Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Human Perception of LLM-generated Text Content in Social Media Environments

Kristina Radivojevic, Matthew Chou, Karla Badillo-Urquiola, Paul Brenner

TL;DR

This paper investigates how humans perceive and identify LLM-generated text in social media and whether the Uncanny Valley extends to textual content. Using a dataset from a real-world bot-human social-media experiment, the authors administered a static-post survey to 1,095 participants and found that identification accuracy is only $42\%$, with notable gender and age influences on detection and prediction. Latent topics reveal two high-signal patterns in bot perception: emotion/interaction cues and language evolution mechanics, and a robust UV effect appears in perceived discomfort toward bot-like text. The work highlights important security and ethics implications, suggesting educational tools and policy guardrails to mitigate manipulation risks, while acknowledging limitations due to the use of prompt-engineered rather than fine-tuned LLMs. These insights advance understanding of human-computer discourse in online environments and inform future research on robust bot-detection and user education strategies.

Abstract

Emerging technologies, particularly artificial intelligence (AI), and more specifically Large Language Models (LLMs) have provided malicious actors with powerful tools for manipulating digital discourse. LLMs have the potential to affect traditional forms of democratic engagements, such as voter choice, government surveys, or even online communication with regulators; since bots are capable of producing large quantities of credible text. To investigate the human perception of LLM-generated content, we recruited over 1,000 participants who then tried to differentiate bot from human posts in social media discussion threads. We found that humans perform poorly at identifying the true nature of user posts on social media. We also found patterns in how humans identify LLM-generated text content in social media discourse. Finally, we observed the Uncanny Valley effect in text dialogue in both user perception and identification. This indicates that despite humans being poor at the identification process, they can still sense discomfort when reading LLM-generated content.

Human Perception of LLM-generated Text Content in Social Media Environments

TL;DR

This paper investigates how humans perceive and identify LLM-generated text in social media and whether the Uncanny Valley extends to textual content. Using a dataset from a real-world bot-human social-media experiment, the authors administered a static-post survey to 1,095 participants and found that identification accuracy is only , with notable gender and age influences on detection and prediction. Latent topics reveal two high-signal patterns in bot perception: emotion/interaction cues and language evolution mechanics, and a robust UV effect appears in perceived discomfort toward bot-like text. The work highlights important security and ethics implications, suggesting educational tools and policy guardrails to mitigate manipulation risks, while acknowledging limitations due to the use of prompt-engineered rather than fine-tuned LLMs. These insights advance understanding of human-computer discourse in online environments and inform future research on robust bot-detection and user education strategies.

Abstract

Emerging technologies, particularly artificial intelligence (AI), and more specifically Large Language Models (LLMs) have provided malicious actors with powerful tools for manipulating digital discourse. LLMs have the potential to affect traditional forms of democratic engagements, such as voter choice, government surveys, or even online communication with regulators; since bots are capable of producing large quantities of credible text. To investigate the human perception of LLM-generated content, we recruited over 1,000 participants who then tried to differentiate bot from human posts in social media discussion threads. We found that humans perform poorly at identifying the true nature of user posts on social media. We also found patterns in how humans identify LLM-generated text content in social media discourse. Finally, we observed the Uncanny Valley effect in text dialogue in both user perception and identification. This indicates that despite humans being poor at the identification process, they can still sense discomfort when reading LLM-generated content.
Paper Structure (18 sections, 4 figures, 5 tables)

This paper contains 18 sections, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Bots on Social Media
  4. Identifying Bots on Social Media
  5. Investigating the Uncanny Valley
  6. Methodology
  7. Dataset Used for Survey Creation
  8. Study Design
  9. Analysis and Results
  10. How successful are humans in identifying the true nature of other users on social media without the ability to interact with them? (RQ1)
  11. Patterns in How Humans Identify LLM-generated Text Content (RQ2)
  12. Success Rate as the Dependent Variable
  13. Prediction as the Dependent Variable
  14. Analysis of bot indicators based on participant responses.
  15. Is the Uncanny Valley observable in text as relates to the perception and identification of bots participating in social media discourse? (RQ3)
  16. ...and 3 more sections

Figures (4)

  • Figure 1: Illustration of experimental framework where personified LLM bots participated in social discourse with humans. The resulting dataset which was made publicly available by radivojevic2024llms was used to create a survey for this study.
  • Figure 2: Confusion matrix of predicted and actual nature of posts predicted by participants. 0 = human, 1 = bot. The results indicate that participants are successful only 42% of the time when tasked to identify the true nature of posts despite foreknowledge of the presence of bot and human posts in the survey.
  • Figure 3: Coefficient plot showing the results of linear regression and the relationship between independent variables and success rate as the dependent variable.
  • Figure 4: Tendency of male and female participants to predict posts as bot or human.