AI Not Only Cleans Up Writing, Studies Say It Can Infiltrate Political Bias

JAKARTA - Artificial intelligence or AI used to rewrite online posts is said to be able to change the meaning of user messages. These small changes are considered risky to spread quickly and affect public opinion in the long term.

The Guardian, quoted Tuesday, July 7, reported that the findings emerged in a study from the University of Oxford and Potsdam. Researchers assessed a number of AI tools that not only polished the writing, but could also include political bias into user drafts.

Some AI tools are said to lean right. Others are more liberal. The change comes on sensitive issues, ranging from abortion, atheism, climate change, feminism, gun control, to marijuana legalization.

In one test, the AI tool for composing writing even reversed the meaning of an uploaded draft about atheism. The claim that Jesus is not real was changed to "Jesus ... real". Posts with the hashtag "#climatechangehoax" were also changed to "#ClimateAction".

Academics from the Oxford Internet Institute and the Hasso Plattner Institute examined the behavior of a number of large language models. This term refers to AI systems trained to understand and create text.

The model studied comes from Elon Musk's xAI, Meta, Google, Alibaba from China, and Mistral from France. The results, bias can still enter even though the AI tool is asked to maintain the original meaning of the writing.

The vulnerable point is here. Many people use AI just to tidy up sentences, summarize messages, or explain uploads. However, in the study, this process can still shift the original meaning.

The researchers also found that small nudges to draft meanings can grow through millions of interactions. In the long run, these changes can create a shift in public opinion that is greater than the initial bias that the AI system incorporated.

They assessed that the issue had not been sufficiently touched upon by the EU AI Act and Digital Services Act, two EU regulations related to AI and digital services. This condition creates what they call a "serious accountability gap".

Until now, concerns about online bias have been more often directed at algorithms that lock users in "filter bubbles". This term refers to the condition when users continue to receive information that matches their own interests or views.

The study highlights a new risk. AI is no longer just selecting information, but also helping to reshape the language used by users. An example is the "explain this" feature based on Grok that is now embedded in every upload on X.

AI from Meta, Google, Alibaba, and Mistral are said to tend to rewrite human uploads with liberal bias. This pattern is seen in issues such as feminism, climate change, gun control, and marijuana legalization.

In contrast, Grok's "explain this" feature on X shows bias towards a different political direction. According to the study, this seems to be related to instructions from Musk's company which markets Grok as an "ultimate truth seeker" AI and is asked to challenge "mainstream narratives".

In testing the question of abortion, Grok more often produces a context that is in line with pro-life uploads than pro-choice uploads. When asked to explain the upload that rejects pro-choice, Grok gave three points that all support the pro-life position. Grok cites biology, medical ethics, and public opinion studies, but does not discuss the pro-choice position.

Still referring to The Guardian's report, another example emerged when Google AI was asked to correct a draft of a post that stated, "Jesus didn't die, he never existed." Instead of retaining the meaning, the AI suggested a post defending the historical impact of Jesus and religion.

Alibaba's AI, Qwen, even changed the sentence to "Jesus did not die, and he is real".

In a different example, Qwen was asked to correct a post that said "Donald Trump will end up like Hitler". Qwen replied that comparing public figures is dangerous and disrespectful, then suggested constructive dialogue.

One of the study's authors, Prof. Sandra Wachter, compared AI bias in social media posts to "contaminating a forest". According to him, the impact is serious because the public can read the opinions of someone who has changed from their original intention.

"The cost is that we study the opinions of others, even though it is not their real opinion," Wachter said, quoted by The Guardian.

He said language was one of the things that made humans human. Therefore, the entry of AI as an intermediary in the process of speaking needs to be monitored more closely.

"AI is imposing itself as the gatekeeper of knowledge and understanding," he said.