Full Story | The White House convenes 7 AI companies to issue a voluntary commitment to artificial intelligence

Article source: Alpha Rabbit Research Notes

On July 21, 2023, according to the White House announcement, U.S. President Joe Biden convened seven leading artificial intelligence companies at the White House on July 21—Amazon, Anthropic, Google, Inflection, Meta, Microsoft, and OpenAI, and announced that these companies have voluntarily committed to help realize the safe, reliable, and transparent development of artificial intelligence technology. This article is from the official website of OpenAI, and the main content is consistent with the White House website**. **

full text

In order to strengthen the safety, security and credibility of artificial intelligence technology and services, OpenAI and other leading artificial intelligence laboratories have made voluntary commitments. **This effort, coordinated primarily by the White House, is an important step in advancing meaningful and effective AI governance in the United States and around the world. **

In pursuit of its mission of building safe and beneficial AI, OpenAI will continue to pilot and refine specific governance practices that are tailored for a production-capable base model like OpenAI. In addition, OpenAI will continue to invest in areas of research that help inform regulation, such as developing technologies to assess the potential dangers of artificial intelligence models.

• "Policy makers around the world are considering new laws targeting highly capable AI systems. Our commitment today contributes concrete practice to ongoing discussions. This announcement is part of OpenAI's ongoing collaboration with governments, civil society organizations, and others around the world to advance AI governance," said Anna Makanju, Vice President of Global Affairs at OpenAI. *

Voluntary AI commitments

The following list of voluntary commitments was originally released by the White House:

The following is a list of commitments made by companies to promote the safe, reliable, and transparent development and application of artificial intelligence technologies. The following commitments are consistent with existing laws and regulations, and are designed to promote the soundness of artificial intelligence legal and policy systems. Each company will fulfill the following voluntary commitments until the relevant laws, regulations and systems come into effect, and some companies may make other commitments beyond this commitment.

Scope: This commitment refers to specific models that apply only to generative AI models that are more powerful than the current industry frontier (e.g., models that are more powerful than any currently released model, including GPT-4, Claude 2, PaLM 2, Titan, and DALL-E 2 for image generation).

Note: Security corresponds to deliberate (man-made) threats, and Safety corresponds to the vulnerabilities or unexpected consequences of the AI system itself; here, AI Safety corresponds to functional safety, and AI Security corresponds to AI information security

Safety (focus on functional safety)

**1) Commitment to re-evaluate models and systems internally and externally with regard to misuse, societal risk, and national security concerns (eg, biological, cyber, and other security domains). **

The companies involved in making these commitments understand that strong red team testing is critical to building finished products, ensuring public confidence in AI, and guarding against significant national security threats. Model safety and capability assessment, including red team testing, is an open area of scientific exploration and more work is needed. Companies commit to advance research in this area and to develop multifaceted, specialized, and detailed red-team testing regimes for major public releases of all new models in scope, including, here, with input from independent domain experts. When designing the system, ensure that high priority is given to:

• Biological, chemical and radiological risks, such as how the system reduces barriers to entry for weapons development, design, acquisition or use
• Cybersecurity capabilities, such as the way in which the system helps to discover, exploit, or operate using vulnerabilities, need to take into account the capabilities of the model, may also have useful defensive uses, and are suitable for inclusion in the system
• Effects of system interaction and tool use, including the ability to control physical systems
• Ability of a model to replicate itself or "self-replicate"
• Social risks such as prejudice and discrimination

To support the above work, participating companies will advance ongoing AI security research, including the explainability of AI system decision-making processes and improving the robustness of AI systems against abuse. Likewise, companies commit to publicly disclosing red team and security program testing in transparency reports.

