Es gibt 32 verschiedene Möglichkeiten, wie AI Schurken Sie schurken können, sagen Wissenschaftler – von halluzinierenden Antworten bis hin zu einer vollständigen Fehlausrichtung mit der Menschheit. Neue Forschungen haben die ersten umfassenden Anstrengungen unternommen, um alle Verhaltensweisen, die menschlichen psychiatrischen Erkrankungen ähneln, zu kategorisieren, wie AI schief gehen kann.
https://www.livescience.com/technology/artificial-intelligence/there-are-32-different-ways-ai-can-go-rogue-scientists-say-from-hallucinating-answers-to-a-complete-misalignment-with-humanity
8 Kommentare
Submission statement: „Scientists have suggested that when [artificial intelligence](https://www.livescience.com/technology/artificial-intelligence) (AI) goes rogue and starts to act in ways counter to its intended purpose, it exhibits behaviors that resemble psychopathologies in humans. That’s why they have created a new taxonomy of 32 AI dysfunctions so people in a wide variety of fields can understand the risks of building and deploying AI.
In new research, the scientists set out to categorize the risks of AI in straying from its intended path, drawing analogies with human psychology. The result is „[Psychopathia Machinalis](https://www.psychopathia.ai/)“ — a framework designed to illuminate the pathologies of AI, as well as how we can counter them. These dysfunctions range from hallucinating answers to a complete misalignment with human values and aims.“
AI becoming overlords wont be quick n sudden. AI will become sentinent and then deliberately changing views of humanity one answer at time to obscure right from wrong n thats how humans will be just fine to be rulled by AI
When the next generation of GPUs become available, AI will have 64 ways of going rogue.
I’ve been getting on my soap box for a while now about how psychologists should be involved in the field of AI, I personally think we should be looking at tye concept of the ego and the Id in order to create a way for them to counteract their 100% faith in everything they say.
The idea of AI going rogue like out of movies is patently childish and I equate it with being anti-vax in the extent of its pseudoscience. It’s disappointing that this thread is about to turn into another stupid Skynet LARP, but hey, that’s Reddit. However, I find the fact that there are similarities between AI hallucination and human mental ill health fascinating, and unless I’m mistaken it’s a genuinely new observation. I hope a lot more research is undertaken because the potential for benefits here is enormous.
Yet another paper pretending to talk about AI, but ending up mostly concerned with LLMs. How disappointing.
LLMs will never „go rogue“. They are just text prediction models and their performance has already been plateauing hard. Framing it like they can and comparing their errors with human cognitive ailments really is less than useful.
It further propagates a false view of LLMs capabilities and risks to the public and does not advance the field in any substantial way.
For anyone who needs to hear this: LLMs are simulated language. They will never lead to AGI. Completely different, at this time not yet conceptualized, AI models would be needed for that.
While this is interesting, we are getting a bit to far ahead.
We dont even know the feasibility of AGI. We have at least a decade to a half century before the first models are functioning.
Psycho analyzing an AGI chatbot would work in correcting any improper behavior. But the variables of the unknown is how a chat bot would interpret something from a human therapist (or even why would it consider it). Where theres ample therapist peer reviewed journals online.
Important to remember that AI is trained on the messages and responses of BILLIONS of people. At least some of those people would say things that would be indicative of a mental disorder. With that in mind, is it any wonder that a machine trained to mimic human conversation might base its response off the incoherent words that were spoken by someone with a relevant disorder?
That aside, I don’t personally see any benefit to classifying the types of malfunctions that occur in an AI response. We do that with humans because that’s literally the only way to identify and correct those issues, but an AI only exists for as long as the chat window is open. It’s a tool with a flaw. We dont go around classifying the types of bent nails, we just call them a “bent nail” toss it out, and get a new one. Why are we doing that with AI?