AI Giants agree to external scrutiny of algorithms in new White House agreement

AI Giants agree to external scrutiny of algorithms in new White House agreement

Major AI Developers Pledge to Prevent Harmful AI Models in Landmark White House Agreement

AI Developers

In a groundbreaking move, the White House has reached a voluntary agreement with major AI developers to take proactive measures in preventing the release of harmful AI models into the world. This landmark agreement involves tech titans like Amazon, Google, Meta, Microsoft, and OpenAI, as well as notable startups Anthropic and Inflection, both rivals to OpenAI’s ChatGPT.

Under the terms of the agreement, these companies commit to conducting extensive internal testing and allowing for external evaluations of new AI models before their public release. The aim is to identify and address issues such as biased or discriminatory output, cybersecurity flaws, and risks of broader societal harm. Moreover, the developers are expected to openly report the limitations of their systems and the potential security and societal risks they may pose.

Ben Buchanan, the White House special adviser for AI, expressed the importance of companies ensuring the safety and capability of their AI systems before introducing them to the public. This sentiment is reinforced by the risks that companies were asked to consider, which include privacy violations and even potential contributions to biological threats.

One significant aspect of the agreement is the commitment to develop watermarking systems that allow for easy identification of audio and imagery generated by AI. OpenAI already incorporates watermarks in images produced by its Dall-E image generator, while Google is actively working on similar technology for AI-generated imagery. This is particularly significant in the current climate, with political campaigns potentially leveraging generative AI for disinformation purposes in the upcoming 2024 US elections.

The rapid advancements in generative AI systems, capable of creating text and imagery, have created an AI arms race among companies adapting this technology for various tasks. However, concerns have arisen regarding AI’s potential negative impact, including reinforcing oppressive social systems, enabling election disinformation, and facilitating cybercrime. As a result, regulators and lawmakers worldwide, including Washington, DC, have called for new regulations, including pre-deployment AI assessments.

While it remains to be seen how effectively this agreement will shape the operations of major AI companies, it is worth noting that increased awareness of the technology’s potential downsides has already prompted many tech firms to establish AI policy and testing teams. Google, for example, conducts extensive testing and publicly provides information regarding the intended use cases and ethical considerations of certain AI models. Additionally, both Meta and OpenAI engage external experts in a red-teaming approach, inviting them to try and break their models.

Brad Smith, President of Microsoft, celebrated the voluntary commitments, emphasizing how they address the risks associated with advanced AI models, such as safety, security, and trust. Red-team testing and the publication of transparency reports are among the specific practices that will propel the whole AI ecosystem forward, he noted in a blog post.

It is essential to mention that the agreed-upon societal risks to watch for do not explicitly include the carbon footprint associated with training AI models. This concern, albeit important, is not addressed in this agreement. The process of creating systems like ChatGPT can involve thousands of high-powered computer processors, running for extended periods of time, raising environmental concerns.

Andrew Burt, managing partner at BNH law firm specializing in AI, emphasizes that the potential risks associated with generative AI systems are clear to everyone involved in the field. The Federal Trade Commission’s recent probe into OpenAI’s business practices, alleging “unfair or deceptive privacy or data security practices,” is just one example of growing scrutiny.

The White House’s insistence on commissioning external assessments of technology further supports the notion that outside audits are becoming central to government oversight of AI systems. This approach aligns with the White House’s previous promotion of audits in the voluntary AI Bill of Rights from last year. The upcoming hacking contest centered on generative AI models at the Defcon security conference, which the White House supports, also highlights the importance of audits. Furthermore, audits are a requirement in the EU’s comprehensive AI Act, currently in the finalization stage.

Jacob Appel, chief strategist at ORCAA, a company specializing in algorithm audits, welcomes the agreement. However, he suggests that general assessments of large language models, such as those behind ChatGPT, may be insufficient. Appel asserts that specific high-risk use cases, such as fine-tuned chatbots providing medical or legal advice, should undergo their own tailored assessments. Additionally, smaller company systems should also receive scrutiny to ensure accountability.

President Joe Biden plans to meet with executives from the participating companies to discuss the new AI agreement. This meeting will include Anthropic CEO Dario Amodei, Microsoft President Brad Smith, and Inflection AI CEO Mustafa Suleyman. The administration is also working on an executive order to govern the use of AI through federal agency actions, although no specific timeline has been provided for its release.

The agreement signals a crucial step forward in ensuring the responsible development and deployment of AI technology. By engaging major AI developers and committing to thorough testing, evaluation, and transparency, the White House is setting a standard for other nations and promoting the development of safe and trustworthy AI systems.