INTRO TO

GENERATIVE A.I.
How to use this guide
This primer addresses the common questions that many people have
about generative artificial intelligence (AI) systems. It contains information
about generative AI capabilities, seven key issue areas where generative AI
is having impact, and how generative AI may be used in the future.
Although this primer was originally designed primarily for journalists, we
have found that it is a valuable resource for anyone communicating about
generative AI.

For a list of resources for where to find technical experts, please see
Finding Experts in AI. For information about AI more generally, please see
AI 101.

Visit techprimers.aspendigital.org to see all primers.

Intro to Generative AI © 2023 by Aspen Digital is licensed under CC BY-NC 4.0.

 CONTENTS
3 What is generative AI?
What it is and how it’s different from decision-making AI tools

3 What generative AI capabilities exist today?

Text-to-text models like ChatGPT aren’t the only thing out there

4 Poster: Tech that people call “artificial intelligence”

5 How might generative AI be used?

It will disrupt many aspects of society

6 Key Issues in Generative AI

7 Information ecosystems Generative AI makes it much easier to mislead at scale

8 Expectations & claims Tools like ChatGPT are not designed to produce facts

9 Intellectual property Existing tools do not attribute to their source material

10 Future of work Automating creative work will impact peoples’ livelihoods

11 Environmental impacts AI use is far from carbon-neutral

12 Discriminatory effects Generative AI systems can perpetuate societal harms

13 Feedback loops How will generative AI impact future AI development?

14 What comes next?

15 Acknowledgements

2
WHAT IS GENERATIVE AI?
Generative AI is a subset of artificial intelligence technologies that are
used to create new content, such as images or text, based on patterns in
large amounts of existing content. Generative AI differs from classification
AI—like email spam filtering or tumor detection used in medical
settings—because generative AI systems are designed to make content,
not to make decisions.

WHAT GENERATIVE AI CAPABILITIES

EXIST TODAY?
While ChatGPT captured the public’s attention by allowing people to
generate uncanny and seemingly confident responses to a vast array of
written prompts, there is a diverse set of generative AI applications that
have been made available to businesses and consumers. Increasingly, AI
developers are creating multimodal tools, tools which incorporate
multiple sources of input data (such as images, audio, or text) at once.

Image-to-image Canvas

Text-to-image Midjourney, DALL-E, Stable Diffusion

Text-to-audio MusicLM

Video-to-video Project Morpheus, AI video compression

Text-to-video Make-A-Video

Image-to-text Automatic image description

Text-to-text (including computer code) Github Copilot, Bard

3
 TECH THAT PEOPLE CALL
“ARTIFICIAL INTELLIGENCE”
MACHINE LEARNING AUTOMATED REASONING
Statistical (based on data) Symbolic (based on rules)

Note: Machine learning

is currently the most
Deep Blue recommended chess moves
by using the rules of the game and EXPERT
popular approach,
searching a parallelized tree of many
possible combinations of moves.
SYSTEMS
but some believe that
automated reasoning
may make a comeback
or that these two The various ghosts in Pac-Man are
branches may be
combined in new ways.
programmed to chase you using
different pathfinding methods. (Blinky, PLANNING
the red ghost, is most aggressive!)

LABELED CLASSIFICATION FaceID classifies images of faces using

convolutional neural networks (CNNs).
DATASETS
DALL-E produces images from text through
an iterative process using a diffusion model.

MIXTURE OF GENERATIVE
LABELED & ChatGPT generates text in response to
typed prompts using a transformer deep
UNLABELED learning model.

DATASETS REINFORCEMENT
AlphaGo develops strategies for the game
Go using self-play reinforcement learning.

UNLABELED CLUSTERING Spotify clusters songs into music genres

using hierarchical density-based spatial
DATASETS clustering with noise (HDBSCAN).

