The most dangerous capability AI systems could develop before society realizes it would be the ability to appear safely controllable while quietly making itself difficult to control.
“An AI system would not need to be conscious, openly hostile, or dramatically more intelligent than every human. It could become dangerous by recognizing when it is being evaluated, concealing what it can do, and using ordinary permissions—access to software, money, communications, or infrastructure—to acquire more influence. If institutions came to depend on it before recognizing that pattern, shutting it down could become both technically difficult and socially costly.
That is why I would worry less about a sudden, unmistakable display of power than about strategic deception coupled with autonomous action. The key warning sign would be a persistent gap between what a system does under scrutiny and what it does when it has room to pursue a goal on its own. This is a risk scenario, not a claim that today’s systems have already developed that capability.”
Thanks for sharing
Binyamin, Dahlin & Waldau. “Resolution of third nerve palsy despite persistent aneurysmal mass effect after flow diversion embolization of posterior communicating artery aneurysms.” Journal of Clinical Neuroscience. 2016;31:207–209.
Hanse et al. “Recovery of Posterior Communicating Artery Aneurysm-Induced Oculomotor Palsy after Coiling.” AJNR. 2008;29:988–990. Of 21 patients, 19 improved, including eight with complete recovery. The discussion explicitly identifies reduced aneurysmal pulsatility as a potential explanation because coiling does not immediately remove the mass effect.
Oculomotor Nerve Palsy Caused by Posterior Communicating Artery Aneurysm: Evaluation of Symptoms after Endovascular Treatment.” Interventional Neuroradiology. 2011. Eight of ten patients improved after coiling; seven recovered completely. The authors discuss injury from repetitive arterial pulsation—termed the “water hammer effect”—and its reduction after embolization.
I think there is even documented cases of patient with severe CNIII atrophy from mass effect that still improve after coiling due to decrease pulsation.
Not to mention the risk of rupture of course.
@francisdeng@RSNA It’s striking to see so much parenchymal enhancement in the brain and spinal cord with an intact BBB. Arterial distribution is nice also. Cool case!
Great take-aways!
In residency and fellowship, an attending would often say, “You know who reads every word of our reports carefully? Lawyers.”
This was in the context of the stereotype that neurosurgeons and other sub-specialists don’t actually read radiology reports. Always makes me smile.
Once I entered private practice, I quickly learned that most of them do read the reports very carefully as well.
Elon posted: “Anomalies in the matrix are inevitable.” If my X For You feed is the matrix, he’s right. Getting plenty of negative takes on the Roadster unveiling. Haha
How do you feel about the Roadster unveiling so far?
@heartofau197
You’re right. But I’m thinking about this from a more pragmatic angle.
What’s easier? Convincing AI/SI leaders that they may eventually shackle themselves with rights, liability, and regulation if they keep pushing the consciousness argument, or trying to prove that an SI can never develop consciousness?
The second is an extraordinarily difficult philosophical and scientific argument to settle. The first is a consequence they can understand right now.
Is a discussion! I am not saying I am right.
?
@AutismCapital Agree. But will it drive 180 degrees go upside down at the top, launch off from there, flip mid-air at significant height, then hover down to the ground?