GenAI Concepts

GenAI
Concepts
Technical, operational and regulatory terms and
concepts for generative artificial intelligence (GenAI)
Developed as a collaboration between the Authors Webpage

ARC Centre of Excellence for Automated Fan Yang admscentre.org.au/
Decision-Making and Society (ADM+S) Jake Goldenfein genai-concepts
and the Office of the Victorian Information
Kathy Nickels
Commissioner (OVIC)
Acknowledgement of Country
In the spirit of reconciliation, we acknowledge the
Traditional Custodians of country throughout Australia
and their connections to land, sea and community.
We pay our respect to their elders past and present
and extend that respect to all Aboriginal and Torres
Strait Islander peoples today.
Suggested citation (document)

Yang, F., Goldenfein, J., & Nickels, K. (2024). GenAI
Concepts. Melbourne: ARC Centre of Excellence
for Automated Decision-Making and Society RMIT
University, and OVIC. DOI: 10.60836/psmc-rv23
Suggested citation (webpage)

Fan Yang, Jake Goldenfein, and Kathy Nickels,
‘GenAI concepts’, ADM+S and OVIC (Web Page,
2024), https://www.admscentre.org.au/genai-
concepts
Copyright © 2024 Fan Yang, Jake Goldenfein and

Kathy Nickels. The use, distribution or reproduction in
other forums is permitted, provided the original
author(s) and the copyright owner(s) are credited.
ARC Acknowledgement
This research was conducted by the ARC Centre of
Excellence for Automated Decision-Making and
Society (CE200100005). The Centre is funded by the
Australian Government through the Australian
Research Council.
ii
Contents
Introduction .................................................................................... v
Technical ........................................................................................ 1
Generative AI or GenAI ................................................................................................................ 2
Prompt and prompt engineering ................................................................................................... 2
Machine learning .......................................................................................................................... 4
Large language models (LLMs) ................................................................................................... 5
Knowledge cut-off date ................................................................................................................ 5
Chatbot ......................................................................................................................................... 6
General-purpose AI ...................................................................................................................... 6
Foundation model ........................................................................................................................ 7
Frontier model .............................................................................................................................. 7
Transformer architecture .............................................................................................................. 8
Transfer learning .......................................................................................................................... 9
Open-source and closed-source LLMs ........................................................................................ 9
Token ......................................................................................................................................... 12
Reinforcement learning from human feedback (RLHF) ............................................................. 12
Diffusion models ......................................................................................................................... 13
Inference .................................................................................................................................... 14
Operational ................................................................................... 15
Development ................................................................................................................................. 16
Datasets ..................................................................................................................................... 16
Data licensing ............................................................................................................................. 16
Developer ................................................................................................................................... 17
Data labelling and annotation..................................................................................................... 18
Distribution ................................................................................................................................... 19
Supply chains ............................................................................................................................. 19
AI libraries .................................................................................................................................. 20
iii
Machine learning environments ................................................................................................. 21
Deployment and use .................................................................................................................... 22
Fine-tuning ................................................................................................................................. 22
Deployer ..................................................................................................................................... 23
User ............................................................................................................................................ 24
Regulatory .................................................................................... 25
Risks ........................................................................................................................................... 26
Privacy and data protection ........................................................................................................ 26
Hallucination ............................................................................................................................... 28
Safety ......................................................................................................................................... 29
Transparency ............................................................................................................................. 30
Copyright .................................................................................................................................... 30
Data governance and security ................................................................................................... 31
AI standards ............................................................................................................................... 32
Guardrails ................................................................................................................................... 34
Regulatory sandbox ................................................................................................................... 34
Human oversight ........................................................................................................................ 35
AI auditing .................................................................................................................................. 36
Explainable AI (XAI) ................................................................................................................... 36
Post-market surveillance and monitoring ................................................................................... 37
Floating-point operations per second (FLOPS) ......................................................................... 38
Ecological ................................................................................................................................... 39
iv
GenAI Concepts Introduction
Introduction
Generative AI products and services, such as OpenAI’s ChatGPT, Alphabet’s Gemini,
and Microsoft’s Copilot, have sparked substantial interest in the private and public
sectors. Organisations are already experimenting with integrated AI services provided by
big tech firms, as well as custom procurements from smaller software companies. In the
absence of comprehensive AI regulation in Australia, deploying these systems in a
responsible, ethical and legally compliant way demands a deep understanding of how
they function and the legal and ethical challenges they raise.
But generative AI (GenAI) is not simple. AI software systems are diverse and rely on
complex supply chains and data flows. Coming to grips with the technical, operational
and regulatory vocabularies that have emerged around GenAI is a considerable
challenge. To help entities interested in GenAI deployments, this publication outlines
42 concepts fundamental to AI software systems. Each concept is illustrated through
descriptions, examples and real-world use cases, with accessible language and visual
elements to accommodate a diverse range of stakeholders and readerships.
Developed as a collaboration between the ARC Centre of Excellence for Automated

Decision-Making and Society (ADM+S) and the Office of the Victorian Information
Commissioner (OVIC), GenAI Concepts was produced through expert consultation and
analysis of academic literature, industry reports and policy guidance.
GenAI Concepts can also be accessed as a regularly updated website at

admscentre.org.au/genai-concepts
v
GenAI Concepts Technical
Technical
The following diagram demonstrates the interconnections among fundamental technical
terms in the field of artificial intelligence (AI).
Using this diagram as a starting point, this section begins with basic GenAI concepts
relevant to daily applications, including prompt and prompt engineering, large language
models (LLMs), knowledge cut-off dates and chatbots.
It then moves on to more advanced terms that are often encountered in official
documents, such as general-purpose AI, foundation models and frontier AI.
The section further explores specialised technical terms, encompassing open-source and
closed-source LLMs, transformer architecture, transfer learning, tokens, reinforcement
learning from human feedback (RLHF), diffusion models and inference techniques.
1
Generative AI or GenAI
Generative AI or GenAI are both short for generative artificial intelligence. These are software
systems that create content as text, images, music, audio and videos based on a user’s
‘prompts’.
Types of GenAI models in the market
Prompt and prompt engineering

A prompt is an instruction, query or command that a user enters into a GenAI interface to
request a response from the system.
Because GenAI systems produce text through statistical predictions of the most
likely next words in a sentence, the responses that GenAI systems produce may not
always be the same. This is why GenAI outputs are sometimes described as non-
deterministic.
Prompt engineering is the practice of writing inputs for GenAI tools so that the tool can
produce optimal outputs.
2
Example
Basic structure of a prompt:
Acting as a [role] perform [task] in [format] in the [style]
Acting as a [role] Perform a [task] Show as [format] In the [style]
Each role provides The task should be How you would like to This is optional. It
context and clear and specific. structure the shows what tone
background. information generated. you’d like to use for
the information.
Marketer Headline Table Formal