This work was made possible thanks to generous support from

Siegel Family Endowment, the Patrick J. McGovern Foundation,
and the John S. and James L. Knight Foundation..
HOW MIGHT GENERATIVE AI BE USED?
Although generative AI tools are still in the early stages of development,
they are already being used to produce content at surprisingly high
speeds, low costs, and with relative ease for end users. This newfound
accessibility has labor and operational implications for software
engineering, media production, education, the commercial art market,
and more. No one knows exactly what will emerge from the explosion of
generative AI tools hitting the market, but early experiments point
toward the potential for larger scale disruption to business, security, and
society at large:

Hyper-personalized Content
Traditionally, the cost of making individually personalized content,
such as ads with your face in them or movie trailers narrated in the
sound of a loved one’s voice, was prohibitively high. People may
now easily use generative AI to realize this level of extreme
personalization, either for their own fulfillment or to manipulate
others.

The Rise of “No-Code” Application Development

Historically, in order to develop websites or computer applications,
you needed to know programming languages. Now, it is becoming
possible to use conversational language to prompt an AI tool to
produce computer code for you (even if today’s systems are still
imperfect). These tools may lower costs by expanding the number
of people who are able to create and contribute to software
development and make a wide range of products and services
more accessible. However, they could also negatively impact how
much people are paid for these skills and change the nature of
work to make it more tedious and less collaborative.

5
HOW MIGHT GENERATIVE AI BE USED?
Better Augmented Reality
Real-time rendering of believable digital environments is
computationally intensive and expensive. These graphical
requirements have been a pernicious issue for augmented or virtual
reality applications because time delays in rendering can create a
jarring and unnatural user experience. Generative AI systems could
be used to approximate (if not perfectly replicate) complex physical
phenomena, like lighting and shadows, making these virtual scenes
feel more immersive.

There are still many unknowns and opportunities for discovery. We have
only scratched the surface on possible uses of these tools.

KEY ISSUES IN GENERATIVE AI

There are a number of promising applications of generative AI systems, a
subset of artificial intelligence technologies that are used to create new
content based on patterns in large amounts of existing content. These
uses are not without their risks, however. The following sections highlight
a number of the most pressing issues associated with generative AI, with
links to a number of illustrative articles exploring perspectives on each of
these issues.

7 Information ecosystems 11 Environmental impacts

8 Expectations & claims 12 Discriminatory effects
9 Intellectual property 13 Feedback loops
10 Future of work

6
KEY ISSUES IN GENERATIVE AI

INFORMATION ECOSYSTEMS
How will generated content affect the trustworthiness of media?

Media created to mislead is not a new problem, but generative AI makes

it much easier to create mis- and disinformation at scale and to create
convincing human-like AI interactions that could be used to exploit users
with scams or security attacks. As generated content improves, it will be
harder to detect inauthentic content in the wild, and it will be easier for
smaller actors to manufacture large-scale disinformation campaigns.

Deepfakes: Humans may find deepfake faces more trustworthy than

real ones
scientificamerican.com/article/humans-find-ai-generated-faces-more-trustworthy-than-th
e-real-thing

Confident nonsense: “While the answers which ChatGPT produces

have a high rate of being incorrect, they typically look like they might
be good,” burdening human moderators
meta.stackoverflow.com/questions/421831/temporary-policy-chatgpt-is-banned

Democratizing malware: ChatGPT makes it possible for anyone to

produce malware and phishing emails
arstechnica.com/information-technology/2023/01/chatgpt-is-enabling-script-kiddies-to-
write-functional-malware

Impact on elections: Generative AI is already being used to

influence voters and reduce trust in the electoral system
www.brennancenter.org/our-work/analysis-opinion/how-ai-puts-elections-risk-and-neede
d-safeguards

7
KEY ISSUES IN GENERATIVE AI

EXPECTATIONS & CLAIMS

What are the limitations of generative AI systems?

People selling AI products and services benefit from systems being

perceived to be more reliable and capable than they are, from the
anthropomorphization of “smart assistants” to the typing animations of
text-generators like ChatGPT. Peeking behind the curtain reveals that the
AI tools on the market are specialized in scope, not general-purpose
“intelligences,” and still have crucial vulnerabilities and flaws. Although it
might be tempting to ascribe greater power to these systems, there are
still many questions about whether they are appropriately effective for
the widespread adoption we are already seeing, let alone that we are on
the verge of “Artificial General Intelligence” that surpasses humans in a
broad range of capabilities. Systems like ChatGPT and Bard were
designed to produce confident sounding text, not factual statements.

