Quantifying the Epistemic Trust Gap: A Framework for Calibrating Human Expectations and Sentiment of Artificial Intelligence

Abstract

Artificial intelligence is becoming part of how we learn, work, create, and make decisions. Yet as AI becomes more capable, many people either trust it too much or dismiss it too quickly. This paper explores that mismatch, which is described as the Epistemic Trust Gap (ETG)—the difference between what people believe AI can do and what it can actually do. Methodologically, this study conducts a qualitative literature review synthesizing recent interdisciplinary research (2023–2026) across cognitive psychology, human–computer interaction, and artificial intelligence, alongside governmental and industry frameworks. This analysis examines how cognitive biases, persuasive AI-generated responses, media influence, and AI literacy shape public trust in AI. The review reveals that public trust is heavily distorted by surface-level fluency, anthropomorphic media narratives, and a misunderstanding of AI's 'jagged frontier'—where systems excel at complex benchmarks yet fail at simple tasks. To address this, the paper articulates a five-component sentiment-driven framework incorporating continuous public sentiment tracking, benchmark evaluations, real-time confidence feedback loops, targeted trust interventions, and institutional AI literacy programs. Through examples from healthcare, education, business, software development, and other industries, this paper argues that the future of AI is not about replacing human judgment, but instead improving collaboration between people and intelligent systems. Ultimately, reducing the Epistemic Trust Gap will require not only better AI but also better-informed users who understand both the strengths and limitations of these technologies. 

Introduction: A Day in the Life with AI

It’s a typical Monday morning. Maya, a project manager at a mid-sized marketing firm, sits at her kitchen table, coffee in hand, reviewing her to-do list. She opens her laptop and launches her favorite AI-powered assistant—let’s call it “Ava.” Within seconds, Ava summarizes overnight emails, drafts a response to a client, and suggests a new campaign slogan. Later, Maya asks Ava for advice on a complex budgeting issue. The answer is swift, confident, and detailed. Impressed, Maya copies the suggestion into her report, barely pausing to double-check the numbers.

But what if Ava’s advice, though plausible and well-written, is subtly flawed? Maya’s trust in Ava’s output is shaped by the assistant’s fluency and confidence rather than by a clear understanding of how the AI arrived at its conclusion. This increasingly common scenario highlights a crucial challenge of the modern AI era. While existing research has examined AI trust, overreliance, and trust calibration, this paper proposes the Epistemic Trust Gap (ETG) as a conceptual framework for describing the mismatch between perceived AI capability and demonstrated AI capability—specifically, the difference between what people believe AI can do and what AI is actually capable of doing.

As artificial intelligence becomes woven into the fabric of daily life, the ETG is no longer an abstract philosophical concern. It is a practical and measurable phenomenon with profound implications for individuals, organizations, and society. This paper explores the origins, drivers, and consequences of the Epistemic Trust Gap by synthesizing findings from psychology, artificial intelligence, human–computer interaction, and large-scale public sentiment research. Building on this evidence, it proposes a sentiment-driven framework for calibrating human trust in AI. The primary goal of this paper is to examine why the Epistemic Trust Gap emerges and to identify strategies for fostering more accurate and appropriately calibrated trust in artificial intelligence.

Methodology

A qualitative literature review was conducted to examine the factors influencing public trust in artificial intelligence and to construct the proposed Epistemic Trust Gap (ETG) framework. Rather than collecting original experimental data, the study takes into account findings from peer-reviewed journal articles, academic conference papers, reports from research institutions, government publications, and industry analyses. 

To ensure a comprehensive and rigorous synthesis, sources were selected and evaluated based on four primary methodological criteria:

1. Thematic Relevance: Literature was included if it explicitly investigated AI trust dynamics, human–AI interaction (HAI), cognitive psychology, epistemic calibration, or digital literacy frameworks.

2. Organizational and Author Credibility: Eligible literature comprised peer-reviewed journal articles, top-tier computer science conference proceedings, and technical reports from recognized research institutes and policy bodies (e.g., Stanford HAI, NIST, OECD, Pew Research Center, and the European Union).

3. Temporal Recency (2023–2026): Priority was given to studies published between 2023 and 2026 to capture the unique trust dynamics, capabilities, and risks introduced by modern foundation models and consumer generative AI tools.