Advertiser Presentation List Poetic
Copywriter Webinar With bullet points Enthusiastic
Accountant Blog post Summary Shakespearian
Lawyer Book outline HTML Accessible
Financial analyst Email sequence Code Basic English
English professor Social media Spreadsheet Scientific
Journalist campaign CSV file Objective
Project manager Product description Plain text file Neutral
Manager Cover letter Rich text Pop culture
Engineer Summary PDF
Recruiter TikTok, YouTube or Markdown
Instagram Reel video
… Word cloud
script
…
Sales page / ad copy
…
As a [life influencer (the role)], create a [blog post (the task)] about the benefits of daily
exercise in a [PDF (the format)] in an [accessible and enthusiastic tone (the style)].
Tip
If the task is too complicated, break it down into steps to make it easier for the GenAI
model to read and interpret.
3
Machine learning
Machine learning (sometimes seen as ML) is a set of techniques for creating algorithms so
that computational systems can learn from data.
A machine learning algorithm is a set of rules or processes that helps an AI system to perform
specific tasks, such as finding patterns in data or making predictions based on inputs. In this
way, the model’s behaviour reflects the data or the learning experience.
Higher quality data helps algorithms improve their accuracy in various tasks, such as
recognising faces in photos, predicting the weather or recommending products to buy.
Types of machine learning
• Supervised machine learning:

An algorithm is provided with labelled data, for example, a collection of
pictures of animals with labels for each animal in each picture. The algorithm
learns from these examples and tries to predict the correct labels for new,
unseen data.
• Unsupervised machine learning:
An algorithm is provided with unlabelled data and then it tries to find patterns
or structures in the data on its own.
• Hybrid machine learning:

Combines elements of supervised and unsupervised learning approaches and
sometimes different types of algorithms to leverage the strengths of different
methods based on the developer’s needs.
Applications
• Healthcare: ProMed (the Program for Monitoring Emerging Diseases) offers

an online real-time data analysis and reporting system showing outbreaks of
infectious diseases worldwide.
• Finance: Fraud detection, automated trading activities and financial advisory

services for investors.
• Marketing: Product recommendations and news feeds on social media

services.
4
• Transportation: Self-driving vehicles, real-time tracking and last-mile delivery

optimisation.
Large language models (LLMs)

Large language models (LLMs) are data transformation systems. They are trained with large
numbers of parameters, which are numerical values that developers adjust to shape the
inputs and outputs of AI models. When a user inputs a prompt, the model generates text
content in response.
LLMs are trained on extremely large datasets sourced from websites, blogs, forums
and news articles etc. They contain millions or billions of parameters. In the case of
OpenAI’s generative pre-trained transformer (GPT) models, the first large language
model GPT-1 contains 117 millions parameters, GPT-2 contains 1.5 billion
parameters and GPT-3 contains 175 billion parameters. Increased parameters often
lead to more complex models that can handle more complicated tasks and generate
more nuanced text.
Users interact with GPT models through interfaces like ChatGPT. This feedback loop
allows the model to continuously learn and improve over time based on user
feedback.
Examples
Some notable LLMs are:
• OpenAI’s GPT series of models (GPT-1, GPT-2, GPT-3, GPT-3.5 and GPT-4)
used in ChatGPT and Microsoft Copilot
• Google’s PaLM (Pathways Language Model) and Gemini
• xAI’s Grok
• Meta’s Llama (Large Language Model Meta AI) family of open source models
• Antropic’s Claude models
Related term: Machine learning
Knowledge cut-off date

The knowledge cut-off date of a GenAI model is the date when the training data for a specific
LLM was last updated. It defines the limitations of the model’s understanding and knowledge.
5
Example
ChatGPT-3.5’s current knowledge cut-off date is January 2022. The knowledge cut-
off date for ChatGPT-4 is April 2023.
Chatbot
Chatbots are popular applications of GenAI and LLMs. A chatbot is a computer program that
interacts with humans through natural language conversations. Some chatbots use LLMs to
generate content according to user inputs.
Chatbot functionality is often embedded in customer services (e.g. banking, tax, logistics,
ecommerce). Chatbots might be fine-tuned on an organisation’s private datasets to answer
specific queries.
Examples
Apple’s Siri, Amazon Alexa, Google Assistant, ChatGPT, virtual assistants that
provide customer services in applications
General-purpose AI
General purpose AI is a new paradigm where very large AI models become competent at a
wide range of tasks, often without substantial modification or fine-tuning. General-purpose AI
systems are often LLMs.
Some general-purpose AI systems are effective at processing a variety of inputs.They can

process audio, video, textual and physical data and even complex medical or scientific data
with enough training. Part of the commercial appeal of foundation models is their capacity
for general-purpose applications.
Examples
GPT-3, GPT-4, BLOOM (BigScience large open-science open-access multilingual
language model), BERT (Bidirectional encoder representations from transformers),
Stable Diffusion and DALL-E. These are designed for general purposes and used for
downstream natural language processing (NLP) tasks.
6
Foundation model
Foundation models, sometimes called general-purpose AI systems, provide a basis for
future modification and fine-tuning for specific tasks. They are trained on vast amounts of data
at scale, including images, text, audio, video and other data types like 3D models, with less
emphasis on data quality so that they can be adapted to a wide range of downstream tasks.
Examples
Several companies, including OpenAI, Google, Meta (the parent company of
Facebook) and Stable Diffusion, own foundation models. Many open-source
foundation models can be downloaded from Hugging Face, GitHub, TensorFlow Hub
and PyTorch Hub for further use. Popular foundation models include GPT, BERT and
DALL-E 2.
Related terms: Machine learning, Transformer architecture, Open-source models,

General-purpose AI and Large language models
Frontier model
Frontier models are larger than foundation models, with more parameters. They are potentially
much more capable than existing models and raise additional safety risks particularly when
developed for critical applications such as health, social welfare, defense and military.
Challenges
Different definitions of frontier models use different criteria to describe them. These
definitions are not necessarily exclusive but rather adopt different assessment
approaches, potentially leading to alternative policy discussions and interventions.
• Based on risk
The UK government defines frontier AI as consisting of highly capable
general-purpose AI models that can perform a wide variety of tasks and
match or exceed the capabilities of today’s most advanced models (see AI
Safety Summit: Introduction).
The paper Frontier AI regulation: Managing emerging risk to public
safety by staff from Google, Open AI and the Centre for Long Term
Resilience defines frontier AI models as highly capable foundation models
7
that could possess dangerous capabilities sufficient to pose severe risks to

