Generative Artificial Intelligence: Overview, Issues, and Questions for Congress
June 9, 2023
Generative Artificial Intelligence: Overview, Issues, and
Questions for Congress
Generative artificial intelligence (GenAI) refers to AI
use mathematical techniques called attention or self-
systems, in particular those using machine learning (ML)
attention to detect how data elements, even when far away
and trained on large volumes of data, that are able to
sequentially, influence and depend on each other. These
generate new content. In contrast, other AI systems may
methods make GPT models faster to train, more efficient in
have a primary goal of classifying data, such as facial
understanding context, and highly scalable.
recognition image data, or making decisions, such as those
used in autonomous vehicles. GenAI systems, when
Other critical components to the recent GenAI advances are
prompted (often by a user inputting text), can create various
the availability of large amounts of data and the size of their
outputs, including text responses (e.g., OpenAI’s ChatGPT
language models. Large language models (LLMs) are AI
and Google’s Bard), images (e.g., Stability AI’s Stable
systems that aim to model language, sometimes using
Diffusion and Midjourney’s self-titled program), videos,
millions or billions of parameters (i.e., numbers in the
computer code, or music.
model that determine how inputs are converted to outputs).
Repeatedly tweaking these parameters, using mathematical
The recent public release of many GenAI tools, and the race
optimization techniques, and large amounts of data and
by companies to develop ever-more powerful models, have
computational power, increases model performance.
generated widespread discussion of their capabilities,
potential concerns with their use, and debates about their
Notably, GenAI models work to match the style and
governance and regulation. This CRS InFocus describes the
appearance of the underlying training data. They also have
development and uses of GenAI, concerns raised by the use
been shown to have capability overhang, meaning hidden
of GenAI tools, and considerations for Congress. For
capabilities that their developers and users did not
additional considerations related to data privacy, see CRS
anticipate but that are emerging as the models grow larger.
Report R47569, Generative Artificial Intelligence and Data
Privacy: A Primer, by Kristen E. Busch.
Beneficial Uses and Concerns
Background
The increase in size of recent GenAI systems (with
hundreds of billions of parameters) has led to drastically
AI can generally be thought of as computerized systems
improved capabilities over previous systems (with millions
that work and react in ways commonly considered to
or a few billion parameters). For example, as released in
require human intelligence, such as learning, solving
2020, OpenAI’s GPT-3 could translate sentences from
problems, and achieving goals under uncertain and varying
English to French with few to no training examples and
conditions, with varying levels of autonomy. AI is not one
outperformed previous models that were explicitly trained
thing; AI systems can encompass a range of technologies,
to solve that task. GPT-4, released in 2023, improved on
methodologies, and application areas, such as natural
various benchmarks, such as scores on simulated graduate
language processing, robotics, and facial recognition.
and professional exams, and on traditional ML benchmarks.
The AI technologies underpinning many GenAI tools are
OpenAI’s ChatGPT, reportedly the fastest growing
the result of decades of research. For example, recurrent
consumer application in history, uses GPT-3.5 or, for paid
neural networks (RNNs, a type of ML loosely modeled
subscribers, GPT-4, which is the latest, most powerful
after the human brain that detect patterns in sequential data)
version. ChatGPT is a conversational AI chatbot that can
underwent a period of much development and improvement
generate a variety of text outputs, such as emails, essays,
in the 1980s-1990s. RNNs can generate text, but they are
and casual conversations. Microsoft has incorporated GPT-
limited in retaining contextual information across large
4 into its Bing search engine as a chatbot called Bing Chat,
strings of words, are slow to train, and are not easily scaled
and numerous other chatbots also use the GPT-3 and GPT-4
up by increasing computational power or training data size.
models. In addition to GPT models, other LLM chatbots
include Google’s Pathways Language Model 2 (PaLM 2) in
More recent technical advances—notably the introduction
the Bard chatbot, and Hugging Face’s BLOOM model in
of the Transformer architecture by Google researchers in
SambaNova’s BLOOMChat.
2017 and improvements in Generative Pre-Trained
Transformer (GPT) models since around 2019—have
Despite the impressive abilities of GenAI, its growing
contributed to dramatic improvement in GenAI
prominence is raising concerns. For example, the tendency
performance. Transformer models process a sequence of
of GenAI to make things up, sometimes referred to as
whole sentences (rather than analyzing word by word),
hallucinating, might result in the tools generating and
which helps them to “remember” past information. They
amplifying misinformation or being used to create and
https://crsreports.congress.gov
Generative Artificial Intelligence: Overview, Issues, and Questions for Congress
spread disinformation. OpenAI notes that even a powerful
with office tasks such as creating and summarizing content,
model like GPT-4 “is not fully reliable” and “great care
writing speeches, and drafting bills. At the same time,
should be taken when using language model outputs,
Congress has begun considering whether and how to
particularly in high-stakes contexts, with the exact protocol
implement guardrails for GenAI technologies. As that work
(such as human review, grounding with additional context,
continues, along with further development and growing use
or avoiding high-stakes uses altogether) matching the needs
of GenAI tools, Congress might consider a range of
of a specific use-case.” Additionally, because the models
questions and potential actions.
are generally trained on large amounts of data scraped from
the internet, they can incorporate, reflect, and potentially
• Bias and ethics of use. How are the private and public
amplify biases in such data.
sectors using federal guidance documents and
frameworks for AI evaluation and risk management to
Particular GenAI use cases, such as in education and
address bias and manage risk in GenAI systems?
scientific research, have also raised questions about ethical
Should the deployment of GenAI models in high-risk
and transparent use, whether restrictions should be
scenarios (e.g., mental health therapy or generating
implemented and whether they are effective, and the
forensic sketches) be restricted?
accuracy of detection tools.
