Questions for the Record

Responding to Senator Kennedy

1. What are the most important factors for Congress to consider when crafting legislation to regulate artificial intelligence?

Any new laws related to AI will become part of a complex legal and policy landscape. A wide range of existing laws already apply to AI, including to our products. And in sectors like medicine, education, and employment, policy stakeholders have already begun to adapt existing laws to take account of the ways that AI impacts those fields. We look forward to contributing to the development of a balanced approach that addresses the risks from AI while also enabling Americans and people around the world to benefit from this technology.

We strongly support efforts to harmonize the emergent accountability expectations for AI, including the efforts of the NIST AI Risk Management Framework, the U.S.-E.U. Trade and Technology Council, and a range of other global initiatives. While these efforts continue to progress, and even before new laws are fully implemented, we see a role for ourselves and other companies to make voluntary commitments on issues such as pre-deployment testing, content provenance, and trust and safety.

We are already doing significant work on responsible and safe approaches to developing and deploying our models, including through red-teaming and quantitative evaluation of potentially dangerous model capabilities and risks. We report on these efforts primarily through a published document that we currently call a System Card. We are refining these approaches in tandem with the broader public policy discussion.

For future generations of the most highly capable foundation models, which are likely to prove more capable than models that have been previously shown to be safe, we support the development of registration, disclosure, and licensing requirements. Such disclosure could help provide policymakers with the necessary visibility to design effective regulatory solutions, and get ahead of trends at the frontier of AI progress. To be beneficial and not create new risks, it is crucial that any such regimes prioritize the security of the information disclosed. Licensure is common in safety-critical and other high-risk contexts, such as air travel, power generation, drug manufacturing, and banking. Licensees could be required to perform pre-deployment risk assessments and adopt state-of-the-art security and deployment safeguards.

There remain many open questions in the design of registration and licensing mechanisms for achieving accountability at the frontier of AI development, such as precisely how to define such models, and how to ensure that smaller firms are not unduly burdened with unnecessary regulation. We look forward to collaborating with policymakers in addressing these questions.

2. Can you provide a list of the dangerous capabilities evaluations you use when testing a model?

When we released GPT-4, we published a document that we call a System Card, which described our approach to assessing the potential for dangerous capabilities in GPT-4, gave specific examples of capabilities we investigated, described what we found, and explained what we did to address the identified risks. I am enclosing a copy of the GPT-4 System Card for inclusion in the Senate record. You will find a concise list of safety challenges at page 4 of the System Card, details on our evaluations at pages 4-20, and details on mitigations implemented at pages 21-25.

The development of safety best practices, including standardized evaluations for dangerous capabilities, is an active focus of technical research and engineering effort not only at OpenAI but also at other frontier labs such as Anthropic, Google Deepmind, and Microsoft. We anticipate that these technical efforts will evolve in tandem with public policy work to define meaningful and effective yardsticks for measuring dangerous capabilities and mitigating risks.

3. Can you provide a specific list of tests that a model has to pass prior to being deployed into the world?

We are constantly refining our process for deciding which models to deploy. The description below is a snapshot of our current approach. We intend to share more about our approach publicly in the future, and would be pleased to discuss further with you or your staff at any time.

We use a wide range of evaluations for both the capability and safety dimensions of our models. These include not only quantitative evaluations but also qualitative assessments such as red teaming and engagement with external experts in specific risk areas.

Depending on the novelty and level of capability of a particular model, we apply different evaluation processes. For example, any completely new model will undergo extensive testing; small tweaks to existing models are subject to a subset of these tests.

We currently distinguish between two different types of review for deployment decisions. Some models undergo heightened and specialized review by a safety board that performs (or instructs employees to perform) further technical and social impact analyses. These are generally models likely to have significantly greater and more novel capabilities than our previously deployed models, often due to significantly superior compute resources or technical parameters employed in training them. Models with technical features comparable to already deployed models are not routed through this review board, though they may still receive a range of evaluations, as described below and in the attached System Card.

