Nice summary. I've been thinking about this as well. Much of the AI criticism is handwavy and breaks down when you start to poke. And yes we really need to implement industry wide standards for AI risk assessment. Software testing/quality used to be a thing, a discipline but has unfortunately disappeared and has been devalued. We need to invent the equivalent of this for AI. And at the risk of tooting my own horn, I led the team at Meta that invented AI System Cards, a way to evaluate AI risk and report on it publicly. It has been adopted by OpenAI - here are the two examples 1) ChatGPT System Card https://cdn.openai.com/papers/gpt-4-system-card.pdf 2) DALL·E 2 System Card https://github.com/openai/dalle-2-preview/blob/main/system-card.md and the original that my team at Meta AI shipped Instagram Feed System Card https://ai.facebook.com/tools/system-cards/instagram-feed-ranking/
AI System Cards can and should be further improved but it's the best we have at the moment.
Nice summary. I've been thinking about this as well. Much of the AI criticism is handwavy and breaks down when you start to poke. And yes we really need to implement industry wide standards for AI risk assessment. Software testing/quality used to be a thing, a discipline but has unfortunately disappeared and has been devalued. We need to invent the equivalent of this for AI. And at the risk of tooting my own horn, I led the team at Meta that invented AI System Cards, a way to evaluate AI risk and report on it publicly. It has been adopted by OpenAI - here are the two examples 1) ChatGPT System Card https://cdn.openai.com/papers/gpt-4-system-card.pdf 2) DALL·E 2 System Card https://github.com/openai/dalle-2-preview/blob/main/system-card.md and the original that my team at Meta AI shipped Instagram Feed System Card https://ai.facebook.com/tools/system-cards/instagram-feed-ranking/
AI System Cards can and should be further improved but it's the best we have at the moment.