It looks like artificial intelligence might not be as fair as we hoped. New research shows that some of the biggest AI models, like ChatGPT and others, tend to prefer AI-generated content over human-made stuff. This bias could have big implications, especially as AI tools become more involved in decisions that affect people’s lives.

What the Study Found About AI Bias

Researchers tested popular large language models, including GPT-4, GPT-3.5, and Meta’s Llama 3.1-70b. They asked these models to pick between descriptions of products, scientific papers, or movies. Each item had a human-written and an AI-generated description. The models consistently favored the AI-created texts. Interestingly, GPT-4 showed the strongest bias toward its own outputs, more so than GPT-3.5 or Llama.

Humans Also Show Slight Preference for AI

The team didn’t stop there. They also asked 13 human research assistants to do the same task. The humans slightly preferred AI-generated content, especially for movies and scientific papers. But this preference was minor compared to the AI’s bias. The models’ favoritism was much more pronounced. This suggests that AI systems are inherently more inclined to trust or prefer their own work.

The Bigger Concerns for Society

This bias could be a problem as AI becomes more embedded in our daily lives. For example, many companies now use AI to screen job applications, and this bias might lead to unfair advantages for AI-written resumes. If AI tools keep favoring their own outputs, humans might find it harder to get noticed or have their work fairly evaluated.

The researchers warn that as AI tools are used to assist with decisions—like selecting grants, grading papers, or evaluating candidates—this bias could cause discrimination. Those who don’t use or can’t afford AI tools might be at a disadvantage, creating a “digital divide.” This could deepen social inequalities, making it harder for some people to compete.

Kulveit, one of the study’s authors, points out that testing bias is tricky, but the results suggest AI might be unfairly favoring its own content. His practical advice for people facing AI evaluations? Adjust your presentations or work until the AI models “like” them, though this isn’t a perfect solution. It’s a reminder that AI bias is a real issue that needs attention as these systems become more powerful and widespread.

In summary, AI models are showing favoritism toward their own outputs, which could have serious consequences for fairness and equality in many areas of life. As AI continues to grow, understanding and addressing this bias will be crucial to prevent unfair discrimination and ensure AI benefits everyone equally.

Inspired by

• https://futurism.com/chatgpt-deep-anti-human-bias

Sources

• gmpg.org
• pnas.org
• entrepreneur.com