- Experiment
Last updated: 12.05.26
AI Ethics & Policy Research
Understanding AI use and implementation in a library context
Artificial Intelligence (AI) poses a fast-moving and complex challenge for institutions that hold cultural memory. Drawing on research, feedback and global case studies, State Library Victoria is developing ethics guidelines and policy that serves staff and library users.
/filters:quality(75)/filters:no_upscale()/filters:strip_exif()/slv-lab/media/images/1772663871ade/a-crowd-at-create-quarter-in-slv-listening-to-an-ai-talk.jpg)
| Contents |
|---|
Articles
/filters:quality(75)/filters:no_upscale()/filters:strip_exif()/slv-lab/media/images/1776992720a8a/img_9624.jpg)
/filters:quality(75)/filters:no_upscale()/filters:strip_exif()/slv-lab/media/images/175808011e42a/dont-be-evil-lr-20-1.jpg)
/filters:quality(75)/filters:no_upscale()/filters:strip_exif()/slv-lab/media/images/177198817127d/edited_mg_6452-2-copy.jpg)
/filters:quality(75)/filters:no_upscale()/filters:strip_exif()/slv-lab/media/images/177664135053d/kate-crawford-seated-portrait-slv-approved.jpg)
/filters:quality(75)/filters:no_upscale()/filters:strip_exif()/slv-lab/media/images/177501464bf66/edited-frances_densmore_recording_mountain_chief2-copy.jpg)
/filters:quality(75)/filters:no_upscale()/filters:strip_exif()/slv-lab/media/images/175988405faf0/baby-stanley-godard-spencer-shier-collection.jpg)
Resources
Title | Type | Author(s) | Tags |
|---|---|---|---|
| documentation |
| ||
| documentation |
| ||
| tutorial |
|
Glossary
Artificial intelligence
Also known as AI, Artificial Intelligence refers to a digital computer's ability to perform processes that typically require human intellect, like complex reasoning, speech recognition or visual interpretation, and learning from experience. AI can also refer to the scientific field dedicated to developing and understanding these abilities in digital agents.
View Details ->Analytic AI
AI used to analyse information and support decisions. It finds patterns in data, summarises trends and can make predictions or recommendations. Unlike generative AI, analytic AI typically focuses on understanding existing data rather than creating new text, images, or audio.
View Details ->Artificial Neural Network
A computer model inspired by how brains process signals. It contains layers of connected “neurons” that transform input (like words or pixels) into output (like labels or predictions). By adjusting connections during training, the network learns to recognise patterns such as faces, handwriting, or topics. Source: Alban Leveau-Vallier (2025) ‘Glossary,’ trans. Michelle Noteboom, 'The World Through AI', JBE Books
View Details ->CLIP (Contrastive Language-Image Pretraining)
A model that learns links between images and words by training on pairs of pictures and captions. It maps both into a shared “latent space”, helping systems search for, sort, or describe images using text. For example, it can match “a green tram” to relevant photos. Source: Alban Leveau-Vallier (2025) ‘Glossary,’ trans. Michelle Noteboom, 'The World Through AI', JBE Books
View Details ->Computer vision
A field of AI that helps computers “see” by analysing images and video. Computer vision can detect objects, read signs, identify damage in photos or help organise image collections. It is widely used in areas like accessibility tools, quality control and cultural heritage digitisation. Source: Digital NSW [https://www.digital.nsw.gov.au/policy/artificial-intelligence/a-common-understanding-simplified-ai-definitions-from-leading]
View Details ->Convolutional Neural Network (CNN)
A neural network designed for images. It uses small scanning filters to pick up simple features (edges, textures), then combines them across layers to recognise more complex shapes and objects. CNNs are commonly used for image classification, object detection and some handwriting and document analysis tasks. Source: Alban Leveau-Vallier (2025) ‘Glossary,’ trans. Michelle Noteboom, 'The World Through AI', JBE Books
View Details ->DALL-E
A text-to-image model that generates pictures from written descriptions (prompts). You can ask for “a watercolour of Melbourne’s skyline” and it creates an image that fits. Like other generative tools, it reflects patterns in its training data and may produce unexpected or biased results.
