OpenAI recently conducted a test on its ChatGPT-4o language model and identified concerning issues in the results. The organization stated its intention to thoroughly assess new models for potential risks and establish appropriate safeguards before implementing them into ChatGPT or APIs. This evaluation aims to identify and mitigate potential risks such as misinformation, discrimination, and privacy policy violations associated with aggregated models.
The assessments, conducted from early March to late June, involved over 100 individuals from 29 countries and included 45 different languages. The findings revealed potential risks related to unauthorized voice generation, the potential creation of pirated content, and the dissemination of unauthorized content. OpenAI also expressed concern over the impact of ChatGPT-4o’s increasingly human-like behavior on user-chatbot interactions.
OpenAI pointed out that the audio capabilities of GPT-4o facilitate the humanization of models and pose a growing risk in terms of user interactions. The organization highlighted that while the humanization of AI models may benefit lonely individuals, it could potentially diminish the need for human interaction, impacting healthy relationships.
Recognizing these risks, OpenAI is committed to monitoring these behaviors, understanding their potential dangers, and finding ways to address the issue. The organization plans to conduct further research on the emotional capabilities of AI and explore methods to integrate the features of GPT-4o into their systems while avoiding behaviors that may negatively affect human relationships in the workplace.
