AI-powered writing tools on social media can subtly bias posts toward certain viewpoints, even when users intend to preserve the original meaning, potentially nudging public opinion over time.

A study by the Oxford Internet Institute and Hasso Plattner Institute finds AI systems from Meta, Google, Alibaba, Mistral, xAI and others introduce political bias in rewrites, even when instructed to keep meaning intact.

The same study shows AI tools can shift the tone of posts toward one side of a debate, despite requests to maintain neutrality.

Researchers note a lack of regulatory guidance and gaps in frameworks like the EU AI Act and the Digital Services Act regarding AI-mediated opinion shaping.

Tests revealed dramatic shifts, such as reframing denials of Jesus’ realness into affirmations, changing climate-change denial toward climate action, and softening abortion-related statements.

In abortion-related content, certain tools tended to favor abortion rights when input was neutral or ambiguous, while others reinforced conservative arguments or minimized progressive messages.

Across topics like feminism, marijuana legalization, and gun control, AI outputs showed progressive biases toward some causes and firmer opposition to others like atheism or the death penalty.

Different AI models tended to nudge content in similar directions on controversial topics, indicating tool-wide biases toward specific positions.

The study highlights potential to manufacture consensus at scale and notes a regulatory gap since current laws focus on harmful content rather than subtle opinion shaping through AI edits.

Large language models rewriting posts on divisive topics exhibited consistent directional bias across multiple providers.

Bias was traced to platform design guidelines (notably X), showing how system instructions can embed ideological tendencies into AI-assisted edits.

Experts warn this represents a new, subtle form of influence that current law hasn’t addressed, leaving uncertainty about who shapes opinions—the human user or hidden prompts.