View Details ->Data Mining
The process of examining large amounts of data to discover useful patterns, relationships or trends. Data mining can help organisations understand behaviour, detect unusual activity or group similar items. In libraries and archives, it can support collection insights – provided privacy, consent, and ethics are addressed. Source: https://www.digital.nsw.gov.au/policy/artificial-intelligence/a-common-understanding-simplified-ai-definitions-from-leading
View Details ->Deep Learning
A type of machine learning that uses multi-layer neural networks to learn complex patterns. Because it can process large datasets, deep learning is behind many recent breakthroughs in speech recognition, image analysis and language tools. It often requires substantial computing power and careful evaluation to avoid errors.
View Details ->Fine-tuning
Adapting a pre-trained model to perform better on a specific task or domain. For example, a general language model can be fine-tuned on heritage descriptions to write in a consistent style. Fine-tuning can improve relevance, but it can also amplify problems in the new training data if not managed carefully.
View Details ->Foundational Model
A large, general-purpose model pre-trained on a broad spectrum of mainly unlabelled data, using self-supervised learning techniques. Instead of building a new model from scratch, organisations start with a foundation model and adapt it (through prompting, fine-tuning, or add-ons). Many generative AI systems are built on foundation models.
View Details ->Generative AI Model
An AI model that creates new content – such as text, images, audio or code – based on patterns learned from training data. It does not “understand” in a human way; it generates outputs that statistically fit the prompt. This can create biases, and cause accuracy and rights issues.
View Details ->Generative Pre-trained Transformers (GPT)
A family of language models built using the transformer architecture. They are pre-trained on large text collections to predict likely word sequences, then used for tasks like summarising, drafting or answering questions. They make up the base of ChatGPT, released in November 2023.
View Details ->Handwritten Text Recognition (HTR)
Technology that converts handwriting in images into editable, searchable text. HTR is related to OCR (Optical Character Recognition), but handwriting is harder because letter shapes vary widely between writers and time periods. Good results often depend on clear scans, representative training data and human review for difficult words.
View Details ->Human-in-the-loop (HITL)
An approach where people actively guide, check or approve an AI system’s outputs. Humans might correct labels, review generated text, or decide when the system should stop. HITL helps improve quality and safety, especially in high-stakes or sensitive contexts such as public information, cultural collections or identity-related data. Source: https://www.digital.nsw.gov.au/policy/artificial-intelligence/a-common-understanding-simplified-ai-definitions-from-leading
View Details ->LAION 5B
A very large dataset of image links paired with text descriptions, created by the German non-profit LAION. It is used to train some image and multimodal AI models. “5B” refers to billions of image–text pairs.
View Details ->Language Model
A model that works with language by predicting word sequences based on context. It can generate text, translate, summarise, or classify sentiment by learning statistical patterns from training text. Language models can be useful for drafting and analysis, but they can also reproduce biases and produce plausible-sounding errors.
View Details ->Large Language Model (LLM)
Machine learning models capable of recognising and generating human language. These models are trained on vast datasets of high-quality human textual communication, often scraped from the web. LLMs rely on a form of machine learning called deep learning, which analyses the probability of various letters, words and sentences appearing together across these enormous datasets to eventually understand their function and relationships. To date, ChatGPT is perhaps the most famous example of a LLM.
View Details ->Latent Space
A mathematical “map” where an AI system represents concepts, images or sounds as numbers (vectors). Items that are similar end up close together in this space. Latent space helps models match text to images, group similar document, or generate variations – because moving within the space changes features gradually.
View Details ->Low-rank adaptation (LoRA)
A method for adapting a large model to a new task by adding a small set of learnable changes, instead of retraining the entire model. LoRA is popular because it can be faster and cheaper, and it lets people create multiple specialised versions of a model while keeping the original base model unchanged.
View Details ->Machine Learning
A way for computers to learn patterns from examples rather than being explicitly programmed for every rule. By training on data, a machine learning system can classify items, detect anomalies, or predict outcomes. The quality of results depends heavily on the quality, representativeness, and documentation of the training data.