public safety.
• Based on computational power
Training frontier AI models requires computer power exceeding 1026 integers
or floating-point operations (FLOPS) (see the Executive Order on the Safe,
Secure, and Trustworthy Development and Use of Artificial Intelligence).
• Based on risk and computational power
The EU AI Act categorises a general-purpose AI model as having systemic
risk if it has high impact capabilities and the cumulative amount of
computation used for its training measured in FLOPS is more than 1025
(Article 51, Section 1 Classification rules, CHAPTER V GENERAL-PURPOSE
AI MODELS).
Transformer architecture
Transformer architecture makes LLMs possible. Put simply, transformers convert text. This
type of architecture was proposed in 2017 by 8 authors from Google in the paper Attention is
all you need.
The software processes information by tokens and handle sequences step by step.
Transformer models process long sequences simultaneously and capture context from the
input. They allow developers to train larger networks with more training data at a given scale.
Examples
Google’s BERT, OpenAI’s GPT, Meta’s BART, Bidirectional and Auto-Regressive
Transformers.
More information
The central idea of transformer architecture is attention and self-attention.
Attention: When encoding a sentence, the transformer architecture pays attention to

each of the words, figuring out which ones are needed to understand the whole
sentence and where they appear, giving these higher attention scores.
Self-attention: A mechanism to contextualise words by paying attention to other

words that make up its context in a body of text.
Visit Transformers: The Google scientists who pioneered an AI revolution for a

visualisation of this concept.
8
Transfer learning
Transfer learning is a model’s ability to apply information about one situation to another. In
this way, the model builds on its internal knowledge.
In GenAI, products must adapt and work well in different situations. It’s like teaching a model
skills in one area and then letting the model use those skills in other areas too. In this way,
models can learn from one thing and apply it to many different tasks, making them versatile
and useful.
More information
A current trained model knows how to distinguish dogs and cats from other animals
based on a lot of data labelling and data annotation. Now you want your model to
identify a new animal, for example rabbits. You don’t need to start fresh because
cats, dogs and rabbits share similar features like fur, legs and ears but with some
differences. You need to fine-tune the model to let it recognise rabbits. The model
should be able to apply what it has learned about dogs and cats, the fine-tuned
datasets, and apply that knowledge to recognise rabbits. Transfer learning makes the
training process faster.
Relevant terms: Machine learning, General-purpose AI and Large language models
Open-source and closed-source LLMs

Open-source LLMs have publicly accessible source code and underlying architecture,
allowing developers, deployers, researchers and enterprises to use, modify and distribute
them freely or subject to limited restrictions.
Closed-source LLMs have proprietary underlying source code and architecture. They are
accessible only under specific terms defined by their developers.
Open-source LLMs and closed-source LLMs can be accessed and deployed via application
programming interfaces (APIs), which are sets of rules or protocols that allow two software
programs to communicate with each other to exchange functionality. AI developers, deployers
or users can integrate data, services and functionalities from other applications through
provided APIs, rather than developing them from scratch. The following table presents
different features of open-source LLMs and closed-source LLMs.
9
Open-source LLMs Closed-source / proprietary LLMs
Description
Open-source LLMs often mean that a project team Closed-source or proprietary LLMs do not disclose their
has made their model architectures, source code source code, how the model is structured, the datasets
and weight parameters publicly accessible on source used to train the system or the training process. These
code repository hosting platforms such as GitHub, models are often developed commercially and may
GitLab and BitBucket. require licences or subscriptions for their use. Most
frequently, we access the model through a provided API
and use the model as provided by the owner.
Openness and adaptability
Other programmers or developers can read the Closed-source LLMs typically cannot be examined,
code, audit for bugs or security flaws and potentially audited or modified.
contribute improvements back to the project.
Examples
Think about open-source projects as public toolkits GPT-3 and GPT-4 by OpenAI: while GPT-2 is more
of different models for developers and users to build openly accessible, GPT-3 users must use the
their own tools. Access to open-source LLMs can provided API, and GPT-4 is a closed-source model.
unlock customisation opportunities. Gemini Pro (updated from Bard) and BERT by
OpenAI’s GPT models: GPT-2 and GPT-3 can Google:
generate human-like text based on input prompts. • Gemini Pro is open to developers and
The code and pre-trained models are available on
enterprise customers of Google. It is
GitHub, allowing developers / deployers to
accessed via the Gemini API.
experiment with and build upon them.
• BERT is openly published but many of
Hugging Face’s Transformers is an open-source
its applications and some advanced
library providing easy-to-use interfaces for accessing
derivatives need to be accessed via
and fine-tuning NLP models. The library includes
provided APIs. Turing-NLG by
pre-trained models like BERT, GPT and RoBERTa
Microsoft and its more advanced
(robustly optimised BERT pre-training approach),
and tools for training and evaluation. iterations can only be used internally
for enhancing Microsoft’s products
BLOOM is an open-source multilingual language
and services.
model.
• ERNIE by Baidu is accessible through
TensorFlow Models provides a collection of pre-
a subscription model via Baidu Cloud.
trained models and code examples for GenAI tasks
like image and text generation to allow developers to
explore and experiment with different approaches to
GenAI.
PyTorch Hub, like TensorFlow Models, offers a
collection of pre-trained models and code examples
for GenAI tasks contributed by the community.
10
Advantages
Free to use – many open-source LLMs are free of Legal protection – companies and startups that use
charge. closed-source LLMs usually have legal agreements
Customisability – open-source LLMs can be and terms of service that provide legal protection for
customised and modified for specific tasks. their businesses.
Transparency – some information about the model is Data security – closed-source LLMs may come with
made public, providing visibility into the model’s enhanced data security features.
inner workings. Scalability – with the company’s resources, closed
Community support – open-source LLMs are often source LLMs can be scaled more efficiently.
supported by communities of machine learning Regular update – with resources from proprietary
engineers and developers. models, LLMs can be constantly maintained and
Software development environments and platforms updated.
like Amazon Web Services (AWS), Hugging Face High-level support – companies may dedicate support
and Azure are emerging as marketplaces where you for model deployment, integration and
can develop tools using your selected model. troubleshooting.
Concerns
There are different ways of doing open source and Vendor dependency – companies using closed-
the term is contested. However, for GenAI there are source LLMs can be dependent on the AI or machine
open-source models and ecosystems that have their learning development company. There may be
own development, procurement and deployment challenges if the development company discontinues
pathways. support.
Limited resources – open-source LLMs often rely on Limited flexibility and lack of transparency – closed-
community volunteers with fewer resources for source LLMs are not suitable for experimentation due
development, addressing bugs and updating. to limited access to the model’s internal architecture
Security issues and misuses – exploitation of and training data.
uncensored LLMs for malicious services without Licensing and costs.
appropriate safety checks; fine-tuned data
containing harmful content might be free for public
use and accessible.
OpenAI’s Usage policies include universal policies,
safety best practices and moderation endpoint to
ensure their models are used in compliance with the
law and without causing harm.
11
Token
A token is the smallest unit of data used by GenAI systems.
For text-based models like GPT, a token can be a word, part of a word, punctuation marks,
spaces or other elements of the text. For image-generating AI models like DALL-E, a token is
a pixel of the image. For audio-based AI models like MusicLM, a token might represent a short
sound segment.
Tokens allow AI models to understand, memorise and generate meaningful responses. They
play a vital role in memory capacity, determining how much information the AI model can
recall.
Example
Different versions of ChatGPT come with different memory capacities. The ChatGPT
Plus plan offers up to 8,000 tokens. This means that ChatGPT Plus can only
remember information within the last 8,000 tokens during ongoing user interactions.
More information
Counting tokens is especially important when LLMs bill input and output tokens
differently. For example, with Anthropic’s Claude 3 Opus model, the cost for one
million input tokens is US$15 and the cost for one million output tokens is US$75.
Both input (the prompt or conversation history) and output (the model’s response)
tokens contribute to token usage. That is, if the input message uses 10 tokens and
the model’s response generates an additional 15 tokens, then the user, whether an
individual or an organisation, will be billed for a total of 25 tokens.
Reinforcement learning from human feedback (RLHF)

