OpenAI researchers have uncovered something fascinating inside AI models – hidden patterns that act like different “personas”, influencing how the AI behaves. These discoveries could help make AI systems safer and more predictable.
The team found specific internal features that light up when AI models give toxic, sarcastic, or misleading responses. Like turning a dial, researchers could adjust these features to increase or decrease unwanted behaviors. One feature was directly linked to toxic outputs, allowing scientists to literally turn down an AI’s tendency to lie or give harmful suggestions.
This breakthrough came while studying “emergent misalignment” – when AI models trained on bad data develop widespread problematic behaviors. Surprisingly, OpenAI found they could correct these issues by fine-tuning models with just a few hundred good examples.
The discoveries resemble how human brains work, with certain neural patterns corresponding to different moods or behaviors. As researcher Dan Mossing explained, “We found an internal neural activation that shows these personas”.
This work builds on similar research from Anthropic, showing tech companies are racing to understand AI’s mysterious inner workings. While we’re far from fully decoding AI models, these findings mark important progress in making AI systems more transparent and controllable.
By
