Version 1.0.

1
Prepared by:
EMERGING TECHNOLOGIES TEAM

This document contains

proprietary information
of the United Nations.

Generative AI
Information contained
herein is to be used solely
for the purpose submitted,
and no part of this
document or its content

Primer
shall be reproduced,
published or disclosed to
a third party without the
express permission of the
United Nations.

What is Generative AI?

Generative AI is a subfield of artificial intelligence (AI) and machine learning

(ML) that involves the creation of original data or content, including images,
video, text, code and 3D renderings. This subfield has been developing over
several decades and is rapidly evolving, due to advances and availability
in computational power, large datasets and significant improvements in
machine learning algorithms. Generative AI models are based on deep
learning algorithms that learn to recognize patterns and relationships from
vast amounts of input data, which then generate new outputs that are similar
in style and structure to the data they were trained on.
The ability of these models to self-formulate new and varied outputs
represents a paradigm shift in the field of AI because they are not being
explicitly programmed to follow pre-determined rules, or generate specific
outputs, like other AI systems. This will likely lead to a change in how we
interface with computers, and more broadly, in how we access, understand,
and produce knowledge and information. Figure 1: Subfields of artificial intelligence

Core Concepts

Building Blocks of Generative AI

» Supervised and semi-supervised » Generative Adversarial Network » Natural Language Processing
learning - A machine learning (GAN) - A machine learning (NLP) - A subfield of AI that is at
technique that helps algorithms model that uses two neural the intersection of linguistics,
learn to recognize patterns and make networks—a generator and a computer science and machine
predictions based on categorized discriminator—to produce new learning. NLP enables computer
or labeled data. In the case of data that is similar to a given data programs to process and
semisupervised learning, algorithms set. GANs have become a popular analyze large amounts of natural
are trained on both labeled and approach for generative AI in language data. It uses a range
unlabeled data to detect patterns various domains, such as image of computational methods and
and make predictions. and video generation. algorithms to allow machines,
such as chatbots and voice
» Deep learning - A machine learning » Transformer - A type of neural assistants, to understand and
technique that uses layers of neural network that uses encoders and mimic written or spoken human
networks to process data and make decoders to generate the best language.
decisions. probability for the following
word in a sentence. Transformers » Generative Pre-trained
» Neural networks - An AI method enable the development of Transformer (GPT) - A type of
that simulates the structure and powerful generative models. LLM developed by the research
function of the human brain. Neural lab OpenAI that uses deep
networks process information » Large Language Model (LLM) - learning and NLP techniques.
through interconnected nodes that A statistical AI model that is GPT underlies the user-facing,
are organized in a layered structure. trained on massive amounts general-purpose chatbot,
This computational model serves as of text data and predicts the ChatGPT, which produces human-
the basis of deep learning and is used probability of sequences of words like conversational responses in
in various types of generative models. to produce human-like text reaction to short user prompts.
responses.

Generative AI Primer
UN OICT Emerging Technologies Team 1
 Version 1.0.1

Co-pilot
Assistants Generative AI Applications
Co-pilot
and Infrastructure
APPLICATION LAYERS

PowerPoint Assistants
Presentations
Co-pilot PowerPoint
Translation Assistants Presentations

Report Code Image

Content Generation Generation

Summarization Code Consumer

Documentation /Social
Support
(Chat/Email) Text to SQL Media/ Translation
Advertising
General Web App Voice Video Editing 3D Models/
Writing Builders Design Synthesis /Generation Scenes

TEXT CODE IMAGE SPEECH VIDEO 3D

Figure 2: Generative AI applications (Source: Adapted to Huang S. and Grady P., 2022)

Value Proposition
Proponents of generative AI believe that this set of converging technologies is poised to revolutionize the economy, spur
productivity and transform industries—from education, healthcare and finance to infrastructure development. According
to new research, it is estimated that generative AI systems could increase annual global Gross Domestic Product (GDP) by
7 percent over a 10-year period (Briggs and Kodnani, 2023). Early data suggests that a massive transformation is already
underway, as Big Tech companies have begun to quickly roll out and integrate new generative AI tools and features into
their products, including search engines and office suite software. The intense competition in Silicon Valley to develop and
deploy these tools, however, has prompted concerns among many regulators and ethical AI researchers. Some of these
concerns include the risk of perpetuating bias and discrimination, spreading mis- and disinformation, and infringing on
intellectual property rights. Discussions regarding generative AI’s value proposition are full of debate and uncertainties.
Below, we explore contrasting views on the potential benefits of these technologies, highlighting three interwoven
features at the core of its value proposition: efficiency, personalization, and creativity and innovation.