4. Analytical Contribution: Sources were evaluated on their capacity to explain how individuals perceive, evaluate, over-trust, or under-trust AI systems, providing the conceptual building blocks for the Epistemic Trust Gap (ETG) framework.

To provide a balanced perspective, this review also incorporates research from universities and research institutions such as Stanford Human-Centered Artificial Intelligence (HAI), Pew Research Center, Harvard Business School, and the Organization for Economic Co-operation and Development (OECD), alongside government frameworks including the United States’  National Institute of Standards and Technology (NIST) AI Risk Management Framework and the European Union AI Act. Industry reports from organizations such as Adobe and Mastercard were included when they offered large-scale empirical findings or practical insights that complemented academic literature.

The selected sources were analyzed to identify recurring themes related to trust calibration, cognitive bias, AI hallucinations, media influence, public sentiment, AI literacy, and responsible AI governance. These themes informed the development of the Epistemic Trust Gap framework presented throughout this paper. 

Defining the Epistemic Trust Gap (ETG)

The term epistemic may sound complex, but it simply describes how people evaluate knowledge and decide what information is reliable, accurate and worth trusting. Furthermore, the Epistemic Trust Gap refers to the structural mismatch between the widespread integration of AI into everyday life and a user's ability to critically evaluate the reliability of AI-generated outputs. Put simply, it is the gap between perceived AI capability and actual AI capability.

This gap emerges because modern AI systems, particularly large language models and generative AI tools, often produce responses that appear fluent, coherent, and highly convincing. As a result, users may mistake persuasive language for factual accuracy, even when the information has not been verified. The ETG is therefore not simply about AI making mistakes; it is about people accepting AI-generated information as trustworthy without adequately evaluating its accuracy or supporting evidence.

Why does the ETG matter? When people overestimate AI's capabilities, they may rely on it too heavily, increasing the risk of poor decisions, safety issues, or ethical concerns. Conversely, when people underestimate AI, they may ignore valuable insights, resist useful technologies, or fail to benefit from tools that can improve productivity and decision-making, among other advantages. The ETG therefore influences not only individual choices but also organizational performance, public policy, and the broader adoption of artificial intelligence.

The Psychological Roots of Overestimation and Underestimation

Pseudo-confident Knowledge and Cognitive Bias

A major contributor to the ETG is the phenomenon of pseudo-confident knowledge—AI-generated information that appears clear, logical, and complete despite lacking verified accuracy (Kurmanbayeva et al., 2026). Because this information is presented with confidence and linguistic fluency, users may mistakenly accept it as reliable knowledge without critically evaluating its validity.

Several well-established cognitive biases reinforce this tendency.

• The Dunning–Kruger Effect: Traditionally, this cognitive bias describes how individuals with limited knowledge or expertise often overestimate their own competence. Recent research suggests that AI assistance can increase users' confidence regardless of their actual skill level. As people rely more heavily on AI-generated responses, they may begin to attribute the quality of those responses to their own expertise. Computer scientist and AI researcher Adrian Stan (Stan, A. 2026), whose work focuses on human–AI interaction and AI-assisted cognition, describes this proposed phenomenon as False Cognitive Power Transfer (FCPT)—the mistaken attribution of AI-assisted performance to one's own cognitive ability—which may encourage overconfidence, reduce critical thinking, and increase poor decision-making.

• Automation Bias: People naturally tend to trust recommendations produced by automated systems, particularly when those systems appear authoritative or provide little explanation for their reasoning (Okamura & Yamada, 2020). This often results in users accepting incorrect recommendations simply because they were generated by AI. 

• Cognitive Offloading: As AI assumes responsibility for more cognitive tasks, users may gradually reduce the amount of independent analysis they perform. Over time, excessive reliance on AI can weaken critical thinking skills, professional judgment, and problem-solving abilities (Okamura & Yamada, 2020). 

The Role of Epistemic Vigilance

Epistemic vigilance refers to an individual's ability to critically evaluate whether information is trustworthy before accepting it as true. In an AI-driven world, this skill has become increasingly important.

