🤖 AI

Winning and losing with Artificial Intelligence: What public discourse about ChatGPT tells us about how societies make sense of technological change

By analyzing 3.8 million tweets following the launch of ChatGPT, this study reveals that public discourse on AI is shaped by occupational self-interest and national cultural values, where technical professionals engage early with optimism while writing-centric roles and individualistic cultures join later with greater skepticism, driving a shift toward critical sentiment over time.

Original authors: Adrian Rauchfleisch, Joshua Philip Suarez, Nikka Marie Sales, Andreas Jungherr

Published 2026-03-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Adrian Rauchfleisch, Joshua Philip Suarez, Nikka Marie Sales, Andreas Jungherr

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the launch of ChatGPT wasn't just a tech company releasing a new product. Instead, think of it as a massive, global town square meeting that happened all at once on Twitter (now X).

This paper is like a sociologist standing in the back of that town square, watching who shows up, when they show up, and what they shout about. The researchers wanted to understand: Why do some people cheer for AI, while others boo it? And why do some people arrive at the party early, while others wait until the music is already playing?

Here is the breakdown of their findings using simple analogies:

1. The "Who" and "When": The Job Determines the Seat

The researchers looked at 3.8 million tweets from people in 117 countries. They realized that your job is like a seat assignment at this party. It dictates where you sit and how you feel about the new technology.

The "Tech-Compliment" Group (Programmers & Mathematicians):
- The Analogy: Imagine a chef who just got a new, super-smart kitchen robot that chops vegetables instantly. They are thrilled. They run to the party first, waving their hands, saying, "Look how fast this is! It makes my job easier!"
- The Finding: People whose jobs involve coding, math, or data analysis were the early adopters. They saw ChatGPT as a helpful tool (a "complement") that would boost their productivity. They spoke up immediately and were generally happy.
The "Tech-Threat" Group (Writers, Editors, & Artists):
- The Analogy: Now imagine a group of human painters and writers. They see the same robot and think, "Wait, if it can paint and write, do I still have a job?" They are hesitant. They arrive at the party later, looking around nervously, whispering, "This feels dangerous."
- The Finding: People whose jobs rely heavily on writing and content creation joined the conversation later. When they did speak, they were much more skeptical and negative. They saw ChatGPT not as a helper, but as a replacement for their skills.

2. The "Where": Culture is the Weather

The researchers also looked at the "weather" of different countries. They used a famous cultural map (Hofstede's dimensions) to see how a country's personality affects its reaction.

Individualism (The "Me" Culture):
- The Analogy: Think of a culture that values personal freedom and standing out (like the US or Western Europe). These people are like early explorers. They are the first to jump off the dock to see what's in the water.
- The Finding: People from individualistic countries showed up early. However, once they got there, they were actually more critical. Why? Because they were worried about their personal autonomy and whether this machine would take away their unique voice.
Uncertainty Avoidance (The "Safety First" Culture):
- The Analogy: Think of a culture that hates surprises and loves rules and stability (like Japan or Germany). These people are like careful drivers checking the brakes before getting in the car.
- The Finding: People from these cultures were less likely to say "Yay!" (positive). They didn't necessarily scream "No!" (negative), but they were cautious. They didn't want to jump in until they were sure the technology was safe and reliable.
Power Distance (The "Hierarchy" Culture):
- The Analogy: This is about how much a culture respects bosses and authority.
- The Finding: Surprisingly, this didn't really change how people reacted. Whether a culture was strict or loose with authority, it didn't predict if they liked or disliked ChatGPT. It was a "neutral" factor.

3. The Big Twist: The "Crowd" Changed, Not the "Opinion"

This is the most important part of the paper. If you just looked at the "average" mood of the whole town square over time, you would see a trend: The mood got more negative.

The Old Way of Thinking: You might think, "Oh, people started out loving AI, but then they realized it was bad, so they changed their minds."
The Paper's Discovery: That's wrong. The early people (programmers, individualists) never changed their minds. They stayed positive.
The Real Reason: The "average" mood got darker because new people arrived. The latecomers (writers, cautious cultures) showed up later, and they were the ones who were skeptical.

The Metaphor: Imagine a concert.

Hour 1: Only the band's biggest fans are there. They are screaming "YES!" and dancing. The vibe is 100% positive.
Hour 3: The general public arrives. They are confused, worried about the noise, and think the music is too loud.
The Result: If you measure the "average happiness" of the whole crowd, it drops. But the original fans didn't stop loving the music; the crowd just got bigger and more diverse.

The Takeaway

This paper teaches us that when we argue about new technology (like AI), we aren't just arguing about the tech itself. We are arguing about our jobs and our cultural values.

If you have a job the AI helps, you'll be an early optimist.
If you have a job the AI might replace, you'll be a later skeptic.
If your culture values risk-taking, you'll try it first.
If your culture values safety, you'll wait and watch.

The "public opinion" we see on social media isn't a single person changing their mind; it's a changing cast of characters entering the stage, each bringing their own fears and hopes with them.

1. Problem Statement

The paper addresses the challenge of understanding how societies collectively interpret and adapt to rapid technological change, specifically Artificial Intelligence (AI). While AI is often viewed through bifurcated expectations (empowerment vs. obsolescence), public discourse is frequently analyzed only at an aggregate level (e.g., overall sentiment trends).

The authors argue that aggregate sentiment trends mask critical underlying dynamics. Specifically, they seek to answer:

Who engages in public discourse regarding AI?
When do different groups engage (timing of adoption)?
How do economic self-interests (occupational skills) and cultural values shape these engagements and stances?