OpenAI will sometimes set specific targets for what constitutes sufficiently safe performance on a particular quantitative metric, such as refusing a certain percentage of outputs in a particular risk category, or achieving a sufficient degree of parity in performance across different types of inputs. At other times we use more holistic and qualitative criteria such as the extent to which our understanding of risk (via red teaming, expert engagement, etc.) has “converged” to a reasonably stable point that we believe to be sufficiently safe and our mitigations sufficient to address the identified risks.

Over time, in line with our iterative deployment approach, lessons from early access (e.g. researcher access, red teaming, alpha testing) and larger scale access feed back into improvement of mitigations for those same deployed models, as well as institutional learning for improving the deployment process for future models.

Responding to Senator Tillis

1. What do you see as the principal threats, posed by generative AI, to creators of copyrighted works? How do you plan to address these threats?

We have spoken with and heard from many creators who hold a broad spectrum of opinions about the opportunities as well as the potential risks of AI. Many creators are excited about the opportunities that AI brings to simplify and streamline their work, and to grant them new capabilities in their fields. AI systems can help create initial drafts for consideration, help do preliminary research for written works, or allow creators to quickly iterate on ideas and sketches. For example, DALL-E 2 helped a group of designers and artists create a new cover for an issue of Cosmopolitan Magazine, through an iterative, human-driven process.¹

We are also aware that many creators are concerned about being displaced or replaced by AI. We’ve spoken with prominent artists, musicians, writers, actors, filmmakers, and other groups about these issues. Creators are concerned that improperly developed AI systems may directly copy or create derivatives of their work. Many creators also wonder what it will mean to be a human creator in a world where AI can “create” things at significantly greater scale and speed. Some creators are also concerned that the value created by AI systems will accrue to large companies (such as technology platforms or industry rights groups), diminishing the value and power of individual creativity.

Other creators have provided prominent counterpoints to these concerns. They have noted that the role of a human creator is to be able to see and bring to life something that no one else can see, and AI will not replace that human spark, nor will AI be able to “see” such things in the first place. They have also observed that AI is simply the latest in a long line of technological tools that have assisted and empowered creators; the newest generation of creators is already cutting its teeth on these tools, and they will quickly become ubiquitous and essential to creative workflows as many technological tools, such as CGI or auto-tune, are today.

OpenAI does not want to replace creators. We want our systems to be used to empower creativity, and to support and augment the essential humanity of artists and creators. We design and continue to refine our systems to support these goals.

We continue to engage with a broad variety of creators and creative groups about these issues, and are committed to taking part in industry and policy discussion on these topics and to updating our approaches and thinking based on this feedback.

2. For the purposes of determining the presence of copyrighted works in training data and models, to what extent has OpenAI kept records regarding what has been ingested by each iteration of ChatGPT and Jukebox?

The training data for ChatGPT is based on publicly accessible content, licensed content, and data from human trainers. OpenAI keeps records related to these development activities, and has previously disclosed that the vast majority of the publicly accessible content originates from public sources such as Common Crawl and Wikipedia.

The training data for Jukebox was discussed in a paper published about Jukebox at the time of its release in 2020.² This model is a non-commercial research project that was designed to advance scientific progress and understanding of music, and the license attached to the project further stipulates that it may only be used for non-commercial purposes. The open source license is viewable at the Github repository for this project.³ OpenAI engaged in a dialogue with a leading music industry organization about Jukebox around the time of its release, and has not undertaken any significant work on this model since its release. While there may be many interesting commercial applications in generative AI music that simultaneously respect the work of creators and advance the arts and sciences, it is not a current focus area for OpenAI’s activities.

3. Moving forward, are there plans to give creators control over inclusion of their works in training data and how their works are used by a given model? What about works that have already been used in existing training data and models without any creator control?

As noted above and in numerous public statements, OpenAI is committed to engaging with and respecting creators and the creative process. OpenAI’s training data comes from a variety of sources, including from publicly available resources like Common Crawl and also by licensing content from content owners.