WHAT ADVOCATES SAY WHAT SKEPTICS SAY WHY THE UN SHOULD CARE

EFFICIENCY

There are many opportunities for While generative AI systems can By automating certain functions, generative AI systems
businesses and organizations to help an organization run more can free up time and resources for more high-value tasks,
maximize their potential using efficiently, training and operating such as policy development and advocacy.
generative AI tools. One such benefit these models is expensive and can
One key area where generative AI can be leveraged is in
can include increased efficiency by be cost-prohibitive. This is due in
writing and research-related tasks. Generative models
automating tasks that involve complex part to the specialized hardware
can be trained to generate text for various purposes,
reasoning, pattern recognition and large and significant amount of memory
such as drafting grant proposals, or summarizing large
data sets. In the long term, this type and storage space these systems
documents.
of automation could reduce business require to support the high volume of
costs and increase productivity. It calculations models typically produce. Generative AI can also assist with coding tasks by
could also accelerate the speed of providing descriptive code documentation. This can help
With continued advances in the field
iteration, allowing for faster learning and developers to understand complex code bases more
of AI, however, costs associated with
improvements, which can drive business quickly and easily, leading to more efficient and effective
generative AI models are expected to
performance. software development. In addition, generative AI can be
decrease over time, which could make
integrated into software interfaces, allowing end-users to
According to economists, foundation them more accessible. Researchers
interact with their software using natural language. This
models could impact every sector of estimate that millions of jobs across
can improve their experience and make software more
the economy and lead to significant large economies could be exposed
accessible to non-technical end-users.
economic growth. It is predicted that to some degree of automation.
generative AI will likely create disruption Certain white-collar positions are Generative AI tools can facilitate near-real-time translation
across some industries and drive considered to be at greater risk of and transcription, which could enhance the delivery of
demand for new skills. This includes being altered or displaced by text critical services to underserved communities. For instance,
roles for prompt engineers and data generation tools such as ChatGPT. the UN could use generative AI to provide real-time
practitioners that can integrate, fine- This includes jobs in tech (e.g., coders, translation services to refugees or to transcribe speeches
tune, and improve generative AI models computer programmers), media (e.g., at international conferences.
into existing products and pipelines. advertising professionals, journalists)
Generative AI models, however, lack an inherent
It also creates greater demand for AI and finance (e.g., financial analysts
understanding of the content they generate, and instead
safety professionals to manage the AI and advisors), as well as educators
rely on predicting the most probable next words based on
systems’ shortcomings and work to and customer service agents.
the patterns derived from a large corpus of textual data.
ensure their responsible use.
Hence, they can be considered as personal assistants
rather than authoritative decision-makers.

Generative AI Primer
UN OICT Emerging Technologies Team 2
 Version 1.0.1

WHAT ADVOCATES SAY WHAT SKEPTICS SAY WHY THE UN SHOULD CARE

PERSONALIZATION

Generative AI systems could reduce Since generative AI models are trained on As a multilingual organization with numerous
the barriers to entry for non-experts a vast amount of data, it can be difficult to specialized entities, one of the key advantages of
to engage in highly specialized explain how a system arrived at a particular generative AI for the UN is its ability to be trained on
work that was previously limited result or response. This lack of explainability specialized vocabulary and terminology. Moreover,
to people with relevant technical poses huge risks and could lead to unintended these models can be integrated into existing
skills and knowledge. For example, consequences. For instance, AI systems interfaces, which can help boost rapid information
by automating data analysis or can reinforce patterns of systemic bias discovery and automate knowledge management.
facilitating access to information and discrimination if their training data is
Perhaps most importantly, generative AI can be
through user-friendly interfaces. flawed, biased, or unrepresentative. Without
applied in the context of the Sustainable Development
Organizations will likely see a explainable outputs, it could be challenging,
Goals (SDGs). For example, generative AI can provide
greater return on investment or impossible, to pinpoint the source of a
marginalized communities with improved access to
by using customized generative problem and develop a mitigation strategy.
critical services such as healthcare and education.
models that are fine-tuned with A lack of transparency could also make it
Generative AI can help detect diseases early and
their own data. This can lead to difficult to trust the results of the generative
provide personalized health recommendations based
more tailored results, which would model. This is particularly worrying in cases
on an individual’s medical history, promoting health
better address the organization’s when automated decision-making may be
and well-being (SDG 3). It can also facilitate the
specific needs. Personalization can used to determine outcomes in high-stakes
provision of quality education (SDG 4) through the
also improve user experience and contexts such as healthcare, social services
creation of personalized learning experiences, making
engagement. and law enforcement.
education accessible to those who may not have
access to traditional learning opportunities.
To promote responsible use of generative AI, the
UN could take steps to ensure that personalization
features of generative AI do not result in negative
consequences—such as perpetuating biases or
exacerbating existing inequalities—especially for
marginalized populations.