• Testing and transparency. The biggest models
As GenAI use grows, analysts have begun considering how
deployed today, such as GPT-4 and PaLM 2, are closed
it might affect jobs and productivity. For example, will
source, proprietary models. While many companies
these tools complement workers’ skills in existing jobs and
state that they conduct internal testing and are
create new jobs? Will GenAI automate some jobs,
evaluating options for external validation and testing,
displacing workers? While these have been long-standing
Congress might consider whether and how to support
concerns for automation and AI technologies, the speed,
or require independent testing and reporting of results.
capability, and widespread use of the latest GenAI models
have heightened them.
• Economic and workforce impacts. Researchers in
industry, the private sector, and academia have begun
LLMs have been characterized as foundation models (also
analyzing GenAI’s potential widespread effects on
called general-purpose AI), meaning models trained on
labor. The National Academy of Sciences is currently
broad data that can be adapted to a wide range of
updating a study on automation and the U.S.
downstream tasks. (In contrast, many other AI systems are
workforce. In addition to assessing those findings,
built, trained, and used for particular purposes.) As
Congress might consider the appropriate federal role in
described by the Stanford University Institute for Human-
supporting U.S. workforce reskilling or upskilling in
Centered AI, foundation models may be built upon or
response to shifting job tasks caused by the
integrated into multiple AI systems across a variety of
implementation of GenAI. It might also consider
domains, with the potential for both benefits (e.g.,
whether and how to increase AI expertise in the
concentrating efforts to reduce bias and improve
government’s own workforce.
robustness) and drawbacks (e.g., security failures or
inequities that flow out to downstream applications,
• Research and competition. Estimates put training
amplifying their harms).
costs for GenAI models like GPT-3, with 175 billion
parameters, at over $4.6 million. Some analysts have
Federal AI Laws and GenAI Legislation
argued that cost, use of proprietary data, and access to
vast computing power will create a divide between
Numerous bills focused on AI, or including AI-focused
those who can train the most cutting-edge LLMs (e.g.,
provisions, have been enacted in prior Congresses. For
large technology firms) and those who cannot (e.g.,
example, the National Artificial Intelligence Initiative Act
nonprofits, startups, universities). Congress might
of 2020 (Division E of P.L. 116-283) codified the
consider ways to support access to data, training, and
establishment of a national AI initiative and associated
computing resources, such as through codifying
federal offices and committees. In the 118th Congress, at
least 50 AI-related bills have been introduced.
recommendations in the final report of the National AI
Research Resource Task Force.
Specifically regarding GenAI, the Identifying Outputs of
• Oversight and regulation. How might Congress
Generative Adversarial Networks (IOGAN) Act (P.L. 116-
regulate GenAI technologies while supporting
258) directed federal support of research on generative
innovation and international competitiveness? Do
adversarial networks. In the 118th Congress, at least four
federal regulatory agencies have the authorities and
bills have been introduced that specifically include GenAI,
resources to adequately oversee and regulate GenAI
including those pertaining to transparency and
tools to minimize risks while supporting benefits? If
accountability of GenAI use in political advertisements,
not, what additional authorities are needed? How might
disclosure of GenAI outputs, and federal oversight of
digital platforms.
federal oversight and regulation for GenAI be distinct
from that for AI technologies more broadly?
Potential Questions for Congress
Laurie A. Harris, Analyst in Science and Technology
Like the private sector, federal agencies and Congress have
Policy
begun testing GenAI uses, including for its potential to help
https://crsreports.congress.gov
Generative Artificial Intelligence: Overview, Issues, and Questions for Congress
IF12426
Disclaimer
This document was prepared by the Congressional Research Service (CRS). CRS serves as nonpartisan shared staff to
congressional committees and Members of Congress. It operates solely at the behest of and under the direction of Congress.
Information in a CRS Report should not be relied upon for purposes other than public understanding of information that has
been provided by CRS to Members of Congress in connection with CRS’s institutional role. CRS Reports, as a work of the
United States Government, are not subject to copyright protection in the United States. Any CRS Report may be
reproduced and distributed in its entirety without permission from CRS. However, as a CRS Report may include
copyrighted images or material from a third party, you may need to obtain the permission of the copyright holder if you
wish to copy or otherwise use copyrighted material.
https://crsreports.congress.gov | IF12426 · VERSION 1 · NEW