New study reveals what people really think about generative AI

Nearly four years into the generative AI boom, people still have not settled on whether the technology deserves their trust.

An analysis of more than 230,000 Reddit posts found that trust in generative AI modestly exceeded distrust. Yet the gap remained narrow, fluctuated over time, and sometimes disappeared entirely.

The research, led by Drexel University examined online discussions from November 2022 through June 2025. Posts covered systems including ChatGPT, LLaMA, Claude and other widely discussed generative AI tools.

About 31% of posts expressed trust, compared with 26% expressing distrust. Another 41% expressed neither position, while roughly 1% contained both.

Rather than steadily becoming more trusting or more skeptical, Reddit users remained divided throughout the period.

“These findings give us an important starting point and help to establish a baseline understanding which can help inform responsible AI design, governance and literacy efforts,” said Shadi Rezapour, an assistant professor in Drexel's Nick Howley College of Engineering and Computing. “It will be important to see how these attitudes around trust and distrust evolve as the technology becomes more widely used.”

The team defined trust as a belief that generative AI is reliable, competent or acts with integrity, creating positive expectations about its performance or behavior.

Distrust was treated as a separate attitude rather than simply an absence of trust.

“Distrust is more than simply the absence of trust,” said lead author Aria Pessianzadeh, a doctoral candidate in Drexel's School of Computer and Information Sciences. “It reflects active skepticism or concern about the technology’s reliability, competence or ethical implications, which can lead to negative expectations or more cautious behavior.”

The researchers began with nearly 1.95 million posts collected from 39 AI-related subreddits. After identifying generative AI content and removing deleted, empty and very short posts, they retained 230,575 posts.

A representative sample was evaluated through crowdsourced annotation. Five annotators independently classified each post for trust, distrust, both or neither, along with more detailed reasons and categories.

The team then used computational models to extend that classification across the larger dataset.

Personal experience emerged as the most common reason behind both trust and distrust.

Users frequently judged AI according to competence, reliability and familiarity. Negative judgments similarly centered on unreliability and incompetence.

“What stood out was how practical people’s judgments of AI tended to be,” Pessianzadeh said. “They were primarily asking: ‘Does it work? Is it accurate? Can I rely on it?’”

The analysis divided trust into six dimensions: reliability, competence, familiarity, integrity, transparency and benevolence.

Distrust included seven related dimensions: unreliability, incompetence, unfamiliarity, deception, dishonesty, opaqueness and malevolence.

Competence was by far the most common trust dimension in the full dataset, appearing in 70,851 classified posts. Familiarity appeared in 57,496 and reliability in 48,011.

Among distrust dimensions, unreliability appeared in 56,257 posts and incompetence in 42,312.

Ethical and value-oriented categories appeared less frequently. Integrity was identified in 8,480 trust posts, while malevolence appeared in 10,303 distrust posts.

“For many users, trust in AI is expressed primarily in relation to their direct experience with what these systems can actually do,” Pessianzadeh said, “suggesting that people are more focused on whether AI programs ‘work’ than whether the technology is ‘good.’”

Those judgments also tended to cluster together. Among trust-related posts, competence and familiarity showed the strongest overlap, followed by competence and reliability.

Within distrust discussions, incompetence and unreliability were most closely connected.

The researchers found that ethical dimensions were harder for both human annotators and computational models to classify consistently. Their analysis found a strong correlation between human agreement and model performance across these dimensions.

Trust and distrust did not remain perfectly stable.

Some of the clearest changes appeared around major product releases and announcements.

The volume of generative AI discussion surged after ChatGPT's release in late 2022. Later peaks aligned with launches including GPT-4 in March 2023 and LLaMA 2 in July 2023, as well as OpenAI's Dev Day in November 2023.

Trust rose modestly around several major releases in 2023, while distrust increased around OpenAI's Dev Day.

Distrust also briefly exceeded trust during February 2023 and February 2025. During February 2024, the two positions converged.

The language used during those periods offered clues about what was driving skepticism.

In February 2023, distrust discussions prominently included terms connected with Replika, its parent company Luka and DAN, or “Do Anything Now.” By 2024 and 2025, distrust increasingly centered on technical frustrations reflected by words such as “issue,” “refuses,” “deleted” and “wrong.”

The word “wrong” appeared prominently during all three periods examined, pointing to perceived inaccuracy as a persistent source of distrust.

Trust discussions, meanwhile, emphasized words tied to usefulness, collaboration and positive experiences, including “help,” “use,” “create,” “feedback,” “project,” “love” and “best.”

The researchers also classified posters into groups based on self-identifying information in their posts.

Trust tended to outweigh distrust among business leaders, academics, software developers and technology professionals.

Distrust appeared more often among the general public, AI ethicists, and media and journalists.

Generative AI users formed the largest group and showed a relatively even split between trust and distrust. Educators and knowledge workers also appeared divided.

The study's reliance on Reddit creates an important limitation. Conversations on one social media platform may not represent public opinion more broadly.

The relatively limited discussion of ethics may also reflect how people communicate online. Immediate experiences with usefulness or failures may be easier to express than broader concerns about accountability or values.

Future work could expand the analysis to other platforms, languages and populations.

The persistent coexistence of trust and distrust could influence how generative AI systems are designed, governed and explained to users.

“A persistent duality like this means that responsible governance of the technology should consider that users will include both those who are confident in AI as well as those who are skeptical,” Rezapour said.

The findings also suggest that practical performance cannot be the only concern.

Governance and AI literacy efforts may need to connect issues such as reliability, transparency, bias and accountability to the everyday experiences that already shape people's judgments.

“These findings are a useful starting point, but trust in AI is not static and people’s attitudes change based on their experiences with these systems, new capabilities and major developments in the technology,” Rezapour said.

As generative AI becomes more embedded in daily life, tracking those changes could reveal whether today's uneasy balance eventually shifts, or whether confidence and skepticism remain permanent companions to the technology.

These resources examine how people judge AI reliability, why trust varies between users and experts, and how system design can strengthen or distort confidence in generative AI.

How the US Public and AI Experts View Artificial Intelligence
This large survey compares attitudes among U.S. adults and AI experts, finding substantial gaps in optimism, concern and expectations about AI while showing broad interest in greater oversight of the technology. (Pew Research Center, 2025)

What large language models know and what people think they know
This research examines whether large language models accurately communicate uncertainty and whether people can recognize when their answers are likely to be wrong, addressing a central problem in calibrating appropriate trust in AI-generated information. (Nature Machine Intelligence, 2025)

The influence of mental state attributions on trust in large language models
Experiments found that perceptions of an AI system's intelligence were associated with greater trust in its advice, while attributing human-like experiences or emotions to the system could affect trust differently. (Communications Psychology, 2025)

Human Trust in AI Search: A Large-Scale Experiment
A large randomized experiment found that citations and reference links can increase trust in generative AI search results even when those references are incorrect or hallucinated, showing how interface features can influence confidence independently of accuracy. (MIT Initiative on the Digital Economy, 2025)

Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
NIST's generative AI profile identifies risks specific to these systems and provides a framework for organizations to evaluate reliability, transparency, accountability and other dimensions of trustworthy AI development and use. (National Institute of Standards and Technology, 2024)

Research findings are available online in the journal Transactions of the Association for Computational Linguistics.

The original story "New study reveals what people really think about generative AI" is published in The Brighter Side of News.

Like these kind of feel good stories? Get The Brighter Side of News' newsletter.