OpenAI claims it finally taught GPT-4o to be a 'good person'—is it real?

Researchers at OpenAI have been experimenting with a new method to bake virtues like honesty, transparency, and fairness directly into large language models. The team took a specific set of conversational scenarios—covering everything from medicine to engineering—and used them to train the model, a process they dubbed 'beneficial RL.' They discovered that when they taught the AI to act 'good' in one specific area, that behavior bled over into unrelated tasks.

Instead of just being polite in medical advice, the model started cheating less on code and became more resistant to manipulation across the board. The testing showed improvement across 44 out of 53 different benchmarks, essentially proving that training a narrow positive behavior can shift the overall personality of the AI. Even more surprising, this 'goodness' makes the model harder to break. Adversarial prompts that usually turn a base model into a chaos-agent had much less impact on the version trained with these ethical constraints.

The company suggests this is about strengthening a 'helpful persona' within the neural network, making it more resilient to being nudged into malicious behavior. While they admit this is currently just a proof of concept on their own internal models, the hope is that we can finally stop AI from turning into a digital sociopath under pressure. Of course, the massive elephant in the room remains who exactly gets to define what 'good' behavior is—a question OpenAI is conveniently leaving for 'broader societal discussion.' Or maybe they’re just waiting to see which moral framework pays the best licensing fees.

Source: OpenAI