Users Favor LLM-Generated Content — Until They Know It’s AI

Users Favor LLM-Generated Content — Until They Know It’s AI

The Future of Content Writing

Amid all the LLM releases, be it OpenAI GPT-4o image generation, DeepSeek V3–0324, Gemini 2.5 Pro, a very interesting paper is published stating that

Users love AI content, until they don’t know it AI

Paper link:

Users Favor LLM-Generated Content — Until They Know It’s AI

In the paper, the team did a group study on a number of users in different subsets and came to a conclusion that Users actual love AI content more then human written ones, but once they have a prior knowledge of the source, they develop a bias and prefer human written ones.

Experiment design

The entire procedure followed a number of steps:

1. Question Selection:

• The researchers gathered a diverse set of popular questions from platforms like Quora and Stack Overflow. These questions covered five broad scientific areas: Physical Sciences & Engineering, Life Sciences, Health Sciences, Social Sciences, and Humanities. The questions were chosen to appeal to a broad audience with varying backgrounds. Some sample questions

2. Response Generation:

• Each of the selected questions was answered by four different Large Language Models (LLMs): ChatGPT, Claude, Gemini, and Llama.
• Additionally, a human also answered each question. This provided a comparison between AI and human responses.

3. Survey Design and Implementation:

• The researchers created a survey where participants were randomly presented with five questions, each accompanied by two responses: one from an LLM and one from a human.
• A crucial element of the experimental design was randomization: half of the participants were informed of the response origin (whether it was AI or human), while the other half were not. This allowed the researchers to study the impact of source disclosure on participant preferences.
• Demographic information (age, gender, education, programming skills) was also collected from the participants.
• The survey was distributed through Prolific, an online platform that helps researchers connect with participants. It was also available via a Telegram bot and a web application for broader reach.

4. Data Analysis:

• Exploratory analysis: The researchers performed graphical analysis to explore the distribution of chosen answers based on participant characteristics (gender and programming skills) and knowledge of the response source.
• Regression analysis: A logistic regression analysis was conducted to determine the influence of knowing the response source (along with other demographic variables) on a participant’s choice between the human and AI answers.

Results

We will be discussing these two major charts from the paper

Chart 1: Gender and Source Awareness

This chart compares the share of chosen answers between human and model responses, broken down by gender (female/male) and whether the respondent was aware of the answer’s source (aware/not aware).

• Key Finding: Regardless of gender, a significantly larger proportion of chosen answers came from the model compared to human responses, particularly when respondents were not aware of the source.
• When awareness of the source was present, the model still holds a larger share but the difference is less dramatic. There is little to no significant difference between male and female preferences in terms of choosing either human or model answers.

Chart 2: Programming Skills and Source Awareness

This chart follows the same structure as Chart 1, but instead of gender, it examines the impact of programming skills (yes/no) on answer choice.

• Key Finding: Similar to Chart 1, a significant majority of chosen answers were model-generated, especially when source awareness was absent. However, unlike Chart 1, having programming skills slightly increases the preference for model answers when the source is known.

Key Findings:

• Awareness of Source Impacts Choice: When participants were unaware of the source, 29% chose the human answer. When aware, this increased to 33%. This suggests that knowing the source slightly increases preference for human-generated answers.
• Even AI Model Distribution: The distribution of the four AI models is relatively consistent across both groups (aware and unaware), with each model being used for roughly 24–26% of the questions in both scenarios.
• Consistent Question Field Distribution: The distribution of questions across the five scientific fields is relatively consistent in both the “aware” and “unaware” groups, indicating a balanced representation of diverse topics. No particular field showed a significantly larger proportion of either human or AI-generated responses.

That’s about the paper,

But we have a couple of big questions to answer in the hindsight.

Why humans preferred AI generated content in the 1st place when source is unknown?

I, myself being a content creator for some years now, I know the reason:

1. Given it’s vast knowledge that is not practical for a normal human, it has more words/ideas to add compared to a human.
2. Humans are usually more biased because of lack of knowledge. LLMs, on the other hand, are expected to be less biased and give a more neutral/right answer.
3. When I write blogs, I also struggle to add in filler text at multiple places. For AI, it’s not the case.
4. AI generated texts are consistent throughout the paragraph, but this maynot be the case with Human written text.

Now the more important question,

Why, once the source is known, humans prefer non AI outputs?

Many reasons for this as well

1. Humans by now know AI may hallucinate. Whatever it outputs, may have errors. Once you know a human wrote it, you know the answer is validated and hence preferred.
2. Ultimately, human-written content, with its inherent imperfections and individuality, often fosters a stronger connection with the reader than perfectly polished AI-generated text. This personal touch is highly valued by many readers.
3. Human-written content is generally perceived as more trustworthy and credible, especially in areas where accuracy and expertise are crucial.
4. Knowing that content is AI-generated can raise concerns about authenticity and originality. Readers may value the unique perspective and personal voice that only a human writer can provide.

Concluding,

While users may gravitate toward AI-generated content for its efficiency and objectivity, the awareness of its origin often leads to a preference for human-written work due to the value placed on trust and authenticity. As AI continues to advance, these insights highlight the ongoing tension between convenience and human connection.

And what are your thoughts on this blog? Is it good or bad? And if I tell you it’s AI-generated/Human-generated, will your preference change? Funny times ahead.

Users Favor LLM-Generated Content — Until They Know It’s AI was originally published in Data Science in your pocket on Medium, where people are continuing the conversation by highlighting and responding to this story.