Seven of the top AI companies, including Meta, Google, and OpenAI, have revealed their security plan after attending a White House summit and reaching a deal with the Biden administration. The companies have agreed to implement new guardrails to enhance the safety and security of their systems and users. One of the key measures is the implementation of watermarking for AI-generated content.
During the summit, President Joe Biden emphasized the significance of these promises and their potential to positively impact people’s lives worldwide. He highlighted the crucial role these companies play in leading innovation and urged them to prioritize responsibility and safety in their design processes.
The companies, including Amazon, Anthropic, Meta, Google, Inflection, and OpenAI, made voluntary commitments to invest in research, safety, and security stress testing. They also agreed to support third-party audits to identify system vulnerabilities. While the companies have made specific promises, they have not provided detailed information regarding their company-wide hiring plans. Currently, there are no strict enforcement mechanisms in place to ensure the companies follow through on their commitments.
In a statement, Meta’s global president Nick Clegg emphasized the importance of building and deploying AI technologies responsibly and transparently. He called for close collaboration across industry, government, academia, and civil society to ensure that these powerful new technologies benefit society as a whole.
Anthropic, on the other hand, provided the least descriptive statement, merely expressing the need for everyone to prioritize AI safety. The company announced its intention to share concrete plans regarding cybersecurity, red teaming, responsible scaling, and hoped that other companies would also take swift action.
One significant consumer-focused initiative that emerged from the summit is the agreement to watermark AI-generated content. White House officials revealed that both audio and visual content created by individual users will be required to have watermarks. OpenAI stated that tools and APIs would be developed to identify content created on their systems. However, certain content, such as AI voice assistants, will not be subjected to watermarks.
Google has previously promised to introduce similar disclosure measures, and the company’s global president, Kent Walker, reiterated this commitment in a statement. Google plans to integrate watermarking, metadata, and other innovative technologies into its upcoming system.
Many of the companies involved in the White House summit had already made comparable pledges to protect their technology. Meta recently announced that it would open source its large-scale language model, Llama 2, making it freely available for researchers and most commercial use. OpenAI also made its GPT-4, the language model powering ChatGPT and Microsoft Bing, available to the public. Additionally, Meta agreed to publicly evaluate its model at this year’s Def Con.
In conclusion, the deal reached between the top AI companies and the Biden administration reflects their collective commitment to improving the safety and security of their systems and users. The implementation of watermarking for AI-generated content is a significant step towards accountability and transparency. However, it remains to be seen how effectively these commitments will be enforced and whether they will truly benefit society as a whole.