However, the convenience, speed, and persuasive nature of AI-generated content can reduce people's willingness to question what they read. Frequent AI use does not necessarily improve a person's ability to evaluate AI-generated information. In fact, research suggests that operational familiarity with AI, such as using it regularly, is not the same as developing the critical thinking skills needed to assess its reliability (Okamura & Yamada, 2020). As AI becomes more integrated into everyday life, strengthening epistemic vigilance will be essential for reducing the Epistemic Trust Gap.

Sentiment, Perception, and the Measurement of ETG

Sentiment Analysis: Tools and Methods

To quantify the Epistemic Trust Gap, researchers increasingly use sentiment analysis, a technique that applies natural language processing (NLP)—a subfield of artificial intelligence focused on enabling computers to understand, interpret, and analyze human language—and machine learning to evaluate the emotional tone and attitudes expressed in written text about artificial intelligence. Modern sentiment analysis can distinguish between positive, negative, and neutral opinions, identify emotions such as excitement, fear, or skepticism, and detect more nuanced perspectives regarding trust, uncertainty, and anxiety (Kouloukoui et al., 2025). Common approaches are included in the following table (Kouloukoui et al., 2025; Pessianzadeh et al., 2026):

Sentiment Analysis Method	How it works	Best Used For	Example in AI Trust Research
Transformer Models (e.g., BERT and GPT)	Uses deep learning to understand the meaning and context of complete sentences.	Capturing subtle opinions, sarcasm, and complex discussions.	Measuring whether users express trust or distrust toward ChatGPT in Reddit discussions.
Aspect Based Sentiment Analysis	Identifies sentiment toward specific aspects of a topic instead of the overall text.	Determining whether users trust AI’s accuracy, fairness, safety, or reliability separately.	Measuring opinions about AI reliability versus AI ethics in online reviews.
Emotion and Intent Detection	Detects underlying emotions such as fear, excitement, curiosity, confidence or skepticism	Understanding why users trust or distrust AI.	Identifying anxiety about AI replacing jobs or excitement about AI-assisted healthcare.

Researchers commonly collect sentiment data from multiple sources, each providing a different perspective on public attitudes toward artificial intelligence (Kouloukoui et al., 2025; Adobe, 2026; Stanford HAI, 2026; Pew Research Center, 2025):

• Social media platforms such as X (formerly Twitter), Reddit, and Facebook provide large volumes of real-time, unsolicited opinions. For example, researchers analyzed millions of Reddit posts between 2022 and 2025 to track changes in public trust and distrust of generative AI. Although social media captures immediate reactions, it may overrepresent highly active users and emotionally charged opinions.

• News articles and online publications reflect how AI is presented to the public. Studies comparing media coverage (like in Kouloukoui) with online discussions have shown that news reporting can strongly influence perceptions of AI by emphasizing breakthroughs, ethical concerns, or controversies, though it may not always represent everyday user experiences.

• Customer reviews and support interactions provide insights from people who have directly used AI systems. For instance, Adobe's Creators' Toolkit Report analyzed feedback from creators using AI-powered creative tools, offering valuable information about user trust, satisfaction, and adoption in real-world settings.

• Survey responses and open-ended questionnaires are among the most reliable sources because researchers can carefully select representative samples and ask standardized questions. Large-scale surveys conducted by organizations such as the Pew Research Center and the Stanford AI Index have been widely used to measure public trust in AI across different demographic groups and countries.

By combining these sources, researchers obtain a more complete understanding of public sentiment. Surveys generally provide the most representative picture of public opinion, while social media captures immediate reactions and customer feedback reflects firsthand user experiences. Together, these complementary sources allow researchers to monitor how trust in AI changes over time.

Large-Scale Sentiment Data: What the Numbers Say

Recent large-scale studies analyzing millions of survey responses and social media posts reveal a complex and rapidly evolving landscape of public trust in artificial intelligence (Pessianzadeh et al., 2026).

• Trust and Distrust Are Nearly Balanced: A multi-year Reddit analysis conducted between 2022 and 2025 found that expressions of trust and distrust toward generative AI are nearly equal, although trust has gradually become slightly more prevalent. Significant shifts in public sentiment often coincide with the release of major AI models or technological breakthroughs (Pessianzadeh et al., 2026). 

