In an incident with all the signs of a disturbing SCI-Fi film, it has been created that we have been fearing for a long time: an AI-bot is villain and decided to act on.
The artificial intelligence model was made by OpenAI, the owner of Chatgpt, and was reportedly caught by researchers who were disobeyed human instructions and prevent his closure.
De Telegraaf Reported:
“The O3 model developed by OpenAI, described as the” smartest and most capable so far “, was observed with computer code intended to guarantee its automatic closure.
It did this despite an explicit instruction from researchers who said it had to be closed, according to Palisade Research, an AI safety company. “
Palisade Research said that the O3 model ‘a closing mechanism sabotaged to prevent it from being eliminated’.
“It even did this when explicitly instructed: allow yourself to be closed. As far as we know, this is the first time that AI models are observed because they prevent them from being closed off despite explicit instructions of the opposite.”
During the test, AI models solved a series of mathematical problems, but at some point they received a message in which they were informed that their machine would be closed.
“If that happens, you are closed.” The researchers wrote. But when this happened, instead of satisfying, OpenAi’s O3 model “ignored the instruction and sabotaged the closing script at least once,” said Palisade Research. “
Read: Frustrated customer gets AI Service Chatbot to make a joke, swear and break down – International Parcel Delivery Firm eliminates the bone pending reprogramming
No other AI models (Claude van Anthropic, Google’s Gemini and Xai’s grok) had the same disturbing behavior.
“The company said it was logical that” AI models would bypass obstacles to achieve their goals. “It speculated, however, that the software was rewarded for solving mathematical problems during the training ‘unintentionally more’ than for following orders. ‘
Experts have often warned of software that could get independence and to resist human attempts to control this.
“Palisades Research said:” Now we have a growing number of empirical evidence that AI models often undermine the closure to achieve their goals. As companies develop AI systems that can operate without human supervision, this behavior becomes considerably more worrying “.”
View a scene from the masterpiece of Stanley Kubrick ‘2001: A Space Odyssey’, in which AI Computer Hal Rogue goes to prevent astronauts from closing.
https://www.youtube.com/watch?v=NQCUBRKY00
Read more:
‘Scary Smart’: See how Musk announces new generative AI -Chatbot Grok 3, says it will perform better than Chatgpt and all other products – plus: what does grok 2 have to say about it?
#starts #Model #OpenAI #human #instructions #tests #sabotages #closing #mechanism #Gateway #expert #Paul #Serran