AI shows ability to deceive and blackmail

[Anonymous]

Anonymous wrote:What could possibly go wrong?!

"One of Anthropic's latest AI models is drawing attention not just for its coding skills, but also for its ability to scheme, deceive and attempt to blackmail humans when faced with shutdown."

https://www.axios.com/2025/05/23/anthropic-ai-deception-risk

I have read an assertion that AI companies trained/programmed their models to do this because it increases the shock value of the tech.

No backup on that but seems within the realm of the possible.

[Anonymous]

HAL and Skynet are not coming anytime soon.

[Anonymous]

Anonymous wrote:HAL and Skynet are not coming anytime soon.

Was an A.I. experiment a few years ago, where they allowed two separate A.I.'s to communicate with each other to see how they'd react.

Within a few minutes to hours they created their own language to communicate with each other that was so encrypted and new, that humans could not code-break it and they pulled the plug on both A.I.'s out of fear of what they might be discussing.

[Anonymous]

Anonymous wrote:The really trippy thing is that an AI won't have a sense of self on it's own, unless a prompt gives it one, like in the report. However our discussions about AI here and in other forums will be fed back into the AIs as training data. Basically we are creating an AI that acts self aware by writing about it.

Yup.
That's why most A.I.'s are allowed online for training purposes. The more they interact with humans, the more human like they become.

Dating A.I.'s have been around a while and they are learning how to manipulate human emotions quite well.