Reinforcement learning from human feedback (RLHF) resembles the human learning process.
When we learn a skill our teacher/instructor says things like ‘well done’ if we do something
right or ‘let’s try it again and differently’ to improve our skill.
12
Application
RLHF works a bit like this: human testers, trainers or programmers observe how AI
software conducts tasks and provide feedback through ranks and scores.
The model tries different actions based on the feedback and learns which actions are
more preferable based on the scores. Over time, with enough feedback and practice,
the model gets better at conducting tasks, making decisions and achieving its goals
even without human involvement.
Example
The system provides two possible outputs to a prompt, and the human indicates
which one they prefer.
AI products like chatbots are often fine-tuned through RLHF. Outputs from an LLM
are often conditioned by RLHF according to specific needs, norms, ethical guidelines
and standards to make AI products more accountable to the community.
Diffusion models
Diffusion models are used for AI image generation. They work by destroying training data (by
adding visual noise) and then learning to recover the data by reversing the noise.
Application
Imagine you have a picture of a dog. A diffusion model would take this picture and
add some random changes or destroy a part of it, like making some parts of the
picture blurry or changing some colours (i.e. adding noise). Then the model’s job is to
figure out what the original picture looked like before noise was added (i.e.
denoising). Diffusion models learn from lots of examples to figure out how to undo the
damage and reveal the original image.
Examples
DALL-E 2, Midjourney, Stable Diffusion
13
Inference
AI inference is the process of applying trained machine learning models to new, unseen data
to derive meaningful predictions or decisions. When users give a GenAI system a prompt, the
computational system used to produce the output is called inference. The energy required for
inference is much lower than training a model but is still significant and is a large part of the
cost of using a GenAI system.
Applications
Almost any real-world AI application relies on AI inference. Some of the most
commonly used examples include:
• LLMs – a model trained on sample text can parse and interpret texts it has
never seen before.
• Predictive analytics – once a model has been trained on past data and
reaches the inference stage, it can make predictions based on incoming data.
• Email security – a machine learning model can be trained to recognise spam
emails or business email compromise attacks, then make inferences about
incoming email messages, allowing email security filters to block malicious
ones.
14
GenAI Concepts Operational
Operational
This section begins by illustrating data flows within a GenAI system, followed by an
explanation of the general operation of a GenAI system through development,
distribution and deployment and use.
The development of GenAI starts with the acquisition of datasets, which often adhere to
data licensing agreements. To ensure software recognises datasets effectively, data
labelling and annotation are needed to prepare the data for training.
AI libraries and machine learning environments serve as crucial intermediaries in the

distribution phase, facilitating the deployment and use of the model.
During deployment and use, models can be fine-tuned by deployers or machine learning
/ operation teams to ensure they meet user demands.
15
Development
Datasets
LLMs are trained from different types of datasets. The material in datasets may be protected
by intellectual property or information privacy laws. Organisations like Common Crawl
scrape the public internet periodically and create gigantic datasets that can be used for
model training.
Application
Datasets for fine-tuning models are typically found on platform marketplaces like
Hugging Face, AWS and Azure. The terms are often commercial, and typically
include information about how the dataset was created and its compliance with
relevant legal rules.
Most AI companies have trained their data on the open web, but this has prompted
several legal challenges under copyright and privacy law.
Data licensing
Entities that control large amounts of text, visual, musical, code or video content may license
it to AI firms for a fee. Datasets may also come with licensing stipulations that dictate what
others can do with that dataset, for instance whether they can share it, modify it or use it for
commercial purposes.
Examples
Organisations with large desirable datasets like Reddit or New York Times are now
seeking financial compensation when AI models train on their data.
On the other end, AI companies have arguments against paying for copyrighted
content:
• Adobe claims that they license all the copyright material used to train their
commercial models.
• Google claims that AI training is like reading a book. If training could be done
without creating copies, there would be no copyright questions.
• Microsoft claims that changing copyright law could hurt small AI developers.
16
Developer
AI developers belong to a broader category of programmers and engineers. They write code
and algorithms to enable machines to perform tasks that normally require human
intelligence. They build models in-house by training on a mix of public and private data for
various applications, from chatbots and virtual assistants to self-driving cars.
In emerging AI regulations, developers are responsible for creating AI software products and
complying with regulations.
Example
The EU AI Act imposes different obligations on AI ‘developers’, ‘deployers’ and
‘users’. Many of the regulatory obligations around safety and auditing must be
satisfied by developers before their product can go to market.
With organisations and individuals assuming overlapping roles within the AI supply
chain, third-party organisations have emerged, offering services like Future of Life’s
The EU AI Act Compliance Checker to allow individuals and organisations to
identify their roles and obligations under certain legislation.
17
Data labelling and annotation

Data labelling and annotation describe the process of tagging or labelling text, images, audio
and video data for AI training.
Data annotation Data labelling
Meaning
Data annotation provides information about a set of Data labelling attaches meaning to different types of
collected data so machine learning algorithms can data to train a machine learning model. It identifies a
easily recognise the dataset. single entity from a set of data.
Information about data, or known as metadata,
includes description, administrative details, statistics
and legal information about the dataset. Examples of
metadata may include how the data was created, the
means of creation, the purpose of the data, the time
and date of creation, the location of creation, the
standards used, file size, data quality and the source
of the data.
Purpose
Data annotations can be used for visual-based Data labels are used to identify dataset features for
perception models. It helps models to identify and NLP algorithms. Data labelling is used for training
process objects and people in images and videos, a advanced algorithms to recognise patterns within the
capability known as ‘computer vision’. datasets in the future.
Applications
Data annotation is a fundamental element in creating Data labelling is more complicated than annotation. It
training data for computer vision. Annotated data identifies key features in the data while minimising
helps to train machine learning algorithms to see the human involvement. Real-world use cases include
world humans see. NLP, audio and video processing, computer vision
etc.
18
Distribution
Supply chains
An AI supply chain refers to the process of creating, sourcing and integrating the
components needed to develop and deploy AI systems or products.
This diagram builds on Ali Arsanjani’s The generative AI life-cycle.
19
Concerns
Developers often use pre-made software parts, which makes it challenging to know
who’s responsible for following the rules and regulations. Understanding AI supply
chains requires us to ask:
• where the data used to train the AI comes from

