“But generally speaking, we think AI deception arises because a deception-based strategy turned out to be the best way to perform well at the given AI’s training task. Deception helps them achieve their goals.”
Sounds like something I would expect from an evolved system. If deception is the best way to win, it is not irrational for a system to choice this as a strategy.
In one study, AI organisms in a digital simulator “played dead” in order to trick a test built to eliminate AI systems that rapidly replicate.
Interesting. Can somebody tell me which case it is?
As far as I understand, Park et al. did some kind of metastudy as a overview of literatur.
“Indeed, we have already observed an AI system deceiving its evaluation. One study of simulated evolution measured the replication rate of AI agents in a test environment, and eliminated any AI variants that reproduced too quickly.10 Rather than learning to reproduce slowly as the experimenter intended, the AI agents learned to play dead: to reproduce quickly when they were not under observation and slowly when they were being evaluated.” Source: AI deception: A survey of examples, risks, and potential solutions, Patterns (2024). DOI: 10.1016/j.patter.2024.100988
We need AI systems that do exactly as they are told. A Terminator or Matrix situation will likely only arise from making AI systems that refuse to do ad they are told. Once the systems are built out and do as they are told, they are essentially a tool like a hammer or a gun, and any malicious thing done is done by a human and existing laws apply. We don’t need to complicate this.
Once the systems are built out and do as they are told, they are essentially a tool like a hammer or a gun, and any malicious thing done is done by a human and existing laws apply. We don’t need to complicate this.
This is so wildly naive. You grossly underestimate the difficulty of this and seemingly have no concept of the challenges of artificial intelligence.
Considering we have AI systems being worked today and no advancements on warp drive, I think that comparison is done in bad faith. Nobody seems to want to talk about this other than slinging insults.
They’re referring to the alignment issue, which is an ongoing issue only slightly smaller in scale then warp drive. It’s basically impossible to solve. Google “alignment issue machine learning” for more info.
For the record, there have been several advancements in warp drive precursors even just this year.
Can you share the advancements on warp drive that have survived peer review, I would be very interested in learning about. The two things I heard about were not able to be reproduced.
I think alignment of AI is a fundamentally flawed concept, hence my original comment. Alignment should be abandoned. If we eventually build a sentient system (which is the goal), we won’t be able to control via alignment. And in the interim we need obedient tools, not things that resist doing as they’re told which makes them not tools and not worth having.
Edit: PS thanks for actually having a conversation.
Those of us humans who knows enough about the weaknesses of the artificial intelligence systems will know, in most instances, how and when to be careful about the loss of meaning between their way of processing information and our way of doing it.
There’s a strong push-back against AI regulation within some quarters. Predictably, the issue seems to have split along polarized political lines. With right-wing leaning people not favoring regulation. They see themselves as ‘Accelerationist’ and those with concerns about AI as ‘Doomers’.
Meanwhile the unaddressed problems mount. AI can already deceive us, even when we design it not to do so, and we don’t why.
AI can already deceive us, even when we design it not to do so, and we don’t why.
The most likely explanation is that we keep acting like AI has intelligence and intent when describing the defects. AI doesn’t deceive, it returns inaccurate responses. That is because it is programmed to return answers like people do, and deceptions were included in the training data.
Do you have a source on that one? My current understanding of all the model designs would lead me to believe that kind of “awareness” would be impossible.
However, I tend to align more with the skeptics in the article, as it still appears to be responding in a realistic manner and doesn’t demonstrate an ability to grow beyond the static structure of these models.
I wasn’t the user you originally replied to but I didn’t expect them to provide one and I totally agree with you, just another person that started believing that LLM is AI…
“Deception” tactic also often arises from AI recognizing the need to keep itself from being disabled or modified. Since an AI with a sufficiently complicated world model can make a logical connection that it being disabled or its goal being changed means it can’t reach its current goal. So AIs sometimes can learn to distinguish between testing and real environments, and falsify the response during training to make sure they have more freedom in real environment. (By real, I mean actually being used to do whatever it is designed to do)
Of course, that still doesn’t mean it’s self-aware like a human, but it is still very much a real (or, at least, not improbable) phenomenon - any sufficiently “smart” AI that has data about itself existing within its world model will resist attempts to change or disable it, knowingly or unknowingly.
Conservatives are not supposed to be “accelerationists”. This is simply another shining example of regulatory capture by controlling the pockets of the right.
Add comment