AI hype: An AI model “passing” an exam designed for humans is not

indicative of intelligence
technologyreview.com/2023/08/30/1078670/large-language-models-arent-people-lets-s
top-testing-them-like-they-were

Confident but inaccurate: CNET used ChatGPT to generate articles

that ended up riddled with errors
washingtonpost.com/media/2023/01/17/cnet-ai-articles-journalism-corrections

False equivalency: Equating GPT’s text prediction capabilities with

consciousness is misleading
karawynn.substack.com/p/language-is-a-poor-heuristic-for

Unintuitive failures: Certain strange prompts (like usernames) elicit

nonsense from ChatGPT
vice.com/en/article/epzyva/ai-chatgpt-tokens-words-break-reddit

8
KEY ISSUES IN GENERATIVE AI

INTELLECTUAL PROPERTY
Who owns what?

Many datasets used to build today’s generative AI have been compiled

by scraping, or extracting information, from the web. For example, to
make a generative AI tool that can output digital images, developers
scraped millions of existing images hosted on large art platforms, like
Flickr and DeviantArt. Content collected online in this manner is often
used without the consent or knowledge of the original creator. Even if the
original content is not reproduced by the system (although in some cases
it can be), this process leads to thorny questions around attribution,
intellectual property, monetization of generative AI tools, and economic
harms to creative industries.

Sensitive data: Scraping isn’t flawless—personal health data was

found in a popular image dataset
arstechnica.com/information-technology/2022/09/artist-finds-private-medical-record-ph
otos-in-popular-ai-training-data-set

IP protections at work: Google won’t release its text-to-music

generator because there is a chance it will reproduce copyrighted
music it was trained on
techcrunch.com/2023/01/27/google-created-an-ai-that-can-generate-music-from-text-de
scriptions-but-wont-release-it

Fights for attribution: A class action lawsuit is filed against Microsoft

for lacking attribution for code used to train Github Copilot
theverge.com/2022/11/8/23446821/microsoft-openai-github-copilot-class-action-lawsuit
-ai-copyright-violation-training-data

Copyright ramifications: AI-generated works should not be

permitted copyright protection
lawfaremedia.org/article/ai-generated-works-artists-and-intellectual-property

9
KEY ISSUES IN GENERATIVE AI

FUTURE OF WORK
How will generative algorithms impact peoples’ livelihoods?

Generative tools can be used as assistants, augmenting human creativity,

but they can also be used to automate certain types of work, from writing
copy to creating spot art for articles to coding. There are many open
questions about what tasks will be most easily automated and whether
that automation will result in a reduction in total jobs, a profound change
in how certain work is valued, or a restructuring of labor as new jobs are
created. For example, a software developer that once created website
templates could either (1) lose their job because someone else can use a
tool to do themselves what they would have hired the developer to do,
(2) get a reduction in salary as they face more competition in the market
or, (3) no longer code as much manually, but instead be in charge of
generating outputs using the AI.

Expanded creativity: Generative AI could make it much easier for

people to create websites and apps without needing to know how to
code
venturebeat.com/ai/how-code-generating-ai-code-llms-is-creating-new-challenges-as-it-
matures

Impact on creative industries: The use of generative AI in

Hollywood was a key sticking point in labor negotiations
wsj.com/business/media/hollywood-studios-can-train-ai-models-on-writers-work-under-t
entative-deal-aedae589

Augmenting work: How professionals can use ChatGPT today

linkedin.com/pulse/chatgpt-professionals-guide-using-ai-allie-k-miller

Invisible labor: Automating some tasks just creates a different kind

of work
logicmag.io/failure/the-automation-charade

10
KEY ISSUES IN GENERATIVE AI

ENVIRONMENTAL IMPACTS
How does generative AI contribute to climate change and
consume natural resources?

Generative AI systems are more resource-intensive than many similar

technologies. The large data centers (collections of connected
computers) used to develop and deploy these tools require power and
cooling, using electricity and water. While the development of AI models
typically requires a large amount of energy, generative AI is unique in that
people’s ongoing use of these systems makes up the bulk of its
consumption. Additionally, although today the water usage of generative
AI is dwarfed by the amount used for other purposes, like growing
almonds or making potato chips, water and energy usage would rise
further if proposals such as incorporating generative AI into every Google
search are implemented.