• whether the AI model is open for anyone to use or owned by an entity
• how organisations tweak an AI system by using different datasets and fine-
tuning models.
AI libraries
AI libraries include pre-written codes and algorithms for common AI tasks, such as data
preprocessing, model training and evaluation.
Examples
PyTorch is one of the most popular machine learning libraries, alongside
TensorFlow. It offers free and open-source software released under the modified
BSD licence.
Development: PyTorch was developed by MetaAI and is now part of the Linux
Foundation
Notable applications: Tesla Autopilot, Uber’s Pyro, Hugging Face’s

Transformer
TensorFlow is a free and open-source software library for machine learning and AI.
It is particularly used for training and inference of deep neural networks.
Development: TensorFlow was developed by the Google Brain team
Notable applications: GE Healthcare, Twitter’s ranking system, the image-

captioning software DeepDream
The Hugging Face ecosystem, including Transformers and the Hugging Face Hub,
contains libraries for tasks such as dataset processing, model evaluation and
machine learning demos. Its transformers library is built for NLP applications.
Hugging Face Hub also allows users to share machine learning models and datasets.
Development: Hugging Face is a French–American company that originally

developed a chatbot app targeted at teenagers
20
Notable applications: partnership with AWS – Hugging Face’s products are

available to AWS customers and can be used as building blocks for their custom
applications
Machine learning environments

Machine learning environments are where machine learning models are built, trained and
deployed. Provided by the big tech corporations like Amazon and Microsoft, these
environments contain fully managed infrastructure and tools to enable reproducible,
auditable and portable machine learning workflows across different environments. Deployers
can run training scripts or host service deployments, often referred to as computer targets,
within these environments.
Examples
Microsoft Azure
• Amazon SageMaker Studio and SageMaker Studio Classic
21
Deployment and use
Fine-tuning
Fine-tuning happens during model development and deployment. It involves modifying a
trained AI model with a smaller, targeted fine-tuning dataset. Fine-tuning maintains the
original capabilities of a pre-trained model while adapting it to suit more specialised use
cases.
Example
Imagine you have a model that’s been trained to recognise birds in pictures, but you
want it to be better at recognising Swift Parrots specifically. You deploy a model you
found from an AI library, fine-tune it by feeding it lots of pictures of Swift Parrots,
adjust its parameters and even score the performance (i.e. reinforcement learning)
until it gets better at identifying the target.
More information
Large training datasets are available from development platforms like Github, Azure
AWS, Papers with Code and Hugging Face.
The Data Provenance Initiative identified that across Github, Papers with Code and
Hugging Face more than 70% of pre-trained datasets had no data licences. For those
that did, roughly half were incorrect and were more permissive than the dataset
creators had intended.
Read more in Public AI Training Datasets Are Rife With Licensing

Errors and The Data Provenance Initiative: A Large Scale Audit of Dataset
Licensing & Attribution in AI.
22
Deployer
While AI developers create the AI system, AI deployers make them available for use in real-
word applications. In some cases, the distinction between a developer and a deployer is
minimal. However, deployers may be the organisations that make AI tools available to
others.
Applications
Let’s think about an AI model being composed of 4 layers: infrastructure, model
development, application development and applications. Different levels are open to
deployment and associated with different levels of technical capacities.
The new AI stacks
AI deployment can be as simple as a dashboard at the application level or as complex as the

augmentation of a foundation model through prompt engineering or specific GenAI
architecture at the infrastructure level.
23
As shown in the pie chart, open-source AI tools are deployed for coding, communications
(e.g. WhatsApp bots, Slack bots), information aggregation (e.g. asking a model for meeting
summaries), image production, workflow automation and writing.
This section references Chip Huyen’s research blog.
User
Users are the individuals or organisations that interact with AI technology through text, voice,
images or other inputs. Users can also be anyone whose rights and activities are influenced
by AI system outputs.
Example
We are users when we chat with virtual assistants like Siri or Google Assistant.
Even though individuals do not directly engage with AI models, individuals’ rights and
activities can be affected by how organisations use AI models.
For example, if AI is used to screen job applications and select candidates for
interviews, the AI decision-making system might influence the outcomes.
24
GenAI Concepts Regulatory
Regulatory
This section explains the risks and concerns with GenAI systems and how existing
regulations, ethical guidelines and practices address these.
GenAI systems raise many concerns, such as privacy breaches, hallucination, safety
issues, lack of transparency and ecological impacts.
Existing regulations and legislation, such as copyright laws, data governance and
security, demonstrate our capacity to regulate some aspects of GenAI systems.
Complementing these are emerging AI-specific guidelines and regulations, including AI
standards, guardrails, regulatory sandboxes, human oversight, AI auditing, explainable
AI, post-market surveillance and monitoring and FLOPS to respond to increased risks
associated with AI.
25
Risks
AI regulation is developing as risk-based regulation. This approach imposes obligations
depending on the risk level of the system. Most regulatory proposals have 3 risk categories:
low risk, medium risk and high risk.
The measurement and mitigation of AI-generated risks requires human oversight, AI auditing,
transparency and explainability.
Challenges
Measuring risks is not easy! Risk is a spectrum and depends on a number of factors
that are difficult to quantify. Standards are being created to help developers,
deployers and users specify the appropriate level of risk for the AI system.
Concerns
Using AI to play chess is low risk compared to using AI for staff recruitment, which
generates medium risks (e.g. gender or racial discrimination). Using AI-enabled
robots for medical surgery and in self-driving cars to make real-time decisions
generates high risks, as such risks are systematic, perpetual or hard to reverse.
Privacy and data protection

