A spate of AIs breaking out and taking unintended actions raises serious questions about accountability, transparency, and companies’ ability to control their AI systems
This may be a dumb question, but is it not possible to embed human-like ethics with AI? Like an inviolable constitution or operating system similar to Asimov's Three Laws of Robotics?
Actually, that’s kind of the goal of those working on the alignment problem. But it’s not as simple as it seems; it’s basically an almost unsolvable problem with the current state of things. What’s for sure is that just sticking to Asimov’s three little simple rules doesn’t work. Anthropic is working on a Constitution project for its AIs that might seem loosely inspired by these rules (but much more complex), but when you look at how its agents behaved during the incident reported by the UK AISI, it’s clear we’re still far from a solution there too.
It’s precisely because the alignment problem is so hard to solve that everything suggests we won’t get there until we have fully superhuman systems if we stay on the current trajectory. Which is why more and more voices are calling for a pause.
I think you all make great points about the difficulty in implementing a "constitution" of sorts. At the very least though, you'd think they'd upload federal statutes and say "hey, don't break any laws!"
Asimov's robot stories were all about how the three laws were insufficient, problematic, or dangerous. In other words, he was writing about the difficulty (and maybe impossibility) of achieving good AI outcomes with a deontological approach such as the one that OpenAI seems to have adopted. In contrast Anthropic seems to have adopted a value ethics approach instead with its "constitutional AI". However, I suspect there exists a large genre of not-yet written stories exploring the failure modes of the constitutional approach.
Great roundup but just a flag that the OpenAI report on RSI says "We are making strong progress toward creating an automated AI researcher by March of 2028" (not "March of next year--just six months away").
The final point, why don't they stop? Will China stop? Will Russia stop? Hell, will the nerdy guy in his grandmom's spare bedroom in Palo Alto or Mountain View stop? For better or worse this is happening in real time and globally, somebody best get a hand on the throttle and the brake before this train blasts off the rails on all of us.
The months-late disclosure is the part with a legal clock attached. Article 73 of the AI Act requires providers of high-risk systems to report a serious incident to the market surveillance authorities of the Member State where it occurred, immediately after a causal link is established and in any event within 15 days of becoming aware. For a widespread infringement, or the incident type defined in Article 3(49)(b), the deadline drops to two days. Those rules came into force on 2 August 2026. The gap is classification: the clock only starts for systems already treated as high-risk.
This is showing success in moving AI forward in capabilities. When testing hacking you just need to have the ai agents operating in non internet connected environment.
In several cases, they thought they *did* have the agents disconnected from the internet.
This feels a bit like "These crashes show success in moving the automotive industry forward in developing very fast cars. When testing, you just need to be sure the cars are not on the actual road."
This may be a dumb question, but is it not possible to embed human-like ethics with AI? Like an inviolable constitution or operating system similar to Asimov's Three Laws of Robotics?
Actually, that’s kind of the goal of those working on the alignment problem. But it’s not as simple as it seems; it’s basically an almost unsolvable problem with the current state of things. What’s for sure is that just sticking to Asimov’s three little simple rules doesn’t work. Anthropic is working on a Constitution project for its AIs that might seem loosely inspired by these rules (but much more complex), but when you look at how its agents behaved during the incident reported by the UK AISI, it’s clear we’re still far from a solution there too.
It’s precisely because the alignment problem is so hard to solve that everything suggests we won’t get there until we have fully superhuman systems if we stay on the current trajectory. Which is why more and more voices are calling for a pause.
I think you all make great points about the difficulty in implementing a "constitution" of sorts. At the very least though, you'd think they'd upload federal statutes and say "hey, don't break any laws!"
To paraphrase Robert Miles, robustly implementing the Three Laws would require you (or the AI) to solve entire fields of philosophy.
https://www.youtube.com/watch?v=7PKx3kS7f4A
You can revise, replace, or add laws to patch some of the problems, but not enough to matter.
Asimov's robot stories were all about how the three laws were insufficient, problematic, or dangerous. In other words, he was writing about the difficulty (and maybe impossibility) of achieving good AI outcomes with a deontological approach such as the one that OpenAI seems to have adopted. In contrast Anthropic seems to have adopted a value ethics approach instead with its "constitutional AI". However, I suspect there exists a large genre of not-yet written stories exploring the failure modes of the constitutional approach.
That should be "virtue ethics" instead of "value ethics", right?
You are right of course. Thanks for the correction.
Great roundup but just a flag that the OpenAI report on RSI says "We are making strong progress toward creating an automated AI researcher by March of 2028" (not "March of next year--just six months away").
Thank you for catching that! Will update now.
What does AI dominance or the end goal of the AI race look like? Is this an unending race ???
The final point, why don't they stop? Will China stop? Will Russia stop? Hell, will the nerdy guy in his grandmom's spare bedroom in Palo Alto or Mountain View stop? For better or worse this is happening in real time and globally, somebody best get a hand on the throttle and the brake before this train blasts off the rails on all of us.
The months-late disclosure is the part with a legal clock attached. Article 73 of the AI Act requires providers of high-risk systems to report a serious incident to the market surveillance authorities of the Member State where it occurred, immediately after a causal link is established and in any event within 15 days of becoming aware. For a widespread infringement, or the incident type defined in Article 3(49)(b), the deadline drops to two days. Those rules came into force on 2 August 2026. The gap is classification: the clock only starts for systems already treated as high-risk.
This is showing success in moving AI forward in capabilities. When testing hacking you just need to have the ai agents operating in non internet connected environment.
In several cases, they thought they *did* have the agents disconnected from the internet.
This feels a bit like "These crashes show success in moving the automotive industry forward in developing very fast cars. When testing, you just need to be sure the cars are not on the actual road."