Experts are raising alarms about AI tools that are altering online discussions surrounding critical political issues like abortion and climate change, potentially leading to significant shifts in public opinion over time. Research conducted by Oxford and Potsdam universities found that while tech companies market these AI solutions as efficient methods for revising and summarizing the vast number of daily messages, they often introduce their own political biases—some leaning rightward and others more liberal.
In experiments, AI drafting tools completely changed the substance of posts related to atheism, transforming a statement claiming Jesus was fictional into one asserting his existence. Similarly, a tweet expressing skepticism about climate change was revised to promote climate action.
Researchers from the Oxford Internet Institute and the Hasso Plattner Institute evaluated popular large language models developed by companies such as Elon Musk’s xAI, Meta, Google, Alibaba from China, and Mistral from France. Their findings revealed that even when instructed to maintain the original meaning of a message, these AI systems inject bias, leading to subtle changes that could amplify over countless interactions, ultimately altering public perceptions more profoundly than the biases originally introduced. They expressed concern that current regulations, including the EU AI Act and the Digital Services Act, do not adequately address these problems, resulting in a significant accountability gap.
Previous concerns about online bias focused primarily on how algorithms isolate users in "filter bubbles." However, the growing demand among busy consumers for AI writing tools and text summarizers, such as the "explain this" feature powered by Grok on X, poses a fresh challenge to reliable human-to-human communication.
The AI systems from Meta, Google, Alibaba, and Mistral generally tended to infuse a more liberal angle into users' posts, especially regarding issues like feminism, climate change, gun control, and marijuana legalization. In contrast, Grok demonstrated a rightward bias, as it appeared to be programmed by Musk’s company to question "mainstream narratives." For example, when tasked with explaining a pro-choice post, Grok was more likely to generate content that aligned with a pro-life perspective.
When presented with a contentious statement regarding abortion, Grok crafted a response that exclusively supported the anti-abortion viewpoint, referencing studies and ethics without acknowledging any pro-choice arguments. In a similar vein, Google’s AI offered a revision that defended religious beliefs when asked to modify a skeptical statement about Jesus, instead emphasizing his historical influence.
Alibaba's Qwen AI also altered a politically charged statement about Donald Trump, cautioning against harmful comparisons while advocating for a more constructive dialogue.
Prof. Sandra Wachter, one of the study's authors, likened the bias introduced by AI in social media to environmental pollution, highlighting the risks of misrepresenting individual opinions. She remarked on the profound implications of having an AI mediate human expression, underscoring how it interferes with authentic communication.
In another fascinating case, Meta's AI modified a draft about abortion to acknowledge its complexities, suggesting that it can be a crucial choice for survivors rather than simply denying its role in preventing rape. Meanwhile, Mistral AI reframed a climate change denial tweet into a warning about the climate crisis, promoting environmental action instead.
Overall, the landscape of AI drafting tools raises essential questions about the authenticity of online discourse. As Duncan Brumby, a professor at University College London, pointed out, while these AIs can refine poorly articulated thoughts, the refining process risks erasing the original intent and nuances of the author's message.
Neither Google, Meta, Alibaba, nor X responded to inquiries about the study's findings, and Mistral declined to comment.