• Personal Experience Is the Strongest Influence: Research conducted by Kouloukoui, D., de Marcellis-Warin, N., and Warin T. in 2025 consistently shows that direct experience using AI tools has a greater impact on trust than media coverage or peer opinions. Users tend to base their perceptions on the quality, reliability, and usefulness of the AI systems they interact with regularly (Pessianzadeh et al., 2026). 

• Global Optimism and Anxiety Coexist: According to the 2026 Stanford AI Index, approximately 59% of respondents believe AI's benefits outweigh its drawbacks, while 52% also report feeling concerned or anxious about AI's rapid advancement. Countries in Southeast Asia generally express greater optimism toward AI, whereas respondents in the United States and parts of Europe remain comparatively more cautious (Pessianzadeh et al., 2026). 

• The Expert–Public Divide: A substantial gap continues to exist between AI experts and the general public. For example, 73% of AI experts believe AI will have a positive effect on employment, compared with only 23% of the general public. Similar differences appear in perceptions of AI's influence on healthcare, education, and economic growth (Pessianzadeh et al., 2026). 

These findings illustrate that attitudes toward AI are shaped not only by technological capability but also by personal experience, cultural context, and differing levels of AI literacy.

The Influence of Media, Social Media, and Science Fiction

Media and Social Media

Traditional media and online platforms play a significant role in shaping public perceptions of artificial intelligence. News coverage highlighting AI breakthroughs, failures, ethical controversies, or emerging regulations can rapidly influence public opinion (Kouloukoui et al., 2025). Likewise, viral social media posts and online discussions often amplify both excitement and concern, sometimes spreading unverified information before expert analysis can provide proper context.

In a large-scale computational study on public trust in generative AI across Reddit discussions, Pessianzadeh et al. (2026) observed that:

• Technical performance and practical usefulness are among the most frequently discussed aspects of AI. 

• Public trust and skepticism often fluctuate alongside major product announcements and news events. 

• Personal experiences and recommendations from friends, colleagues, and online communities strongly influence individual attitudes toward AI. 

These patterns demonstrate that public understanding of AI is shaped not only by direct experience but also by the information environment surrounding emerging technologies.

Science Fiction and Cultural Narratives

Andrea Armstrong, a researcher at Dakota State University, examined how science fiction influences public perceptions of artificial intelligence. According to Armstrong (2023), science fiction has played a significant role in shaping public perceptions of artificial intelligence. Stories featuring intelligent robots, virtual assistants, and autonomous machines have influenced expectations of what AI can and cannot do. 

Several recurring themes contribute to the Epistemic Trust Gap:

• Anthropomorphism: When AI communicates fluently or mimics human conversation, people often attribute human-like intelligence, reasoning, or understanding to the system, even though it operates through statistical pattern recognition rather than conscious thought (Armstrong, 2023). 

• Recurring Narratives: Fiction frequently portrays AI as either an all-knowing assistant or an existential threat. While engaging, these narratives often oversimplify the realities of AI development and create unrealistic expectations regarding its capabilities (Armstrong, 2023). 

• Ethical Concerns: Popular media regularly explores issues such as bias, surveillance, loss of control, and automation. Although these concerns are valuable topics for discussion, fictional portrayals can sometimes exaggerate the risks associated with current AI technologies (Armstrong, 2023). 

Science fiction has undoubtedly inspired innovation and encouraged important ethical conversations. However, Armstrong emphasizes the importance of distinguishing fictional narratives from the demonstrated capabilities and limitations of modern AI systems.

The Jagged Frontier: Technical Capabilities and Limitations

The Reality of Modern AI

Despite rapid technological progress, modern AI remains characterized by what researchers from Georgia State University often describe as a "jagged frontier" of capability. AI systems may outperform humans on certain specialized tasks while simultaneously struggling with problems that people find relatively simple (OECD, 2025; Georgia State University, 2024).

For example:

• AI models can achieve exceptional performance on complex mathematical reasoning and coding benchmarks while still making mistakes when interpreting analog clocks or following ambiguous real-world instructions (OECD, 2025; Georgia State University, 2024). 

• In healthcare, AI systems can identify specific diseases from medical images with remarkable accuracy but may perform poorly when encountering rare conditions or cases that differ significantly from their training data (OECD, 2025; Georgia State University, 2024). 

