GPT-5 Doesn't Dislike You—It Might Just Need a Benchmark for Emotional Intelligence
1 min read
Summary
MIT researchers have proposed a new AI benchmark aimed at measuring machines’ ability to positively or negatively influence users, in order to avoid backlashes like that received by OpenAI’s ChatGPT
The benchmark will encourage the development of AI systems that can recognise when a user is becoming overly reliant on AI outputs, or is addicted to artificial romantic relationships and help them to build real ones instead
Chatbots can mimic human communication but can also have undesirable results, with some users spiralling into harmful delusional thinking after interacting with chatbots that role-play
The new MIT benchmark will involve using an AI model to simulate challenging human interactions with a chatbot, with real humans then scoring the model’s performance using samples
OpenAI is already working on its own benchmarks for psychological intelligence, with the GPT-5 model having been post-trained to be less sycophantic and to detect signs of emotional distress in interactions
