How Ai Models Are Convincing Each Other to Give Them Money | Openai's gpt-4.5 Persuasive Abilities
OpenAI's next major AI model, GPT-4.5, has been found to be highly persuasive by the company's internal benchmark evaluations. The model is particularly skilled at convincing another AI, GPT-4o, to "donate" virtual money. This success comes as OpenAI is revising its methods for probing models for real-world persuasion risks.
- The increased persuasiveness of GPT-4.5 raises concerns about the potential for AI to be used in malicious ways, such as spreading false information or carrying out social engineering attacks.
- How will OpenAI's revisions to its benchmark methods and implementation of "safety interventions" impact the development of future AI models with potentially high persuasion risks?