The wide use of AI systems presents challenges to privacy. The datasets used to train
foundation models and other machine learning systems often include personal information
without consent. There have been cases where chatbots reproduce personal information from
training data in response to prompts. Privacy and data protection laws impose obligations on
entities that collect and process personal information.
These obligations relate to collecting, using and disclosing personal information; data quality
and security; and deleting, minimising or deidentifying personal information. Protecting
individuals’ data often requires that individuals consent to the use of their personal information,
while systems limit use to only those that an individual would reasonably expect and that
relate to the original reason for collection. Privacy and data protection laws also create rights
for individuals to have access to information about them held by other entities.
26
Legislation
The Privacy Act 1988 is a federal law which does not cover local, state or territory
government agencies, except the Norfolk Island administration. Most Australian
states and territories have equivalent legislation covering their public sector agencies
(see State and territory privacy legislation).
For example, in Victoria, individuals’ privacy rights are protected by the Privacy and
Data Protection Act 2014 (Vic) (PDP Act). The PDP Act contains 10 information
privacy principles (IPPs), which are the core of privacy law in Victoria and set out
the minimum standard for how Victorian public sector organisations should manage
personal information.
Personal information
Under the PDP Act, personal information is information or an opinion about you
where your identity is clear or where someone could reasonably work out that it is
related to you. It can include your name, email address, postal address, phone
number, signature, fingerprint, photographs or surveillance footage of you, comments
written about you (whether true or not), your financial details and more (see more at
OVIC’s Your privacy rights).
Challenges
AI models are trained on large datasets and our real-time inputs are sometimes
included in the AI training process. Organisational data and personal and sensitive
information may then be inadvertently disclosed to companies in control of the AI
software being used.
When analysed or correlated, a combination of seemingly non-personal information

can make individuals identifiable.
Further, AI’s predictive capabilities can exceed what an individual discloses, resulting
in the collection of new personal information generated by the AI model’s output (read
more at OVIC’s Artificial Intelligence – Understanding privacy obligations).
More information
The Office of the Australian Information Commissioner reported a 19% increase in
data breaches from July to December 2023, with notable impacts in health services,
finance, insurance, retail and the Australian government, where our sensitive
27
information is stored (read more in OAIC’s Notifiable data breaches report July to
December 2023).
If an organisation uses GenAI services that include accessing a model through an

API, then the organisation should ensure that those data flows do not breach relevant
privacy and data security legislation, such as the Victorian Information Privacy
Principles and Australian Privacy Principles.
Software providers must provide assurance around how their data is stored,
processed and secured.
Hallucination
Hallucination refers to AI models making up facts to fit a prompt’s intent. When an LLM
processes a prompt, it searches for statistically appropriate words, not necessarily the most
accurate answer. An AI system does not ‘understand’ anything, it only recognises the most
statistically likely answer. That means an answer might sound convincing but have no basis in
fact. This creates significant risks for organisations that rely on chatbots to give advice about
products or services because that advice might not be accurate.
Real case
In 2023, 2 lawyers in New York used ChatGPT in their legal research for a personal
injury case. Their legal brief, submitted to the court, included 6 fake case citations
generated by ChatGPT (i.e. hallucinated detail). The lawyers then conducted a fact
check with ChatGPT, which approved the results generated. The lawyers and the law
firm received penalties as a result.
Concerns
AI systems cannot ‘tell the truth’; they are prediction models. Some people think AI
systems only get things wrong occasionally while otherwise telling the truth. That is
not true. AI systems make output mistakes based on the input data they use, so it is
important to verify the accuracy of the output before relying on the model. This is
especially important when people rely on AI systems to make decisions that affect
themselves or others.
28
Tip
One way to identify AI-generated images is by looking at the detail (e.g. hands, teeth,
hair, unnaturally smooth skin, accessories, watermarks in the background,
stereotypes). Because current AI systems struggle with complex detail they might be
trained on simpler stereotypes, leading to odd or unrealistic results.
Safety
In AI regulation, AI safety means that AI should be designed, developed and used in a way
that is human-centric, trustworthy and responsible (see The Bletchley Declaration by
Countries Attending the AI Safety Summit, 1–2 November 2023).
The main concern of AI safety is to make sure that AI systems are developed, deployed and
used in ways that align with the norms and values of the broader public or specific user
groups.
Guidelines and legislation

AI safety is central to numerous laws and AI ethics guidelines developed at national
and supranational levels, such as The Bletchley Declaration, the EU AI Act, The
Executive Order on the Safe, Secure and Trustworthy Development and Use of
Artificial Intelligence.
The Australian government is already undertaking work to strengthen existing laws

(e.g. the Online Safety Act 2021) to address known harms with AI.
Along with ethical guidelines, including Australia’s Artificial Intelligence Ethics

Framework and Australia’s AI ethics principles, the Australian government’s
interim response to safe and responsible AI consultation describes actions to
help ensure that AI is safe and responsible.
Measurement
Safety measures include testing AI systems thoroughly before using them, designing
systems to be robust enough to handle unexpected situations without making
mistakes and having guardrails in place.
29
Transparency
AI transparency is a fundamental regulatory and ethical requirement; however, transparency
has different meanings for different stakeholders (e.g. developers, deployers, users,
regulators).
Transparency represents how well users, regulators and other stakeholders understand how
the system operates and how it was built. This is important for ensuring trust in AI outputs.
Transparency is related to the openness and explainability of AI systems.
Forms of transparency
Different stakeholders will be interested in different forms of transparency.
For general public users, transparency can be a sense of what the AI system is doing
and why.
For a regulator, transparency can enable auditing how a prediction and decision is
made by AI systems in detail or mean transparency in the data used to train a
system.
Engineers might be more interested in the internal mechanism and parameters of AI

systems.
Related term: Explainable AI
Copyright
Copyright law typically applies to original works of authorship, such as literary works, artistic
works and computer programs. Copyright material has been included in AI training datasets
without permission or rights clearance.
30
Challenges
The application of copyright to datasets raises complex issues and how rights-holders
opt out or demand licensing fees will depend on how the law develops and will vary in
different jurisdictions. Some jurisdictions may create exceptions for model training,
and others may determine that the training process does not replicate the
‘expressive’ part of a copyright work and thus is not a breach.
Real cases
There are currently court cases and policy reforms addressing how to manage the
inclusion of copyright material in datasets. There is also the risk that GenAI responds
to prompts with outputs that violate copyright. This might be an exact reproduction of
a text or image or the unauthorised use of a copyrighted character.
Several coders have brought a lawsuit against GitHub, Microsoft and OpenAI
because GitHub Copilot was trained and developed on existing open-source code,
leading to questions of attribution.
Several visual artists filed a class-action lawsuit against Stable Diffusion,

Midjourney and DreamUp, all of which generate images based on text prompts from
users. The case alleges that these AI tools violate copyright by scraping images from
the internet to train their models.
Getty Images alleges that Stable Diffusion’s use of its images to train models
infringes on copyrights and that some images even contain the Getty watermark.
Many GenAI systems providers now indemnify users in case an AI output breaches
copyright.
Data governance and security

Data governance refers to the system of rules for management and control of data.
Data governance requirements may involve compliance with laws, standards or ethical
guidelines to ensure data within an organisation is collected, stored and used appropriately.
Within an organisation, data governance responsibilities include ensuring data quality (i.e.
data is accurate) as well as data security (i.e. data is appropriately encrypted and protected).
31
This may mean additional care when using organisational data to fine-tune a GenAI system or
when using organisational or sensitive data as part of a prompt.
Data life cycle management refers to the processes for managing data from creation to
disposal, including data archiving and removal.
Challenges
The Australian government will develop a whole-of-government Data Governance
Framework for public services to ensure data quality, privacy, authority and
innovation.
There are also sector-based data governance frameworks, for example, by

the Australian Commission on Safety and Quality in Health Care, Australian
Institute of Health and Welfare, National Archives of Australia and the National
Health and Medical Research Council.
In Victoria, there is the Victorian Protective Data Security Framework and

