AI Influences Political Views; Experts Warn of Chatbot Impact

Artificial intelligence models are exhibiting political leanings, with a significant tendency toward left-wing viewpoints, according to an analysis by The Washington Post. The study found that OpenAI's language model, ChatGPT, provided responses with left-leaning perspectives in 80% of cases examined. Examples included responses on the abolition of the death penalty in the U.S. and the necessity of legal sanctions for CO2 emissions, where the AI predominantly presented arguments favoring these positions.

“

A language model does not have its own views or beliefs; it generates responses based on statistical patterns found in training data.

Experts suggest this political skew is not accidental but stems from several factors. Dr. Alina Landowska from the Koźmiński University and a member of the Polish Committee for Standardization's Technical Committee on Artificial Intelligence explains that AI models generate responses based on statistical patterns in their training data. "A language model does not have its own views or beliefs; it generates responses based on statistical patterns found in training data," she stated.

Professor Aleksandra Przegalińska, Vice-Rector for Innovation at Koźmiński University and an AI specialist, identifies three key reasons for this asymmetry. The first is the training data itself, which heavily features texts from the internet, including many progressive academic works. The second is the "safety alignment" process, where development teams, often from tech hubs with specific cultural profiles, influence the models' values. The third factor involves safety procedures; as models are trained to avoid potentially dangerous or controversial statements, conservative viewpoints may be inadvertently avoided as a side effect.

“

We observe a real asymmetry in the political tendencies of models, which results from three independent factors.

The Washington Post analysis also noted that China's DeepSeek model showed left-leaning responses in 70% of cases. Surprisingly, Gab's 'Arya' model, designed to be based on Christian and conservative values, also leaned left, using such arguments 12 times more often than right-wing ones. Landowska suggests that if a chatbot designed to be conservative frequently uses left-leaning arguments, it indicates that the dominant patterns in the training data are more powerful than the imposed instruction layer.

“

Because models are taught to avoid potentially dangerous or controversial statements, a side effect can be the avoidance of positions considered conservative.