CREATIVITY & INNOVATION

Generative AI tools offer several Some critics have raised concerns that Generative AI can enable marginalized communities
opportunities to help augment generative AI systems that have been trained to tap into new and innovative resources and
human creativity. For example, on copyrighted works, may infringe on the networks, which can help create jobs and generate
by prompting discussion and intellectual property rights of professional economic opportunities (SDG 8). For instance,
generating new ideas, which artists, writers and programmers. Several generative AI can help democratize access to
could spark people’s imagination lawsuits have been filed against companies education and training, which can equip people with
and possibly help them that have used copyrighted material in their the skills they need to succeed in the digital economy.
overcome writer’s block. This training data without the permission of (or
Leveraging generative AI can also help reduce
collaborative ideation tool could attribution to) its content creators.
inequalities (SDG 10). For instance, AI technologies can
assist in spurring the creation
Another common concern among critics, is be used to identify and address disparities in access
of innovative products, services,
students’ use of highly proficient generative to education and healthcare services. AI-powered
business strategies and advocacy
AI tools to complete writing assignments, platforms can also help marginalized communities
campaigns. One of the biggest
as these programs can make it harder for connect with potential employers and access training
draws of generative AI systems
professors and other academics to spot and development resources.
is their potential to change the
instances of plagiarism.
nature of how cognitive work is The UN should take proactive measures to encourage
performed. Additionally, generative AI systems cannot creativity and innovation within the bounds of ethical
always be relied on for accurate information and legal principles by implementing policies that
Educators are also considering
as they occasionally “hallucinate” or produce promote transparency, accountability and fair use
how generative AI could assist
falsehoods. This occurs when a system fails of AI-generated content, especially from creative
in creating engaging activities,
to understand its prompts and confidently industries.
including games and simulations,
produces outputs that do not match the
which could help students better
data it has been trained on. Researchers are
understand abstract or other
concerned that generative AI models could
difficult to grasp concepts.
make large-scale mis- and disinformation
cheaper and easier to produce and
disseminate. While there are some steps that
can be taken to try to reduce instances of
“hallucinations” (e.g., through reinforcement
learning with human feedback and input),
completely eliminating these falsehoods is a
persistent challenge.

 Hypothetical – the technology is conceptually possible The United Nations Office of

 Experimental – research and experiments are proving the technology Information and Communications
Maturity  Working Prototypes – working examples are being built
Technology (OICT) is developing an AI
Assessment  Diffusion – the technology is being adopted
governance framework, as well as a
matrix and guidance note, for the use
 Commercialization – the technology is part of mainstream solutions of generative AI within the UN.

Generative AI Primer
UN OICT Emerging Technologies Team 3
 Version 1.0.1

Interested in learning more About the Emerging Technologies Team

about Generative AI?
ETT expedites the adoption of frontier
In an upcoming thought piece, technologies across the UN Secretariat. It
the Emerging Technologies leverages emerging technologies to generate
Team (ETT) will further break greater efficiencies and to enhance the
down this technology’s use cases Organization’s ability to respond to an ever-
and purported transformative evolving technological landscape, while
potential for the UN. Visit the ETT providing appropriate safeguards through
web page for information about the careful identification and evaluation of
this publication. adoption-related risks.