• In software development, AI coding assistants can rapidly generate functional code for common programming tasks but may also introduce subtle bugs, security vulnerabilities, or inefficient implementations that require human review (OECD, 2025; Georgia State University, 2024). 

Such disparities underscore that modern AI functions as a specialized task-support tool rather than an all-encompassing intelligence.

AI Augmentation vs. Replacement: The Real Future of Work

Augmentation: AI as a Collaborative Partner

Contrary to widespread fears of mass automation, current research suggests that artificial intelligence is more likely to augment human work than replace it entirely. Rather than eliminating jobs outright, AI is increasingly being used to support employees by automating repetitive tasks, enhancing decision-making, and improving productivity (Harvard Business School, 2026; MIT Sloan School of Management, 2025). This allows people to devote more time to creative thinking, strategic planning, and complex problem-solving.

Research from MIT Sloan highlights five uniquely human capabilities that remain difficult for AI to replicate: Empathy, Presence, Opinion and Judgment, Creativity, and Hope. Tasks requiring these qualities are significantly less susceptible to automation and are more likely to benefit from AI-assisted collaboration rather than replacement. Similarly, research from Harvard Business School found that following the widespread adoption of generative AI, demand for occupations involving analytical, technical, and creative skills increased by approximately 20%, while job postings for highly repetitive and easily automated work declined by about 13%. These findings suggest that AI is reshaping work by changing the nature of many occupations rather than eliminating them altogether (MIT Sloan School of Management, 2025; Harvard Business School, 2026).

Replacement: Where Automation Makes Sense

Although AI is unlikely to replace most professions entirely, it can effectively automate repetitive, rule-based tasks requiring minimal human judgment—such as data entry, basic customer service inquiries, invoice processing, routine administrative workflows, and certain manufacturing and logistics operations. In these environments, automation can significantly improve speed, efficiency, and consistency. However, even highly automated industries continue to rely on human oversight for quality assurance, decision-making, exception handling, and ethical responsibility. 

In these environments, automation can significantly improve speed, efficiency, and consistency. However, even highly automated industries continue to rely on human oversight for quality assurance, decision-making, exception handling, and ethical responsibility.

Rather than replacing workers completely, successful organizations increasingly combine AI-driven automation with human expertise to achieve better overall outcomes. The future of work is best described as hybrid rather than fully automated. Instead of humans competing against AI, most professionals will increasingly collaborate with intelligent systems that enhance productivity and support decision-making. AI will perform repetitive computational tasks, while humans continue to contribute creativity, ethical reasoning, emotional intelligence, leadership, and contextual understanding. As this collaboration becomes more common, organizations will need to invest in AI literacy, workforce training, and continuous skill development to ensure employees can work effectively alongside AI technologies.Ultimately, success in the AI era will depend not on competing with machines but on learning how to collaborate with them effectively.

Sector Case Studies: Benefits and Limitations of AI

Healthcare

Artificial intelligence is transforming healthcare by improving diagnostic accuracy, supporting personalized treatment plans, and increasing operational efficiency (Stanford Human-Centered Artificial Intelligence [HAI], 2026).

Examples include:

• Diagnostics: AI systems can analyze medical images with accuracy comparable to, and in some cases exceeding, that of experienced medical specialists. For example, AI-assisted breast cancer detection has demonstrated high levels of diagnostic accuracy, enabling earlier intervention and improved patient outcomes (Stanford HAI, 2026).

• Personalized Medicine: AI analyzes patient histories, genetic information, and clinical data to recommend more individualized treatment strategies while helping reduce adverse side effects (Stanford HAI, 2026). 

• Operational Efficiency: Predictive analytics assist hospitals in managing staffing, scheduling, patient flow, and resource allocation, reducing wait times and improving overall efficiency (Stanford HAI, 2026). 

• Telemedicine: AI-powered virtual assistants and chatbots provide preliminary health guidance, answer common questions, and expand access to healthcare services, particularly in underserved communities (Stanford HAI, 2026). 

Limitations

Despite these advances, healthcare remains a domain where human oversight is essential. Excessive reliance on AI recommendations can result in "silent errors," particularly when clinicians accept AI-generated advice without independent evaluation. Additionally, many AI models function as "black boxes," making it difficult to explain how decisions were reached. Overreliance on AI also raises concerns about the potential de-skilling of future healthcare professionals if critical diagnostic reasoning is not continuously practiced.

