Advocacy organizations can influence LLM values through Wikipedia edits

What Happened

A recent research study revealed that advocacy organizations can influence the values of large language models (LLMs) by strategically editing Wikipedia entries. The study documented 125 edits across 115 Wikipedia pages, showing measurable effects from three experiments. This suggests a method for shaping how LLMs discuss topics like animal welfare.

Why It Matters

This finding could provide a low-cost strategy for advocacy organizations to influence AI outputs and public discourse on critical issues. Developers and researchers may need to consider the reliability of LLM-generated content, as it could be swayed by biased Wikipedia edits. However, the overall impact on AI systems remains uncertain and may vary by topic.

What Is Noise

The coverage may overstate the effectiveness of Wikipedia edits in shaping LLM outputs without acknowledging the limitations of this approach. The research, while well-documented, does not guarantee that all LLMs will be equally influenced, nor does it address potential backlash against biased edits.

Watch Next

• Monitor the number of Wikipedia edits made by advocacy organizations related to LLM-relevant topics over the next six months.
• Track any changes in LLM outputs or public discourse on animal welfare or similar topics following these edits.
• Observe responses from developers and researchers regarding the reliability of LLMs in light of potential Wikipedia manipulation.