Stanford researchers have developed a scoring system called the Foundation Model Transparency Index to rate the transparency of large AI language models. The index evaluates models based on criteria such as disclosure of training data sources, hardware information, labor involved, and downstream indicators. While Meta’s Llama 2 scored the highest, followed by BloomZ and OpenAI’s GPT-4, none of the models received high marks. The researchers argue that increased transparency is crucial as AI models become more powerful and integrated into daily life, allowing for better understanding, regulation, and awareness of potential risks.
Yoshua Bengio and Geoffrey Hinton, two of the so-called AI godfathers, have joined with 22 other leading AI academics and experts to propose that companies and governments allocate a third of their AI research and development budgets to AI safety. The paper also calls for a comprehensive insight into AI development, model registration, and accountability for harms caused by advanced AI models. The authors emphasize the risks posed by autonomous AI systems and the need for research breakthroughs in oversight, robustness, interpretability, risk evaluation, and addressing emerging challenges.
The idea of mandatory safety audits for AI models is gaining popularity as a way to ensure a safe future with powerful AI. Audits, similar to financial audits, would independently assess the risks of new AI systems. The concept has been included in the EU’s Digital Services Act and a proposed AI policy framework by senators Josh Hawley and Richard Blumenthal. Pre-deployment and post-deployment audits would examine AI models’ plans and functioning. However, auditing AI models is complex due to their inner workings, and auditors would rely on access to training data and observations of inputs and outputs. Despite limitations, the idea of independent oversight through audits is seen as a way to address concerns about AI’s impact.
As AI models like GPT-4 become integrated into business solutions and workflows, MSPs will find themselves responsible for explaining these technologies to customers. Transparency in AI, as measured by Stanford’s index, is not just an ethical concern but a practical one. Being informed about the transparency score of an AI model could help MSPs guide decision-making for clients who are concerned about ethical and operational risks.
A greater focus on safety could result in new compliance and governance policies that MSPs would need to understand and implement.
Should safety audits for AI models become a standard, MSPs would need to be equipped to understand what these audits entail and how to interpret them. The implications for governance, compliance, and even service offerings could be significant. MSPs offering AI solutions might themselves become part of the auditing process, necessitating a solid understanding of AI mechanics and ethics.
I love this entire space.