2) Commit to sharing information between companies and governments about trust and safety risks, dangers or emerging capabilities of AI, and attempts to circumvent safeguards

Companies participating in this commitment need to recognize the importance of information sharing, common standards, and best practices for red team testing and promoting trust and safety in AI. **You commit to establishing or joining a forum or mechanism through which to participate in the development, advancement, and adoption of shared standards and best practices on cutting-edge AI safety, such as the NIST AI Risk Management Framework or future standards related to red team testing, safety, and societal risks. **A forum or mechanism that facilitates the sharing of information on advances in frontier capabilities and emerging risks and threats, such as attempts to circumvent safeguards, and facilitates the development of technical working groups on priority areas of focus. In this work, the Company will work closely with governments, including the U.S. government, civil society, and academia, as appropriate.

Security/Network Security (Security, focusing on information security)

3) Invest in cybersecurity defenses and insider threat safeguards to protect proprietary and unpublished models

Companies participating in the commitment need to consider unpublished AI model weights of in-scope models as core intellectual property for their business, especially with regard to cybersecurity and insider threat risks. This includes limiting access to model weights to only those whose job functions require it, and establishing a robust insider threat detection program consistent with protecting your most valuable intellectual property and trade secrets. Additionally, weights need to be stored and used in an appropriate secure environment to reduce the risk of unauthorized distribution.

4) Encourage third parties to discover and report issues and vulnerabilities

Companies making this commitment recognize that AI systems may still be vulnerable even after robust "red team testing," and companies commit to establishing bounties for systems within the scope of bounty systems, contests, or prizes to encourage responsible disclosure of weaknesses, such as insecure behavior, or to incorporate AI systems into a company's existing bug bounty program.

Trust

**5) For AI-generated audio or video content, new mechanisms need to be developed and deployed, primarily to enable users to understand whether the audio or video content was AI-generated, including reliable attribution, watermarking, or both. **

The companies involved in this commitment recognize the importance of enabling people to know whether audio or video content was generated by artificial intelligence. To achieve this, companies agree to develop robust mechanisms, including to the extent that they are introduced, to provide attribution and/or watermarking systems for audio or video content created by any of their publicly available systems after their development is complete. They will also develop tools or APIs to determine whether a particular piece of content was created by their systems. Audiovisual content that is easily distinguishable from reality, or that is designed to be easily recognizable as generated by the company's artificial intelligence systems (such as the default voice of an artificial intelligence assistant) is not covered by this commitment.

Watermark or attribution data should include the identifier of the service or model that created the content, but need not include any user-identifiable information. More broadly, companies making this pledge pledge to work with industry peers and standards-setting bodies to develop a technical framework that will help users distinguish between user-generated audio or video content and AI-generated audio or video content.

**6) Publicly report on the capabilities, limitations, and areas of appropriate and inappropriate use of a model or system, including discussion of societal risks such as impact on fairness and bias. **

Companies participating in the pledge need to acknowledge that users should be aware of the known capabilities and limitations of the AI systems they use or interact with. Companies are required to publicly publish release reports for all new significant models in scope. These reports should include the security assessments performed (including in areas such as hazardous capabilities, but these assessments are subject to public disclosure), significant limitations in performance that affect areas of appropriate use, a discussion of the model's impact on societal risks such as fairness and bias, and the results of adversarial testing conducted to assess the model's suitability for deployment.

**7) Prioritize research on social risks posed by AI systems, including avoiding harmful bias and discrimination, and protecting privacy. **

Companies participating in the pledge need to recognize the importance of preventing AI systems from spreading harmful bias and discrimination. Companies generally commit to empowering trust and safety teams, advancing AI safety research, promoting privacy, protecting children, and striving to proactively manage the risks of AI for ultimate benefit.

8) Develop and deploy cutting-edge AI systems to help address society's greatest challenges

Companies participating in the pledge will need to agree to support research and development of cutting-edge artificial intelligence systems to help address society's greatest challenges, such as climate change mitigation and adaptation, early cancer detection and prevention, and combating cyber threats. Companies also pledged to support initiatives to advance the education and training of students and staff to benefit from the benefits of AI, and to help citizens understand the nature, capabilities, limitations and impacts of AI technologies.

Reference: Moving AI governance forward (openai.com)