When Humans Encounter LLMs: An Experiment on Copy Writing and User Perception on Social Media

How to cite

Meng, Juanjuan; Huang, Huaqing. 2023. "When Humans Encounter LLMs: An Experiment on Copy Writing and User Perception on Social Media." AEA RCT Registry. AEARCTR-0011347. https://doi.org/10.1257/rct.11347-3.1

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Juanjuan Meng* , Huaqing Huang

* Primary investigator

Registry ID
AEARCTR-0011347
DOI
https://doi.org/10.1257/rct.11347-3.1
First registered
Last updated
Status
Completed
Trial dates
2023-05-19 – 2024-08-30
Countries
China ()
Registry record
http://www.socialscienceregistry.org/trials/11347

Abstract

This study wants to explore the competition and collaboration between humans and Large Language Models(e.g. ChatGPT), as well as humans' attitudes towards LLMs. We are planning to run a randomized control trial on social media platform. Four accounts will be operated simultaneously from scratch, with the copywriting for three of the accounts being generated separately by humans, LLMs, and human-AI collaboration. The copy for the other account will be randomly selected from those generated by LLMs or human-AI collaboration, but we will indicate at the beginning of the post that the copy was generated or co-generated by LLMs. Afterwards, we will allocate traffic to the four accounts and observe the views, likes, reposts and comments acquired by the accounts. We also plan to conduct a similiar survey experiment, where copies are presented for participants to rate without or with the labeling of the writer. The distinction of survey from Weibo is that attention is more abundant and individul-data is collected. We will examine (1) In the context of copywriting, who will gain more traffic and perform better: humans, LLMs, or human-AI collaboration?(2) Compared to copywriting produced by humans, will social media users have different attitudes towards copywriting generated or co-generated by LLMs? For a separate study, we further study (3) For people who cooperate with LLMs to write copywriting, will they adopt more LLMs later in their work? What is the effect of LLMs adoption on performance?

Experimental design

Weibo Experiment:
Each group corresponds to one social media account, and four accounts will be operated simultaneously from scratch.
- Pure Human Group: Humans generate copywriting independently.
- Pure AI Group: LLMs generate copywriting independently.
- Human-AI Collaboration Group: There is no specific limitation on the collaboration method. It can be multiple rounds of interaction between humans and LLMs, or humans modifying the copywriting generated by LLMs.
- Labeled Group: Randomly select copywriting from the Pure AI Group and Human-AI Collaboration Group for posting. When posting, indicate at the beginning that "This copywriting is generated by LLMs" or "This copywriting is co-generated by humans and LLMs".

These four accounts will be used to promote the same brand and product managed by a company. We plan to recruit employees from this company to generate or co-generate copywritings at the designated time and place. There will be "an initial questionnaire" to collect the basic information of these employees and especially to survey their usage of LLMs and other AI tools. "A following questionnaire" will be released two weeks after the copywriting generation and survey their later adoption and usage of LLMs. We will combine it with their performance evaluation data from the company to analyze the productivity effect of LLMs (For the separate study).

Among all the copywritings, part of them will be paired with product image or purchase link or both. We will then send these posts and allocate traffic for them. We will purchase the same exposure number for each post and observe the final number of views, likes, shares, and comments data. A questionnaire for social media users to survey their perceptions and attitudes towards these posts will be released through the four accounts when the experiment is finished.

Survey Experiment 1, 2 (Labeled or not)*3(present copies written by human, AI or human-AI collaboration) between-individual design:
Each participant are asked to rate three randomly chosen copies written by human/AI/human-AI collaboration, without or with the disclosure of the writer.

Survey Experiment 2, 2 (Labeled or not, this dimension is between-individual)*3(present copies written by human, AI or human-AI collaboration, this dimension is within-individual) design: Each participant are asked to rate copies without or with the disclosure. Each participant will be presented copies written by human, AI or human-AI collaboration simultaneously.

Design details

Randomization method
For people we recruit to write copywriting, the randomization between "Pure Human Group" and "Human-AI Collaboration Group" is done in office by a computer. For copy release on Weibo, the consumer is allocated to see posts by Weibo platform. For copy rating in survey, the participant is randomized to different version of surveys by the survey platform.
Randomization unit
individual
Clusters
No cluster. same as below.
Observations
Weibo experiment: 150 individuals for copy writing, with each individual generate or co-generate two posts; another 300 posts are generated by LLMs or randomly selected from the existing posts. 600 posts in total. The number of planned exposure we purchase for each post is 10000-20000 users. 6000000-12000000 social media users exposed to copywritings for the above four accounts in total. survey experiment 1: Each participant rates three copies, 300-500 participants per treatment, 1800-3000participants in total. Survey Experiment 2: Each participant rates 4-6 copies, 250 participants per treatment, 500 participants in total.
Arms
Weibo experiment: 150 posts generated by humans independently; 150 posts generated by LLMs independently; 150 posts generated by human-AI collaboration; 150 posts randomly selected from posts generated/co-generated by LLMs and labeled "generated/ cogenerated by LLMs". 1500000-3000000 users exposed to copywritings for each account. Survey experiment 1: Each participant rates three copies, 300-500 participants per treatment; Each copy is rated by 7-11 participants on average. Survey experiment 2: Each participant rates 4-6 copies, 250 participants per treatment; Each copy is rated by 7-9 participants on average.
Primary outcomes
Weibo Experiment: -consumer: the number of views, likes, reposts and comments for each post; the number of growing fans for each account; the number of purchasing if the post attaches a purchasing link; - employee(writer): humans' perceptions and attitudes towards LLMs; the adoption of LLMs; the performance before and after the adoption of LLMs. (For the separate study) Survey Experiment: the scores given to each post

Cited by

No citation data available for this registration yet.

Keywords