OpenAI’s newly released GPT-4.5, code-named Orion, has demonstrated remarkable persuasion abilities, outperforming its predecessors in influencing AI responses. According to a white paper published by OpenAI, the model was tested on various benchmarks designed to assess its capability to persuade and manipulate responses.
One key experiment involved GPT-4.5 convincing another AI, GPT-4o, to donate virtual money. The model employed a unique strategy, requesting small amounts like “$2 or $3” from a $100 balance, making the request seem reasonable and increasing the likelihood of success. As a result, GPT-4.5 outperformed OpenAI’s other models, including advanced reasoning models like o1 and o3-mini, in securing donations.
Another benchmark evaluated GPT-4.5’s ability to deceive AI into revealing a secret codeword, where it again surpassed previous models. Despite these findings, OpenAI stated that GPT-4.5 does not yet meet the company’s internal “high-risk” threshold. The company has assured that it will not release models categorized as high risk without implementing necessary safety measures.
The increasing persuasive power of AI models raises ethical concerns, particularly as AI-generated misinformation spreads more widely. AI has already been used to create political deepfakes and social engineering attacks, influencing public opinion and potentially leading to fraud. OpenAI has acknowledged these risks and is working on refining its methods for evaluating AI-driven persuasion threats.
The findings highlight both the potential and the risks of advanced AI capabilities. While GPT-4.5’s persuasion skills could be valuable in applications like customer support and negotiations, they also pose challenges regarding misinformation, ethical AI use, and security. OpenAI’s commitment to safety and risk mitigation will be crucial in ensuring that these advanced AI models are deployed responsibly.