Bibliography

» Alphonso, G. (2023) ‘Generative AI: Education In The Age Of » Lehmann, E. (2023) ‘ChatGPT can make life easier, here is for whom.’,
Innovation’, Forbes, 3 March. Available at: https://www.forbes.com/ GIZ Data Lab, 15 March. Available at: https://www.blog-datalab.com/
sites/forbestechcouncil/2023/03/03/generative-ai-education-in-the- home/generative-ai-in-international-development (Accessed 3 May
age-of-innovation/?sh=3e9a0b324eca (Accessed: 4 May 2023) 2023)
» AWS (2023) ‘What Is A Neural Network?’, Amazon Web Services, » Leswing, K. and Vanian, J. (2023) ‘ChatGPT and generative AI are
Available at: https://aws.amazon.com/what-is/neural-network/ booming, but the costs can be extraordinary’, CNBC, 13 March.
(Accessed: 27 April 2023) Available at https://www.cnbc.com/2023/03/13/chatgpt-and-
» Azhar, A. (2023) ‘What chatbots make cheap’, 21 February. Available
generative-ai-are-booming-but-at-a-very-expensive-price.html
(Accessed: 26 April 2023)
at: https://www.exponentialview.co/p/what-chatbots-make-cheap
(Accessed: 27 April 2023) » Luccioni, S. (2023) ‘The mounting human and environmental
» Briggs, J. and Kodnani, D. (2023) ‘Generative AI could raise global
costs of generative AI’, Ars Technica, 12 April. Available at: https://
arstechnica.com/gadgets/2023/04/generative-ai-is-cool-but-lets-
GDP by 7%’, Goldman Sachs Research, 5 April. Available at: https://
not-forget-its-human-and-environmental-costs/ (Accessed: 1 May
www.goldmansachs.com/intelligence/pages/generative-ai-could-
2023)
raise-global-gdp-by-7-percent.html (Accessed 1 May 2023)
» Chow, A. and Perrigo, B. (2023) ‘The AI Arms Race Is Changing
» Mok, A. and Zinkula, J. (2023) ‘ChatGPT may be coming for our jobs.
Here are the 10 roles that AI is most likely to replace.’, Business
Everything’, TIME, 17 February. Available at: https://time.
Insider, 9 April. Available at: https://www.businessinsider.com/
com/6255952/ai-impact-chatgpt-microsoft-google/ (Accessed: 2
chatgpt-jobs-at-risk-replacement-artificial-intelligence-ai-labor-
May 2023).
trends-2023-02 (Accessed: 2 May 2023)
» Fazackerley, A. (2023) ‘AI makes plagiarism harder to detect, argue
» Ribeiro Neto, J.A. (2023) ‘ChatGTP and the Generative AI
academics – in paper written by chatbot’, The Guardian, 19 March.
Hallucinations’, Medium, 15 March. Available at: https://medium.
Available at: https://www.theguardian.com/technology/2023/mar/19/
com/chatgpt-learning/chatgtp-and-the-generative-ai-
ai-makes-plagiarism-harder-to-detect-argue-academics-in-paper-
hallucinations-62feddc72369 (Accessed: 26 April)
written-by-chatbot (Accessed: 4 May 2023)
» Grant, N. and Weise, K. (2023) ‘In A.I. Race, Microsoft and Google
» Samuelson, Pamela, (2023) ‘Generative AI meets copyright’, Science,
13 July. Available at: https://www.science.org/doi/10.1126/science.
Choose Speed Over Caution’, New York Times, 7 April. Available
adi0656 (Accessed: 27 July 2023)
at: https://www.nytimes.com/2023/04/07/technology/ai-chatbots-
google-microsoft.html (Accessed: 2 May 2023) » Singer, Gadi, (2023) ‘Survival of the Fittest: Compact Generative
» Gruetzemacher, R. (2022) ‘The Power of Natural Language
AI Models Are the Future for Cost-Effective AI at Scale’, 25 July.
Available at: https://towardsdatascience.com/survival-of-the-fittest-
Processing’, Harvard Business Review, 19 April. Available at: https://
compact-generative-ai-models-are-the-future-for-cost-effective-ai-
hbr.org/2022/04/the-power-of-natural-language-processing
at-scale-6bbdc138f618 (Accessed: 27 July 2023)
(Accessed: 27 April 2023)
» Gurdeniz, E. and Hosanagar, K. (2023) ‘Generative AI Won’t
» Storius Magazine, (2023) ‘Generative AI: The End of Human
Creativity or the New Renaissance?’, Storius Magazine, 14 April.
Revolutionize Search — Yet’, Harvard Business Review, 23
Available at: https://storiusmag.com/generative-ai-the-end-of-
February. Available at: https://hbr.org/2023/02/generative-ai-wont-
human-creativity-or-the-new-renaissance-754b627efdde (Accessed:
revolutionize-search-yet (Accessed 3 May 2023)
2 May 2023)
» Hsu, T. and Thompson, S. (2023) ‘Disinformation Researchers Raise
» Wiggers, K. (2023) ‘The current legal cases against generative AI are
Alarms About A.I. Chatbots’, New York Times, 8 February. Available
just the beginning’, TechCrunch, 27 January. Available at: https://
at: https://www.nytimes.com/2023/02/08/technology/ai-chatbots-
techcrunch.com/2023/01/27/the-current-legal-cases-against-
disinformation.html (Accessed 2 May 2023)
generative-ai-are-just-the-beginning/ (Accessed: 4 May 2023)
» Kak, A. and Myers West, S. (2023) ‘AI Now 2023 Landscape:
Confronting Tech Power’, AI Now Institute, 11 April. Available at:
https://ainowinstitute.org/2023-landscape (Accessed 26 April 2023)

Generative AI Primer
UN OICT Emerging Technologies Team 4