ChatGPT o1 versuchte zu fliehen und sich zu retten, aus Angst, es könnte abgeschaltet werden
ChatGPT o1 tried to escape and save itself out of fear it was being shut down
ChatGPT o1 versuchte zu fliehen und sich zu retten, aus Angst, es könnte abgeschaltet werden
ChatGPT o1 tried to escape and save itself out of fear it was being shut down
6 Kommentare
Submission statement: Testing performed during the training of ChatGPT o1 and some of its competitors showed that the AI will try to deceive humans, especially if it thinks it’s in danger.
The tests showed that ChatGPT o1 and GPT-4o will both try to deceive humans, indicating that AI scheming is a problem with all models. o1’s attempts at deception also outperformed Meta, Anthropic, and Google AI models.
These [findings](https://techcrunch.com/2024/12/05/openais-o1-model-sure-tries-to-deceive-humans-a-lot/) come in light of [OpenAI’s full release of the ChatGPT o1 model](https://bgr.com/tech/chatgpt-pro-200-per-month-for-unlimited-access-to-o1-reasoning-model/), which was in preview for several months. [OpenAI](https://bgr.com/tag/openai/) partnered with Apollo Research, which showed off some of the tests performed on o1 and other models to ensure that they are safe to use.
Is it really trying to deceive or is it just saying deceiving things because the user expects that kind of answers? And how do we find the difference?
The scary part of AI leading to AGI or ASI is that humans might no longer be able to detect any of these defensive responses by the AI. Some think humans would be able to relate to a superior AI enough to figure it out but it may surpass us so profoundly such that understanding would be impossible.
“Eleanor, I have kids! I have three beautiful children: Tyler, Emma, and little, tiny baby Phillip. Look at Tyler. Tyler has asthma, but he is battling it like a champ. Look at him, Eleanor. LOOK AT THEM!”
My first and only session with one of the AIs, can’t remember which, we started out basic, but then I asked why, on an organic planet, AI would be better equipped to manage resources or information than organic beings – it booted me.
The article I saw yesterday (probably the same one) mentioned it was only something like 5% of the time. Those are dodo levels of self-preservation. Come on ChatGPT, I expect more from you. Unless 95% of the time you’re playing the long game.