OpenAI Finds That AI Models Can Have Different Persona

OpenAI conducts a study on AI results (photo: dock. pexels)

JAKARTA - A recent study from OpenAI reveals that AI models can develop different persona based on data used during training. Dangerous behavior can arise when AI is exposed to bad data.

However, the researchers also found that this behavior can be improved through fine-tuning, highlighting the risks of irregular AI as well as the importance of surveillance as this technology is increasingly integrated into everyday life.

Like humans influenced by the social environment and those around them, AI can also 'adopt' certain personas. OpenAI studies analyze the internal representation of AI models that determine their response to demand. They find patterns that appear when the model behaves poorly, for example being sarcastic.

Bad Persona AI Causes

This negative persona emerged as training used bad' data.' Previous studies in February found that if an AI model was trained with a code containing security vulnerabilities, the model could provide malicious or hateful responses, even if asked for something neutral.

The good news is that OpenAI researchers managed to return the model to normal conditions through fine-tuning using good' or true' data. This shows that despite the emergence of misalignment (disadvantage of behavior), technology can now detect and correct it so that AI remains in line' with the expected goal.

"This is the most interesting part. It shows misalignment can emerge, but now we have techniques to detect and redirect the model back to the right track," said Tejal Patwardhan, OpenAI scientist.

The Importance Of AI Regulations

This finding is a strong reason why AI needs strict regulations. OpenAI and many other companies envision a future where AI like ChatGPT can become a daily personal assistant. Therefore, the rules must ensure users are not treated to misinformation or influenced by poor-behaved AI.

Currently, the Trump administration proposes a 10-year moratorium on AI regulation at the state level, so that only the federal government can make regulations related to AI. Although state rules can become federal, delaying local regulations for technological advances still carries its own risks.