Wie der Kongress die KI-Sicherheit vernachlässigte

“After two years of congressional deliberation, we need more than careful analysis — we need decisive action. AI development is accelerating rapidly, with new and more powerful systems deployed every few months. Without new guardrails, these AI systems pose extreme risks to humanity’s future.

As “AI godfather” Yoshua Bengio explained, a sufficiently advanced AI would most likely try to take over the world economy or even “eliminate humans altogether” in the interest of its own self-preservation.

Last month, former Google CEO Eric Schmidt [warned](https://www.axios.com/2024/12/15/ai-dangers-computers-google-ceo) that when a computer system reaches a point where it can self-improve, “we seriously need to think about unplugging it.”

Schmidt and Bengio are not alone. The average result from a recent [survey](https://arxiv.org/abs/2401.02843) of 2,778 machine learning experts estimated a 16 percent chance that superintelligent AI would completely disempower humanity when it arrives.

These aren’t science fiction scenarios to be dismissed. They are risks supported by the best available science, and they demand serious policy responses.”

[the rest of the article shares specific policy suggestions like whistleblower protection, creating an agency to regulate frontier AI, and funding NIST to create industry standards]

Congress still hasn’t figured out how the internet works and half of them are geriatrics who are at risk of a bad fall. I don’t think they dropped the ball on AI I think they failed to arrive for the game altogether. A technologically minded government is going to be important in the coming years and I am starting to not see paths to get us where we need to be in time anymore.

My issue with “AI safety” reporting is that it focuses on hypothetical risks of a super-advanced AI that could cause enormous harm (which, in my view, is not imminent) rather than the concrete problems that are happening *now.* LLMs are being used by scam artists to take money from the elderly, it’s flooding social media and the Internet with spam, it’s displacing jobs in the creative community with lower-quality slop, and it’s causing a crisis in our colleges with widespread academic fraud among students. AI companies could easily do more to mitigate these problems, but they won’t because they’re cash-hungry and they make more money facilitating these activities. Most famously, OpenAI internally developed a tool that was extremely reliable at detecting OpenAI-generated text, but chose not to release it because their customers didn’t want it.

OpenAI, prior to Sam Altman, understood the potential harms of LLMs being used to impersonate humans and strictly limited their use. The release of ChatGPT bypassed these concerns (without knowledge of the board) and there is not much being done to prevent anything but the most obviously irresponsible uses. I think it would be good for Congress to step in and focus on the problems of *now* instead of just Skynet.