The Impact of Large Language Models on Content Quality in Social Media

The increasing availability of large language models (LLMs) is transforming how users create and share content on social media platforms. Beyond enabling text generation, LLMs introduce a new paradigm in which content is deliberately optimized for engagement through algorithmically suggested phrasing, structure, and tone. This paper investigates the emerging shift from authentic self-expression toward engagement-driven optimization in LLM-assisted social media use. It examines whether and how LLM-generated or LLM-assisted posts systematically outperform human-authored content in engagement metrics and at what cost to informational quality, diversity, and authenticity. Using a mixed-methods approach, controlled experiments with human participants are combined with large-scale analysis of social media posts to compare organic and LLM-optimized content. Differences in engagement outcomes (e.g., likes, shares, comments), linguistic features, and perceived credibility and informativeness are evaluated. The findings suggest that while LLM-assisted content consistently increases short-term engagement, it tends to reduce informational depth and perceived authenticity while exhibiting changes in stylistic characteristics associated with engagement-oriented optimization. This creates a potential feedback loop in which users increasingly rely on optimization strategies that privilege attention over substance. The findings suggest that widespread adoption of LLM-driven optimization could contribute to changes in the dynamics of the social media attention economy. Future research is needed to determine whether these effects emerge at scale and persist over longer periods of platform use. Implications are discussed for platform design, content moderation, and the future of human–AI co-creation in digital communication.