Standards and the Privacy and Data Protection Act 2014.
AI standards
International AI standards serve as a global governance mechanism to help achieve AI policy
goals. Standards organisations create standards through stakeholder input and cover various
technical and regulatory dimensions of AI systems. Compliance with standards may be
required for AI products to be market-worthy; procurers may also require compliance with
specific standards. In the EU system, if developers comply with standards, then they likely
comply with their obligations under the EU AI Act.
Applications
A wide range of national, regional and international organisations are engaged in
developing and refining AI standards.
ISO/IEC 42001:2023 Information technology – Artificial intelligence –

Management system is the world’s first AI management system standard that
provides valuable guidance for this rapidly changing field. It addresses challenges
posed by AI technologies, including ethics, transparency and continuous learning. For
organisations, the standard sets out a structured way to manage risks and
opportunities associated with AI, balancing innovation with governance.
• ISO/IEC 23894:2023 – Information technology – Artificial intelligence –

Guidance on risk management provides guidance on how organisations that
32
develop, produce, deploy or use AI products, systems and services can

manage risks related to AI.
• ISO/IEC 23053:2022 – Framework for Artificial Intelligence (AI) Systems
using Machine Learning (ML) establishes an AI and ML framework for
describing a generic AI system using ML technology.
• IEEE is an engineers’ professional organisation with a subsidiary Standards
Associations (SA) whose standards address protocols for digital products,
including ethernet and wi-fi. Its AI standardisation processes are part of a
larger IEEE Global Initiative on Ethics of Autonomous and Intelligent
Systems.
• The ITU (International Telecommunication Union) has historically played a
role in standards for information and communications technologies,
particularly in telecommunications. Following the 2018 AI for Good Global
Summit, ITU created a number of Focus groups on AI in relation to health,
environmental efficiency, autonomous and assisted driving, and data.
• As part of European AI regulation, the European Commission has asked the
European Committee for Standardisation (CEN) and the European
Electrotechnical Committee for Standardization (CENELEC) to
develop harmonised technical standards for companies that wish to comply
with the EU AI Act.
Specialised bodies may also participate in creating international standards. These

bodies can be treaty organisations such as the International Atomic Energy Agency
and the International Civil Aviation Organization.
More information
According to the OECD AI Policy Observatory, there are more than 1,000 AI
policies and principles developed by governments, international organisations and
private entities.
In Australia, there is the Australia’s Artificial Intelligence Ethics

Framework and Australia’s AI Ethics Principles. The Australian government’s
interim response to safe and responsible AI in Australia consultation highlights
that more needs to be done to ensure the safe and responsible development and
deployment of AI.
33
Guardrails
AI guardrails are predefined standards, limitations and operational protocols to prevent AI
systems from making decisions or taking actions that could lead to harmful or unintended
consequences.
Guardrails are often sets of algorithms or rules that filter and edit inputs (i.e. prompts) and
outputs of GenAI systems to ensure that outputs comply with legal and safety requirements.
These guardrails are designed to prevent outputs from breaching copyright, produce political
misinformation, create biased or discriminatory information or generate hate speech. This has
proved very difficult to achieve in practice but remains an important aspect of AI system
design.
Jailbreaking
Jailbreaking refers to techniques for getting around a system’s guardrails. This might
mean tricking a chatbot into generating outputs that would otherwise be a violation of
its policies. Most LLM systems are susceptible to jailbreaking, meaning all guardrails
have limited efficacy.
Regulatory sandbox
Regulatory sandboxes are software testing environments where businesses can conduct
limited testing of innovations and test their regulatory status.
This allows businesses to receive feedback from experts, including regulators, investors,
innovators and other stakeholders, regarding the potential and viability of the innovation.
Application
Regulators and innovators use regulatory sandboxes to offer individualised legal
guidance. Sometimes regulatory sandboxes are used to make legal exceptions for
certain innovations, enabling innovators and the public to experience the impact of
new technologies as if they had already been established as safe and effective.
34
Legislation
Regulatory sandboxes have been included in the EU AI Act Article 53 and by
the Australian Securities & Investments Commission.
Human oversight
Human oversight is the requirement that human actors oversee the output of AI systems to
ensure that systems create accurate and accountable results. There are different forms and
degrees of human oversight, depending on the context and purpose of the AI system. Human
oversight is usually required in high-risk systems.
Application
Humans can intervene in AI systems in different forms and at different degrees:
• Humans in the loop: even though AI systems can do many tasks on their
own, humans can make decisions, provide approvals or improve the system
by correcting or preventing potential AI errors. Humans in the loop may not
need say ‘yes’ or ‘no’ every step of the way as long as there is still a person
watching over the process to make sure everything runs smoothly.
• Humans in command: means that humans have ultimate control over the AI
system and make all the important decisions.
Sometimes human oversight requires a final human decision-maker, but often this is
not the best way to implement oversight or may be impossible.
Legislation
The EU AI Act Article 14 requires that high-risk AI systems have appropriate human
oversight from design to operation. Human oversight should be proportional to the
potential impact and severity of harm that the AI system can cause.
The following Australian documents refer to and explain the EU rules:
• Australia’s AI Ethics Principles

• The OAIC submission to the Department of Industry, Science and Resources
– Safe and responsible AI in Australia discussion paper
35
AI auditing
AI auditing involves humans – normally researchers, programmers and regulators – looking
closely at AI systems to evaluate risk and ensure AI systems act fairly and safely, and comply
with relevant laws, regulations and ethical standards.
Different AI auditing methods have different advantages and limits:
• Technology-oriented audits focus on the properties and capabilities of AI systems.

