OpenAI Forms ‘Preparedness’ Team to Monitor and Test AI Technology for Safety

Leading healthcare entities have signed voluntary commitments to ensure the safe and trustworthy development and use of artificial intelligence in healthcare. These commitments aim to pursue the benefits of large-scale AI models while mitigating risks and protecting patient information. The organizations will align their projects with AI principles outlined by the U.S. Department of Health and Human Services, establish transparency and risk management frameworks, and continue responsible research and innovation. The goal is to deliver coordinated care, improve patient experiences, and reduce clinician burnout.

OpenAI is taking steps to address the potential dangers of AI by assembling a “Preparedness” team to monitor and test their technology for any dangerous capabilities. The team will work alongside other teams focused on safety and alignment. The popularity of ChatGPT and the advancement of generative AI have sparked debates about the risks and benefits of the technology. OpenAI aims to strike a balance between addressing longer-term risks and fixing existing problems. The company plans to hire experts from outside the AI field, collaborate with organizations like the National Nuclear Security Administration, and allow independent third parties to test its technology.

A dataset used to train AI models, including Google’s Imagen, was found to contain child sexual abuse material (CSAM). The dataset, called LAION, was compiled using web scraping tools, which inadvertently collected illegal imagery. The presence of CSAM in datasets raises concerns about the potential for AI models to generate or spread abusive content.

According to researchers, Google’s Gemini falls slightly short compared to OpenAI’s GPT-3.5 Turbo regarding language abilities. Google claims that Gemini Pro performs better than GPT-3.5 and that Gemini Ultra, set to release in early 2024, outperforms GPT-4. The researchers tested Gemini Pro, GPT-3.5 Turbo, GPT-4 Turbo, and Mistral 8x7B on various prompts. They found that Gemini Pro had lower accuracy than GPT-3.5 Turbo in multiple-choice questions and general-purpose reasoning. However, Gemini excelled in word sorting, symbol manipulation, and translation between languages.

The value of IT services is rooted in the guidance and advisory services provided, and knowing which companies have made what committements to safety and data integrity will be a key factor.

Let’s note that the responsibility for addressing and preventing the misuse of AI lies with dataset creators and organizations hosting the datasets. Efforts to clean and filter datasets may be labor-intensive and may only partially eliminate all inappropriate content. Legal and regulatory frameworks are still evolving to address CSAM’s challenges in AI datasets.   If it starts corrupted, it will remain corrupted, and providers are going to need to track this as an element of their services.