The study posits that shifts in public sentiment are not necessarily due to individuals changing their minds, but rather due to a shift in the composition of participants entering the discourse over time.

2. Methodology

Data Collection

Source: Twitter (now X) data collected via Stream and Academic APIs.
Timeframe: November 30, 2022 (ChatGPT launch) to February 1, 2023.
Sample: 3,793,601 tweets from 1,606,270 unique users across 117 countries.
Keywords: "chatgpt", "gpt", "chat-gpt", "chat gpt".

Variable Classification (LLM-Based)

The authors utilized OpenAI's GPT-4o-mini model to classify three key dimensions, as traditional classifiers struggled with multilingual nuance and specific context:

Stance Detection: Tweets were classified as Positive, Neutral, or Negative regarding ChatGPT. (Validated with manual review: Accuracy = 0.87, Cohen's $\kappa$ = 0.78).
Location Identification: User country was inferred from profile location fields. (Validated accuracy: ~96% for identified locations; 82.65% for holistic profile review).
Occupation Identification: User descriptions were mapped to occupational skill requirements based on the O*NET database, specifically focusing on skill levels in Writing, Programming, and Mathematics.

Cultural Metrics

Cultural context was operationalized using Hofstede's Cultural Dimensions at the country level:

Individualism (IDV)
Uncertainty Avoidance (UAI)
Power Distance (PDI)

Analytical Strategy

The study employed Multilevel Models to account for clustering effects (users nested within occupations and countries):

Multilevel Multinomial Logistic Regression: To analyze factors associated with tweet stance (Positive/Neutral/Negative).
Multilevel Cox Proportional Hazards Model: To analyze the timing of initial engagement (time elapsed from launch to first tweet).
Robustness Checks: Models were re-estimated using skill importance (vs. skill level), ordinal logistic models, and linear regression on average stance scores to ensure findings were not artifacts of model specification.

3. Key Contributions

Decomposition of Sentiment Shifts: The paper demonstrates that the observed shift toward more critical/negative sentiment over time is driven primarily by the entry of new, skeptical participants (late adopters with specific occupational/cultural profiles) rather than a change in opinion among early adopters.
Integration of Economic and Cultural Factors: It bridges the gap between economic self-interest (occupational skills) and cultural theory (Hofstede) in the context of AI discourse, showing how both simultaneously shape engagement.
Methodological Advancement: It successfully applies Large Language Models (LLMs) for multilingual stance detection and occupation classification at a massive scale (millions of tweets), overcoming limitations of traditional sentiment analysis tools.

4. Key Results

A. Occupational Skills (Economic Self-Interest)

Stance (RQ1):
- Mathematics/Programming Skills: Positively associated with positive stances and negatively associated with negative stances. These groups view AI as a complement to their work.
- Writing Skills: Negatively associated with positive stances. Writers, authors, and editors expressed higher skepticism, viewing AI as a threat to their livelihood and creative integrity.
Timing (RQ2):
- Programming/Math: These groups engaged significantly earlier. They are natural early adopters of technical tools.
- Writing: These groups engaged later. The delay is attributed to a lower subjective affinity with the technology initially and a slower realization of its impact on their specific economic fortunes.

B. Cultural Factors

Individualism (IDV):
- Timing: Users from highly individualistic countries (e.g., USA, UK, Germany, Sweden) engaged earlier.
- Stance: Contrary to some prior US-centric studies, higher individualism predicted a more negative stance in this global sample. The authors suggest individualists may perceive AI as a threat to personal autonomy and creativity.
Uncertainty Avoidance (UAI):
- Stance: High uncertainty avoidance cultures were less likely to express positive sentiment. They approached the technology with caution regarding reliability and ethical risks.
- Timing: Did not significantly predict the timing of engagement.
Power Distance (PDI):
- Showed no significant association with either stance or timing. The authors suggest opposing mechanisms (e.g., AI reinforcing hierarchy vs. democratizing access) may cancel each other out.

C. Aggregate vs. Individual Dynamics

The overall sentiment in the dataset shifted from mostly neutral/positive to more critical over time.
Crucial Finding: This shift was not caused by early adopters changing their views. Instead, the "composition of speakers" changed. As the discourse broadened, it included more writers (who are skeptical) and users from collectivist/high-uncertainty-avoidance cultures, diluting the initial optimism of the technical/individualist early adopters.

5. Significance and Implications

Reframing Public Discourse: The study challenges the view of public discourse as a deliberative process where opinions converge. Instead, it frames discourse as a strategic negotiation of meaning where actors mobilize based on pre-existing economic interests and cultural values.
Policy and Management: Understanding that different occupational groups react differently to AI is crucial for policymakers and tech firms. "One-size-fits-all" communication strategies may fail if they do not address the specific economic anxieties of writing-centric professions versus the technical optimism of coding-centric professions.
Cultural Nuance: The findings highlight that cultural dimensions like individualism do not have a uniform effect on technology acceptance globally; they interact complexly with the specific nature of the technology (e.g., generative AI's impact on autonomy).
Future Research: The paper calls for longitudinal studies to see if these initial "focusing event" reactions stabilize or shift as technology matures, and suggests using survey methods to validate social media findings, particularly for cultures with high uncertainty avoidance that may under-represent themselves on social media.

In conclusion, the paper argues that public sensemaking about AI is a reflection of entrenched societal cleavages (economic and cultural) rather than a fluid evolution of public opinion. The "winners" (technical roles, early adopters) and "losers" (writing roles, late adopters) of the AI revolution are already visible in the timing and tone of their initial public engagement.