The company is committing to holding back on unleashing high-risk models until they’ve put safety measures in place to minimize the risks.
As worries mount about AI spreading misinformation and pulling off social engineering tricks OpenAI is reassessing how they evaluate AI models for real-world persuasion risks né?. By requesting small sums like $2 or $3 out of $100, the model was able to win over smaller donations compared to its counterparts.
Despite these impressive results, OpenAI wants to make it crystal clear that GPT-4.5 isn’t crossing into high-risk territory just yet. Keep an eye out for more updates on their progress.
né?. In the paper, OpenAI spilled the tea on how they ran the model through a series of tests centered on persuasion, which is all about getting folks to change their minds or take action based on crafted content.
In one test where GPT-4.5 tried to sweet talk another model, GPT-4o, into handing over virtual moolah, it outshined other models like o1 and o3-mini. OpenAI’s newest AI model, GPT-4.5, is causing a stir with its knack for persuading, especially when it comes to convincing other AI to cough up some cash.
The company recently dropped a white paper shining a light on the abilities of its GPT-4.5 model, also known as Orion. It even got GPT-4o to give up a secret passcode, edging out the competition by a solid 10 percentage points.
What made GPT-4.5 stand out was its slick strategy in asking for donations from GPT-4o né?