One method of enabling control is by providing a path for creators to elect not to have a work used to help teach AI models. While we do not believe that as a matter of public policy that anyone’s ability to learn from a work should be restricted, we understand concerns from creators about their works being used to help teach AI models and the fear of displacement or replacement. We continue to have cooperative and productive discussions with creators about the use of their content in mutually beneficial and agreeable ways. Resources like Common Crawl respect industry standard means for content owners to disallow use of their content in that resource, so many content owners already have the ability to restrict their works from being used to teach AI models. We are also working on a number of additional ideas in this space. We continue to engage with a broad variety of creators and creative groups about these issues, and are committed to taking part in both industry and public policy discussion on these topics in the future and to updating our approaches and thinking based on this feedback.

Responding to Senator Blumenthal

1. Training data is crucial to foundational models like GPT-4, where content such as news, art, music, and research papers are used to create and refine AI systems, largely material aggregated from the internet. This content represents the labor, livelihoods, and careers of artists, experts, journalists, and scientists. How should we make sure AI systems respect, acknowledge, and compensate the labor of individuals whose work is used to train AI models?

Ensuring that the creator economy continues to be vibrant is an important priority for OpenAI. Writers, artists, composers and other creators have contributed immeasurably to societies throughout the history of civilization, and they are a vital part of American society and the American economy today. OpenAI is actively engaged in discussions with a wide variety of creators and content owners, geared toward finding mutually beneficial opportunities for creators and technology providers. Those discussions include a recognition by all parties that the technology is still in a nascent stage, and many creators continue to experiment with AI tools to assist in their creation of new works. A few examples:

Karen Cheng, an artist who uses OpenAI’s image generation tool to prompt the AI system to generate creative imagery overlaid to the rhythm of music in the background, created this DALL-E “music video⁶

Tim Boucher, a science fiction writer, has used a combination of AI tools to write a series of books in a volume driven format that previously would not have been possible⁷

Paul McCartney is using AI to create a final Beatles album⁸

AI can also be used to make creative works more accessible and available to consumers. For example, OpenAI works with Be My Eyes to provide AI assisted visual description tools that enable visually impaired persons to view and appreciate art they would not have been able to see before.⁹

In addition, OpenAI provides its DALL-E tool in conjunction with an artist financial assistance program, which enables artists to apply for subsidized access to these technologies to ensure they remain broadly available regardless of economic means.¹⁰

As for enabling artist control, OpenAI’s training data comes from a variety of sources, including from publicly available resources like Common Crawl and also by licensing content directly from content owners. Resources like Common Crawl respect industry standard means for content owners to disallow use of their content in that resource, so many content owners have the ability to restrict their works from being used to teach AI models. We are working on a number of ideas in this space.

In addition, OpenAI has had cooperative and productive discussions with creators and creative platforms both about the use of works that could be used to teach AI models, as well as the best ways for such works to be found, showcased, and inspire others in an AI enabled world. OpenAI has also entered into license agreements to pay for specialized content, such as its partnership with Shutterstock, and it expects to continue to do so in the future.

It is also important to build shared understanding about how training data is actually used in development of our models. Some have suggested that generative AI outputs are simply reproductions and rearrangements of pieces of creative expression received via inputs. Others have even claimed that outputs are akin to “collages” or “remixes.”

This is not the case. Rather, training data is used to teach models statistical patterns at massive scale. For example, training data for large language models is not used to teach these systems to copy the data, but to learn the fundamentals of language – including vocabulary, grammar, sentence structure, and even basic logic. Ultimately, these AI systems are not search engines or databases, and are not designed to repeat or even store the content on which they are trained. In some ways, the process is akin to a person participating in a foreign language immersion environment and developing an understanding of how to speak, write and think in that new language. The fluency that results is not due to memorizing and parroting specific parts of the thousands of works that may be read or seen or heard. Rather, it is the result of learning and mastering the fundamentals of the new language and applying it, independently, to new situations and ideas. This is the fundamental relationship between AI training data and its outputs.

OpenAI does not want to replace creators. We want our systems to be used to empower creativity, and to support and augment the essential humanity of artists and creators. We are continuously designing and refining our systems. We look forward to continuing our important dialogues with content creators, and with policymakers, to help identify the right frameworks that continue to advance scientific progress while supporting the flourishing of beneficial and valuable creative works.