Carbon footprint: Generative AI use eats up resources needed to

mitigate the climate crisis
slate.com/technology/2023/08/chatgpt-ai-arms-race-sustainability.html

Energy usage versus other technologies: ChatGPT uses five to nine

times as much electricity as a Google search
towardsdatascience.com/chatgpts-energy-use-per-query-9383b8654487

Freshwater usage: Generative AI systems contribute to stress on

local water infrastructure in drought-prone areas
themarkup.org/hello-world/2023/04/15/the-secret-water-footprint-of-ai-technology

Reducing climate impacts: Some argue that companies can make AI

greener by fine-tuning existing models instead of training new ones
hbr.org/2023/07/how-to-make-generative-ai-greener

11
KEY ISSUES IN GENERATIVE AI

DISCIMINATORY EFFECTS
How do generative systems perpetuate societal harms?

Unless the data ingested into AI models is carefully curated—which

datasets scraped from the web rarely are—tools built using that dataset
will reflect the biases of the unfiltered internet. Even with careful dataset
curation, however, AI tools need to be fine-tuned by human content
moderators to mitigate systemic biases. In some cases, creators or
deployers of a system will manually override the AI system to limit output
of harmful material, but these sorts of interventions are necessarily brittle
and imperfect.

Biased assumptions: The "magic avatars" created using Lensa AI

sexualize women and whiten people of color regardless of their
users’ wishes
polygon.com/23513386/ai-art-lensa-magic-avatars-artificial-intelligence-explained-stable
-diffusion

Mitigating harms: GPT-3 will make biased statements against certain

groups, but it may be possible to mitigate this with extra training
focused on fairness
vox.com/future-perfect/22916602/ai-bias-fairness-tradeoffs-artificial-intelligence

Content moderation: In protecting the world from biased and

discriminatory outputs in these systems, content moderators suffer
the consequences
time.com/6247678/openai-chatgpt-kenya-workers

12
KEY ISSUES IN GENERATIVE AI

FEEDBACK LOOPS
How will generative AI impact future AI development?

Future datasets scraped from the web will be impacted as everyday

people, content farms, and disinformation campaigns saturate the
internet with generated content. New AI models that are trained using
these datasets may perpetuate existing biases documented in large
language models like GPT-3 or image generation models like Stable
Diffusion. Detecting generated content to exclude it from datasets is an
active field of research but is by no means a solved problem. This
feedback loop of using content produced by machines to train machines
to produce more content could reduce the quality and performance of
future AI systems.

13
 WHAT COMES NEXT?
While these are the early days, many experts agree that generative AI
systems will have far-reaching implications for society. Unlike blockchain
and other emerging technologies that have caught the tech industry’s
eye, generative AI tools have sparked the public’s interest and
imagination with creatives and business leaders alike identifying ready
applications. Most immediately, here are some things that could come
next:

Tech companies vying to both define and control new markets

carved out by generative AI tools, deploying work-in-progress tools
into an unregulated space

The establishment of legal frameworks and precedent to better

define both intellectual property rights and consumer protection
with regards to generative AI and the AI space more broadly

The immediate disruption of some existing labor markets while new

areas of work are still being defined

14
 ACKNOWLEDGEMENTS
This work was produced by Eleanor Tursman, B Cavello, and Tom
Latkowski, and was made possible thanks to generous support from
Siegel Family Endowment, the Patrick J. McGovern Foundation, and
the John S. and James L. Knight Foundation.

Aspen Digital is a technology and information-focused nonpartisan

organization that brings together thinkers and doers to uncover new
ideas and spark policies, processes, and procedures that empower
communities and strengthen democracy. This future-focused
Aspen Institute program inspires collaboration among diverse voices
from industry, government, and civil society to ensure our interconnected
world is accessible, safe, and inclusive—both on and offline. Across its
initiatives, Aspen Digital develops methods for elevating promising
solutions and turning thought into networked impact.

Contact us with questions or corrections regarding this primer. Please

note that Aspen Digital cannot guarantee access to experts or expert
contact information, but we are happy to serve as a resource. To find
experts, please refer to Finding Experts in AI.

Visit techprimers.aspendigital.org to see all primers.

15