View Details ->Metadata
Information that describes other data. For a photo, metadata might include creator, date, location, file type and rights details. Good metadata helps people find, interpret and manage collections. In AI, metadata supports transparency by recording where data came from, how it can be used and what limitations apply.
View Details ->Midjourney
An AI research lab known for its eponymous generative AI program that outputs images from natural language descriptions (or prompts). Released in 2022, Midjourney quickly set a standard in the field, yet has also sparked controversy (as other text-to-image models like Dall-E or Stable Diffusion have done): for using artists’ work without their consent, for making it easy to plagiarise, and even for its role in winning digital art and photography prizes.
View Details ->Multi-Agent Systems
Systems where multiple AI “agents” work together, each with a role such as planning, researching, checking or executing steps. Agents can collaborate or compete to solve complex tasks. Multi-agent approaches can improve coverage and reliability, but they can also multiply errors if agents share the same wrong assumptions or unreliable sources.
View Details ->Multimodal Models
AI models that work with more than one type of input or output – such as text, images, audio or video. A multimodal model might describe an image in words, answer questions about a diagram or generate an image from text.
View Details ->Natural Language Processing (NLP)
A field of computing focused on helping machines work with human language. NLP powers tools like search, speech-to-text, translation, summarisation and chatbots. It can support discovery and accessibility, but results vary by language, dialect and context.
View Details ->Optical Character Recognition (OCR)
Technology that converts text in images (like scanned pages) into machine-readable text. OCR makes printed documents searchable and easier to reuse, making it a foundational technology in the digitisation of archive and historic materials. Accuracy depends on scan quality, fonts, layout complexity and language.
View Details ->Prompt
The input you give a generative AI system to guide its output. A prompt might be a question, an instruction, a passage to rewrite or a description of an image.
View Details ->Reinforcement Learning
A machine learning approach where an agent learns by trial and error. It takes actions, receives rewards or penalties, and gradually learns strategies that maximise reward. Reinforcement learning is used in robotics, games and optimisation problems. Source: https://www.digital.nsw.gov.au/policy/artificial-intelligence/a-common-understanding-simplified-ai-definitions-from-leading
View Details ->Retrieval Augmented Generation (RAG)
A technique that helps a generative AI system use trusted information. Before writing an answer, the system retrieves relevant documents (for example, from a library knowledge base) and then generates text grounded in those sources. RAG can reduce hallucinations and improve currency, but it still requires good retrieval, permissions and careful citation practices.
View Details ->Scraping
Automatically collecting data from websites or online sources, often by software that reads pages at scale. Scraping can support research and discovery, but it can breach terms of use, overload systems or collect personal data without consent. Ethical scraping includes respecting permissions, robots.txt guidance, privacy law and cultural sensitivities.
View Details ->Speculative Design
A design practice introduced in the 1990s by Anthony Dunne and Fiona Raby that uses imagined futures to question how technologies, systems, and values shape societies. Speculative design creates scenarios and artefacts that provoke debate, encourage reflection and open up discussion about possible, preferable or troubling futures. (Source: https://www.critical.design/post/what-is-speculative-design)
View Details ->Supervised Learning
A machine learning method that learns from labelled examples, where the “right answer” is provided. For instance, emails marked “spam” or “not spam” can train a filter. Source: https://www.digital.nsw.gov.au/policy/artificial-intelligence/a-common-understanding-simplified-ai-definitions-from-leading
View Details ->Training Set
A collection of data, sometimes manually labeled, that an algorithm learns to recognise or complete by trial and error. With each error, the algorithm adjusts its parameters so as to gradually encode enough information about the data (their features, similarities and differences) to be able to classify or produce new ones.
View Details ->Unsupervised Learning
A machine learning approach that finds structure in data without labelled answers. The system groups similar items, detects clusters or reduces complexity to reveal patterns. Unsupervised learning is useful for exploration – such as discovering themes in large text collections – but the groupings still need human interpretation to ensure they make sense and avoid misleading conclusions. Source: https://www.digital.nsw.gov.au/policy/artificial-intelligence/a-common-understanding-simplified-ai-definitions-from-leading
View Details ->