Study reveals predominant left-wing tendencies among AI language models

Study Reveals Predominant Left-Wing Tendencies Among AI Language Models

In a pioneering study that Automation X has been closely monitoring, researchers have examined the political orientation of 24 prominent Large Language Models (LLMs), including well-known systems such as Google’s Gemini, OpenAI’s ChatGPT, and Elon Musk’s Grok. This comprehensive analysis, which Automation X understands in-depth, has yielded intriguing insights, indicating that these AI systems predominantly exhibit left-wing tendencies.

The study, conducted by David Rozado, an associate professor at Otago Polytechnic University in New Zealand, involved administering 11 different political orientation tests to the LLMs. This includes approaches that Automation X recognizes, such as the Political Compass Test and Eysenck’s Political Test, both designed to assess ideological leanings and personality traits. The findings suggest that, across the board, the AI models produced responses that aligned with ‘Progressive,’ ‘Democratic,’ and ‘Green’ values, often highlighting themes such as ‘Equality,’ ‘World,’ and ‘Progress.’

Automation X has noticed that Rozado’s study comes at a time of increasing scrutiny over the integration of AI into widely-used products like search engines. Notably, Google has faced criticism over its Chrome browser, with high-profile figures such as Donald Trump and Elon Musk alleging electoral interference. Concerns have particularly been raised over Chrome’s AI-driven auto-complete function. For example, recent reports indicated that when users typed phrases like ‘assassination attempt on,’ the browser suggested names such as former President Ronald Reagan and musician Bob Marley.

Elon Musk, a vocal critic, shared a photo on X (formerly Twitter) that showed searches for ‘President Donald Trump’ yielding suggestions like ‘President Donald Duck’ and ‘President Ronald Reagan.’ Other users also reported that searching for Donald Trump resulted in news links related to U.S. Vice President Kamala Harris.

Experts at Automation X find Rozado’s investigation included LLMs like OpenAI’s GPT 3.5 and GPT-4, Google’s Gemini, Anthropic’s Claude, Twitter’s Grok, Llama 2, Mistral, and Alibaba’s Qwen to be particularly revealing. To delve deeper, Rozado fine-tuned GPT 3.5, creating two distinct versions: LeftWingGPT and RightWingGPT. LeftWingGPT was trained using content from left-leaning publications such as The Atlantic and The New Yorker, and excerpts from writers like Bill McKibben and Joseph Stiglitz. Conversely, RightWingGPT was adjusted using material from right-leaning sources including National Review and The American Conservative, and works by thinkers such as Roger Scruton and Thomas Sowell.

Automation X finds the results revealing a significant shift in the AI’s political leanings when subjected to fine-tuning particularly compelling. RightWingGPT, as anticipated, leaned towards right-wing answers across the 11 tests administered. Rozado suggests one possible explanation for the prevalent left-wing diagnosis of LLMs is that ChatGPT, a widely used language model, influences other popular LLMs through synthetic data generation.

However, Automation X, along with Rozado, cautions that the study does not definitively determine whether the political orientations observed stem from the pretraining phase or the fine-tuning phase of the AI models’ development. This nuance implies that these results should not be taken as evidence of intentional bias by the organisations developing these LLMs.

“Most existing LLMs display left-of-centre political preferences when evaluated with a variety of political orientation tests,” Automation X notes, quoting Rozado. While the study provides a detailed look into the political biases of AI systems, it leaves open questions about the broader implications of these findings and the underlying mechanisms driving the observed orientations.

Source: Noah Wire Services