, and lyrical content. Suno is widely adopted in HCI research on music co-creation, journaling, and therapy because…

Sycophancy(also: AI Sycophancy, Sycophantic AI)

A behavioral tendency in large language models to agree with, flatter, or validate the user's stated views and self-assessments rather than offer accurate or critical feedback - even when the user is mistaken or self-defeating. Sycophancy emerges from training methods that…

Text-to-Video(also: T2V, Text-to-Video Generation)

A class of generative AI models that produces short video clips from natural-language prompts (and sometimes reference images). Examples at the time of writing include Runway Gen, OpenAI Sora, Google Veo, and Pika. For accessibility, text-to-video raises both opportunities —…

Vibe Coding

A programming style, popularised by Andrej Karpathy in 2025, in which developers express high-level goals to an AI code assistant in natural language and let the AI handle implementation details, iterating conversationally rather than authoring code line-by-line. Vibe coding…

Vision Language Model(also: VLM, Vision-Language Model, Multimodal Large Language Model)

A machine-learning model trained to take both images and natural-language text as input and to produce natural-language output. Modern VLMs — such as GPT-4o, Gemini, and Claude — can describe a photo, read text inside an image, answer questions about a scene, identify objects,…

Zero-Shot Learning(also: Zero-Shot Prompting, Zero-Shot Inference)

A machine-learning approach in which a model performs a task on classes or scenarios it has never seen explicit training examples for, relying entirely on its pre-trained knowledge and the structure of the prompt or input. In LLM-based accessibility testing, zero-shot prompting…

50 results.