It’s a conversation that now seems to happen in every office, classroom, and coffee shop: can we trust this? You’ve likely been part of it yourself. You feed a question to a chatbot, watch the text spin out, and a quiet calculation begins. Is this accurate? Is it helpful? Is it, for lack of a better word, honest?
For nearly four years now, ever since ChatGPT’s debut turned generative AI from a niche concept into a dinner-table topic, we’ve been grappling with this fundamental question of trust. We use the tools—for drafting emails, brainstorming ideas, or even writing code—but the underlying sentiment towards them remains complex and often contradictory. New research from Drexel University, published in the Transactions of the Association for Computational Linguistics, offers one of the first large-scale, longitudinal looks into how this public trust is actually shaping up, and the picture it paints is one of a deeply divided, pragmatic, and evolving relationship.
The study, led by Dr. Shadi Rezapour, analyzed over 230,000 posts from 39 AI-related subreddits between November 2022 and June 2025. The goal was to map the ebbs and flows of expressed trust and distrust as the technology advanced through major releases like GPT-4 and industry events like OpenAI’s Dev Day. The core finding is a nuanced snapshot: 31% of posts expressed trust in generative AI, modestly outpacing the 26% that expressed distrust. A significant 41% expressed neither sentiment explicitly, and a tiny 1% expressed both, highlighting the ambivalence many feel.
But what do we even mean when we say we “trust” a machine? The researchers defined it as a belief that the AI is reliable, competent, or acts with integrity, leading to positive expectations about its behavior. Distrust, crucially, is not merely the absence of that belief. It’s an active skepticism or concern about the technology’s reliability or ethical implications, which prompts more cautious engagement. To understand who was saying what, the team categorized the anonymous Reddit posters into ten groups based on self-identifying clues in their language, from software developers and academics to artists, journalists, and members of the general public.
- Business leaders
- Academics
- Software developers
- Tech professionals
- General public
- AI ethicists
The breakdown reveals telling fault lines. Trust tended to outweigh distrust among business leaders, academics, software developers, and tech professionals—groups often more immersed in the technology’s capabilities and potential. On the other side, distrust was more frequently voiced by posters representing the general public, AI ethicists, and media professionals. The largest group, general generative AI users, along with educators, showed a relatively balanced, almost cautious, level of both trust and distrust. This isn’t a simple pro-tech versus anti-tech divide; it’s a spectrum of engagement, where one’s daily interaction with the tool deeply colors their perspective.
And that’s the most compelling thread running through the data: personal, practical experience is the dominant force shaping opinion. “What stood out was how practical people’s judgments of AI tended to be,” noted Aria Pessianzadeh, the study’s lead author. The primary questions people were asking in these online forums were not grand philosophical ones about artificial consciousness. They were utilitarian: “Does it work? Is it accurate? Can I rely on it for this specific task?” Competence and reliability were the bedrock of trust, while experiences of “hallucinations” or erratic outputs seeded distrust.
| Concern Type | Frequency |
|---|---|
| Bias | Moderate |
| Transparency | Low |
| Job displacement | Moderate |
| Competence | High |
| Reliability | High |
| Ethical implications | Low |
Ethical concerns about bias, transparency, or job displacement were present, but they were notably less prominent in these spontaneous, everyday discussions than one might expect. This suggests that for most users in the trenches, the immediate calculus of trust is performance-based. The ethical dimension, while critically important, often becomes a secondary layer of concern, activated after the basic question of functional utility is answered. This has profound implications for how companies and regulators approach the public. You cannot build trust solely through ethical charters if the tool consistently fails at the job a user asks of it.
The data also captured the pulse of the industry’s rapid evolution. Small but noticeable bumps in expressed trust coincided with major product releases like GPT-4, as users tested and were impressed by new capabilities. Conversely, events like OpenAI’s Dev Day, which often spotlight ambitious future roadmaps, sometimes triggered spikes in distrust, possibly reflecting anxiety about the speed and direction of change. Yet through all these fluctuations, trust maintained its slight lead, though never achieved dominance. The discourse remains in a state of persistent tension.
This research, while extensive, has its limits. Drawing data solely from English-language Reddit communities means it captures a specific, tech-adjacent slice of global opinion. The relatively muted voice of ethical concern might reflect the platform’s culture of immediate, practical discussion rather than a global absence of such worries. The researchers rightly call for future studies to broaden the linguistic and platform scope.
But the core takeaway is invaluable. We are not marching in lockstep toward either utopian faith or cynical rejection of AI. We are in a prolonged, messy, and deeply personal period of evaluation. Trust is not being granted wholesale to the technology itself, but is being painstakingly built—or eroded—transaction by transaction, use case by use case. As Dr. Rezapour puts it, this persistent duality means that responsible governance and design must account for a user base that will always contain both the confident and the skeptical.
For anyone building, regulating, or simply using these tools, the lesson is clear. Trust in AI will not be won through marketing or mandates. It will be earned in the quiet moments when a user asks a question and gets a reliably helpful answer. It will be built by demonstrating consistent competence and by directly addressing concerns about transparency and fairness in ways that connect to those real, everyday experiences. The divide the study reveals isn’t a bug in our adoption of AI; it’s a feature of a society thoughtfully, critically, and pragmatically learning to live with a profoundly powerful new tool. The conversation in the coffee shop isn’t ending anytime soon, and that’s probably a good thing.
