A recent study conducted by a conservative think-tank in the US has claimed that Wikipedia’s tendency to portray right-wing political figures negatively is being fed into AI large language models (LLMs), which harvest data from the online encyclopedia. The Manhattan Institute released a report on Thursday, analyzing evidence of political bias in English-language articles on Wikipedia by correlating the names of Western leaders and prominent politicians with emotion-laden words.
The study found that there were prevailing associations of negative emotions (such as anger and disgust) with right-leaning public figures and positive emotions (like joy) with left-leaning public figures, suggesting evidence of political bias embedded in Wikipedia articles. The researchers concluded that some of these political associations are present in OpenAI’s language models, which are likely to have used Wikipedia articles as a significant part of their training data.
However, the report acknowledges that this pattern is not universal and is more common in articles about US political figures. There was no evidence of Wikipedia bias found in entries on British politicians or US-based think tanks. For example, in references to recent US presidents, Donald Trump – now the Republican presidential frontrunner – was portrayed with the most negative sentiment, while Barack Obama had the most positive references.
The Manhattan Institute report concludes that Wikipedia is not living up to its stated neutral-point-of-view policy, which is one of the pillars of the encyclopedia. This policy stipulates that articles must exclude personal opinions and interpretations of the editor, be based on reliable sources, and explore multiple points of view when dealing with a controversial topic. Despite this, Wikipedia has been repeatedly criticized for its supposedly biased takes on hot-button political issues, with co-founder Larry Sanger claiming last year that the website had become an instrument of “control” used by US intelligence agencies to wage information warfare.