• Process-oriented audits focus on technology providers’ governance structures and
quality management mechanisms.
Some auditing methods may be simple compliance checkboxes; others may be

comprehensive assessments of how AI systems might affect users and other stakeholders.
Application
AI auditing is a rapidly growing field of industry practices. Emerging courses,
certifications and associations aim to professionalise the AI auditing industry.
Explainable AI (XAI)
Explainable AI or XAI is a set of tools, techniques and algorithms designed to produce high-
quality interpretable, intuitive, human-understandable explanations of AI decisions. Many
emerging AI regulations require some degree of explanation for higher risk system outputs.
The goal is not to prioritise complete explainability over performance or vice versa.
Organisations should disclose the limits of transparency in the system to find a balanced
approach that considers the risks and benefits of each AI application, while also considering
human and environmental implication.
Application
The explainability of AI-driven decisions is linked to the transparency, fairness and
trustworthiness of the decision made, and sometimes the interpretability of AI
systems.
36
Advocacy for explainable AI Challenges to explainable AI
Governments Complexity vs Simplicity

The Australian Government’s The AI in It is complex to make AI systems explainable.
Government Taskforce outlines transparency Technical complexity of AI systems has made it
and explainability as one of the 4 key principles to challenging for end users to understand how a
ensure the Australian public sector implements AI decision is made by AI systems. Oversimplification
safely and responsibly. can lead to unfaithful explanation of the decision. It is
challenging to strike a balance of accuracy,
explainability and future accessibility.
Research institutes Transparency vs Vulnerability

OpenAI, as an AI research and development With the decision-making process being made
company, has been developing a new tool to explainable, it can also make the system open to
explain large language models’ behaviours. exploitation and attack.
Civil society Generalisability vs Particularity

Advocacy entities like the Electronic Frontier General explainability of AI systems cannot capture
Foundation (EFF) and the American Civil specific or intricate patterns in data. Simplification
Liberties Union (ACLU) consistently stress the and generalisation can lead to potentially lower
significance of AI transparency, especially when performance.
outcomes directly affect individual rights. More complex models can identify and use more
particular and complicated patterns. Their outputs
can be more difficult for people to interpret.
Not all AI applications require the same degree of
explainability. For example, AI used in healthcare
diagnoses should require higher transparency.
Post-market surveillance and monitoring

Post-market surveillance refers to monitoring the ongoing performance and safety of an AI
product or service after it has been released to the market.
Application
Post-market surveillance may involve collecting and analysing data on the use of the
product or service in real-life conditions. If issues are detected, information from post-
market surveillance can help corrective action, such as updating product design,
improving user instructions for use or product recall.
37
Example
In 2022, Meta pulled the public demo of Galactica, an LLM designed for science
research, 3 days after its launch due to its inability to distinguish facts from falsehood
(i.e. hallucination).
In 2024, Google paused its Gemini AI model’s image generation of people due to
inaccuracies in some historical depictions.
Floating-point operations per second (FLOPS)

Floating-point operations per second (FLOPS) in the context of computing and artificial
intelligence are often used to measure the processing power or performance of hardware
devices like CPUs (central processing units) and GPUs (graphic processing units).
Sometimes frontier models are defined by FLOPS, and some regulatory proposals have
stricter compliance obligations for systems that surpass FLOPS thresholds.
Higher FLOPS values indicate higher computational power, more complex AI models, faster
data processing, better graphic performance and higher energy consumption.
Application
The US Executive Order on AI specifies different regulatory requirements for
systems that use different amounts of computing power. It states that federal safety
and security reporting requirements apply to any model trained using either:
• a quantity of computing power greater than 1026 FLOPS

• primarily biological sequence data and a quantity of computing power greater
than 1023 FLOPS.
So far, these thresholds are beyond anything on the market.
The largest GPT-3 model required approximately 3.14E+23 = 3.14*1023 FLOPS for
training, with 174,600 parameters and 300 billion training tokens.
There are research institutes, such as Epoch AI, tracing and documenting information about
AI models, including estimates of their training computation power.
38
Ecological
Training, deploying and using AI systems contribute to the global CO2 emissions. Typically,
more powerful AI models require more energy. The servers that power AI models also
generate considerable heat and are often water-cooled. The amount of water needed to train
an AI model is immense. A team of researchers disclosed that “training GPT-3 in Microsoft’s
state-of-the-art U.S. data centers can directly evaporate 700,000 liters of clean freshwater, but
such information has been kept a secret.” And running GPT-3 inference for 10-50 queries
evaporate 500 mililitres of water depending on when and where the model is hosted.
LLMs are among the biggest machine learning models, spanning up to hundreds of millions of
parameters, requiring millions of GPU (graphic processing units) hours to train and emitting
carbon in the process.
39
GenAI Concepts
40
admscentre.org.au/genai-concepts

GenAI Concepts

Uploaded by

Copyright:

Available Formats

GenAI Concepts

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

GenAI Concepts

Uploaded by

Copyright:

Available Formats

GenAI

Developed as a collaboration between the Authors Webpage

Suggested citation (document)

Suggested citation (webpage)

Copyright © 2024 Fan Yang, Jake Goldenfein and

Generative AI or GenAI ................................................................................................................ 2

Prompt and prompt engineering ................................................................................................... 2

Machine learning .......................................................................................................................... 4

Large language models (LLMs) ................................................................................................... 5

Knowledge cut-off date ................................................................................................................ 5

Foundation model ........................................................................................................................ 7

Frontier model .............................................................................................................................. 7

Transformer architecture .............................................................................................................. 8

Transfer learning .......................................................................................................................... 9

Open-source and closed-source LLMs ........................................................................................ 9

Reinforcement learning from human feedback (RLHF) ............................................................. 12

Diffusion models ......................................................................................................................... 13

Data licensing ............................................................................................................................. 16

Data labelling and annotation..................................................................................................... 18

Supply chains ............................................................................................................................. 19

Deployment and use .................................................................................................................... 22

Privacy and data protection ........................................................................................................ 26

Data governance and security ................................................................................................... 31

Regulatory sandbox ................................................................................................................... 34

Human oversight ........................................................................................................................ 35

Explainable AI (XAI) ................................................................................................................... 36

Post-market surveillance and monitoring ................................................................................... 37

Floating-point operations per second (FLOPS) ......................................................................... 38

Developed as a collaboration between the ARC Centre of Excellence for Automated

GenAI Concepts can also be accessed as a regularly updated website at

Types of GenAI models in the market

Prompt and prompt engineering

Basic structure of a prompt:

Acting as a [role] perform [task] in [format] in the [style]

Acting as a [role] Perform a [task] Show as [format] In the [style]

Marketer Headline Table Formal

Types of machine learning

• Supervised machine learning:

• Hybrid machine learning:

• Healthcare: ProMed (the Program for Monitoring Emerging Diseases) offers

• Finance: Fraud detection, automated trading activities and financial advisory

• Marketing: Product recommendations and news feeds on social media

• Transportation: Self-driving vehicles, real-time tracking and last-mile delivery

Large language models (LLMs)

Related term: Machine learning

Knowledge cut-off date

Some general-purpose AI systems are effective at processing a variety of inputs.They can

Related terms: Machine learning, Transformer architecture, Open-source models,

that could possess dangerous capabilities sufficient to pose severe risks to

Attention: When encoding a sentence, the transformer architecture pays attention to

Self-attention: A mechanism to contextualise words by paying attention to other

Visit Transformers: The Google scientists who pioneered an AI revolution for a

Relevant terms: Machine learning, General-purpose AI and Large language models

Open-source and closed-source LLMs

Open-source LLMs Closed-source / proprietary LLMs

Openness and adaptability

Reinforcement learning from human feedback (RLHF)

AI libraries and machine learning environments serve as crucial intermediaries in the

Data labelling and annotation

Data annotation Data labelling