Reddit’s Four-Year AI Debate Reveals a Public Torn Between Trust and Distrust
Generative artificial intelligence has become one of the most rapidly adopted technologies in modern history, but a large-scale analysis of online conversations suggests that public opinion has not settled into either enthusiasm or rejection. Instead, people appear to be negotiating their relationship with AI in real time, judging systems such as ChatGPT, Claude and LLaMA according to whether they produce useful, accurate and dependable results. A new study from Drexel University, published in Transactions of the Association for Computational Linguistics, analyzed more than 230,000 Reddit posts written between November 2022 and June 2025. The researchers found that trust in generative AI appeared in approximately 31% of the posts, while distrust appeared in 26%. About 41% expressed neither position, and roughly 1% contained both trust and distrust.
The results provide one of the largest longitudinal examinations to date of how people discuss confidence in generative AI. Rather than showing a steady movement toward either acceptance or rejection, the data reveal a persistent balance between the two attitudes. Trust maintained a modest lead across most of the study period, but distrust temporarily overtook it during certain months. These shifts occurred despite the rapid release of new models, features and AI-powered products. The pattern suggests that public attitudes may be more stable and divided than some recent surveys indicate, including reports that people are using AI more frequently while becoming less willing to trust its outputs.
The research team, led by Shadi Rezapour, an assistant professor in Drexel University’s Nick Howley College of Engineering and Computing, examined discussions from 39 AI-related Reddit communities. The posts covered widely discussed systems and developments, including ChatGPT, OpenAI’s GPT-4, Meta’s LLaMA models and Anthropic’s Claude. Using computational language-analysis methods, the researchers identified statements expressing trust or distrust and classified the reasons behind those views. In technical terms, the analysis treated trust not simply as positive sentiment, but as an expectation that an AI system would behave reliably, demonstrate competence or operate with integrity. Distrust was defined as active skepticism or concern about reliability, competence or ethical consequences, rather than merely a lack of confidence.
The distinction is important because people can be uncertain about an AI system without actively distrusting it. A user who says that a chatbot is “sometimes useful” may express neither clear trust nor clear distrust, while someone who warns that the system invents sources or produces dangerous medical advice is expressing active distrust. The researchers also categorized the apparent identity or professional background of commenters into 10 groups, including AI users, software developers, academics, technology professionals, business executives, journalists, educators, artists, ethicists and members of the general public. These categories were inferred from self-identifying information in the posts, allowing the team to compare how different communities described their experiences with AI.
The strongest differences appeared across professional and social groups. Posts associated with business leaders, academics, software developers and technology professionals expressed trust more frequently than distrust. The general public, AI ethicists and media professionals showed the opposite tendency, with distrust appearing more often. The largest category, generative AI users, displayed a relatively even balance between positive and negative judgments. Educators and other knowledge workers also remained divided. These differences may reflect the tasks each group asks AI to perform. Developers and technical professionals may be more likely to use systems for coding, summarization or experimentation, while journalists, educators and ethicists may focus more closely on errors, accountability, bias and the consequences of unreliable information.
Across nearly all groups, personal experience was the most common foundation for both trust and distrust. People tended to form opinions after testing a system themselves, observing its responses or incorporating it into work and everyday tasks. Posts expressing trust often described AI as competent, convenient or surprisingly effective. Distrust was frequently linked to hallucinations, incorrect answers, inconsistent reasoning and poor performance on specialized questions. The researchers found that users generally discussed whether AI “worked” before considering whether it was transparent, fair or morally acceptable. This emphasis on performance indicates that trust in generative AI is often operational: people trust a system when it repeatedly delivers acceptable results for a particular purpose.
“Competence, reliability and familiarity were central to expressions of trust, while unreliability and incompetence were major sources of distrust,” said Aria Pessianzadeh, a doctoral candidate at Drexel and the study’s lead author. Ethical and value-based concerns were present, but they were less prominent than practical judgments about accuracy and usefulness. This finding has technical implications for the design of AI systems. Improvements in benchmark scores may not automatically translate into public trust if users continue to encounter failures in ordinary situations. Conversely, a system may earn confidence in a narrow task even while users remain concerned about privacy, bias, labor disruption or the environmental cost of operating large models.
The study also detected changes in online attitudes surrounding major technology releases and announcements. Trust increased modestly around the public availability of GPT-4 and LLaMA 2, developments that demonstrated improved capabilities to millions of users and developers. Distrust rose around OpenAI’s Dev Day in late 2023, when a major wave of product announcements generated intense public discussion. Such fluctuations suggest that trust is sensitive not only to direct interactions with AI, but also to news events, marketing claims and expectations about what new systems will be able to do. However, none of these episodes fundamentally changed the long-term balance. Trust continued to lead by a relatively small margin, while neither trust nor distrust dominated the broader conversation.
The researchers caution that Reddit cannot be treated as a perfect mirror of public opinion. Its users are not demographically representative of the entire population, and people who post about AI may be more technically engaged, enthusiastic or frustrated than people who use these systems quietly. Online discussions may also encourage immediate reactions and strong opinions, while ethical concerns that require extended explanation may be expressed less directly. In addition, the analysis focused on English-language posts and inferred user categories from self-descriptions, creating possible classification errors. Even with those limitations, the scale and time span of the dataset offer a detailed record of how attitudes formed during the first major phase of consumer generative AI adoption.
The findings could help regulators, educators and AI developers move beyond the assumption that users are either confident in AI or opposed to it. Shadi Rezapour said responsible governance should account for both groups and connect abstract principles such as transparency, bias reduction and accountability to the everyday experiences that shape confidence. Future studies will need to examine other languages, platforms and populations, while tracking how attitudes change as AI becomes embedded in search engines, workplaces, classrooms and personal devices. For now, the Reddit record portrays generative AI as neither an unquestioned breakthrough nor an outright failure. It is a technology people are repeatedly testing, accepting when it performs well and challenging when its limitations become impossible to ignore.
Subject of Research: Trust and distrust in generative artificial intelligence as expressed in Reddit discussions
Article Title: In Generative AI We (Dis)Trust? Computational Analysis of Trust and Distrust in Reddit Discussions
News Publication Date: 1 July 2026
Web References: https://direct.mit.edu/tacl/article/doi/10.1162/TACL.a.744/137431
References: Transactions of the Association for Computational Linguistics; DOI: 10.1162/TACL.a.744/137431
Keywords
Generative AI, artificial intelligence, AI trust, AI distrust, ChatGPT, Reddit analysis, computational linguistics, public opinion, AI reliability, responsible AI governance
Tags: AI acceptance or rejection trendsChatGPT and LLaMA user sentimentsDrexel University AI researchGenerative AI public opinion analysisimpact of AI on social media discussionslarge-scale analysis of AI-related Reddit postslongitudinal study of AI perceptiononline conversations about AI reliabilityperceptions of AI usefulness and accuracypublic attitudes towards AI systemsReddit AI debatetrust and distrust in artificial intelligence