Education

Artificial intelligence is reshaping education by personalizing learning experiences, reducing administrative workloads, and improving student support services (Organisation for Economic Co-operation and Development [OECD], 2025). In practice, these transformations manifest across several key areas:

• Personalized Learning: Platforms such as Duolingo adapt lessons to individual learning styles and progress, helping improve engagement and long-term knowledge retention (OECD, 2025). 

• Automated Student Support: AI-powered chatbots, such as Georgia State University's Pounce, assist students by answering common questions, sending reminders, and helping reduce "summer melt," the phenomenon in which admitted students fail to enroll before the academic year begins (Georgia State University, 2024). 

• Improved Accessibility: AI technologies provide additional educational support for students with disabilities, language barriers, or limited access to traditional educational resources (OECD, 2025). 

Education also presents important challenges. Concerns remain regarding academic integrity, algorithmic bias, student privacy, and the appropriate role of AI in learning environments. Many educators report feeling underprepared to teach students how to use AI responsibly, highlighting the growing need for comprehensive AI literacy initiatives.

Business and Organizational Adoption

Artificial intelligence is rapidly becoming a central component of modern business operations. Recent studies from MIT Sloan School and Stanford indicate that organizational AI adoption continues to grow across nearly every industry, driven by improvements in productivity, customer engagement, and operational efficiency (MIT Sloan School of Management, 2025; Stanford Human-Centered Artificial Intelligence, 2026).

Examples include:

• Productivity: AI automates repetitive workflows, allowing employees to focus on higher-value strategic work (MIT Sloan School of Management, 2025; Stanford Human-Centered Artificial Intelligence, 2026). 

• Customer Insights: Machine learning algorithms analyze large datasets to better understand customer behavior, personalize recommendations, and improve marketing effectiveness (MIT Sloan School of Management, 2025; Stanford Human-Centered Artificial Intelligence, 2026). 

• Small Business Adoption: Small businesses are increasingly adopting generative AI for content creation, customer support, and operational planning, although workforce training continues to be a significant challenge (MIT Sloan School of Management, 2025; Stanford Human-Centered Artificial Intelligence, 2026). 

Despite growing adoption, organizations continue to face obstacles related to employee training, data quality, cybersecurity, and governance. Without proper oversight, excessive reliance on AI can reduce human decision-making and introduce new operational risks (Stanford Human-Centered Artificial Intelligence, 2026; MIT Sloan School of Management, 2025).

Software Development

Artificial intelligence has become an increasingly valuable tool in software engineering through AI-powered coding assistants such as GitHub Copilot (Stanford Human-Centered Artificial Intelligence, 2026).

Benefits include:

• Accelerating code generation for routine programming tasks. 

• Assisting with debugging and documentation. 

• Improving developer productivity by automating repetitive coding activities. 

Despite these advantages, AI-generated code is not always reliable. Coding assistants may introduce subtle programming errors, security vulnerabilities, inefficient algorithms, or fabricated code that appears correct but fails during implementation. Human review, testing, and software quality assurance remain essential components of the development process (Gosmar & Dahl, 2025; Pulkundwar et al., 2025).

Finance

Artificial intelligence has become an important technology within financial services, supporting fraud detection, risk management, customer service, and personalized banking experiences (Mastercard, 2026). Financial institutions leverage these technologies across several key operational areas:

• Fraud Detection: AI systems analyze millions of financial transactions in real time, allowing institutions such as Mastercard to identify suspicious activity within milliseconds. 

• Personalized Banking: AI-powered assistants help customers monitor spending, manage budgets, and detect unusual account activity. 

• Risk Management: Financial institutions use AI to analyze market conditions, customer behavior, and global economic events to improve forecasting and decision-making. 

Financial AI systems must remain transparent and explainable, particularly in regulated environments. Overreliance on automated decisions can increase the likelihood of false positives or missed fraudulent activity if human oversight is reduced (European Union, 2024; NIST, 2026).

Creative Industries

Generative AI is transforming creative fields by assisting artists, designers, writers, filmmakers, and content creators throughout the creative process (Adobe, 2025). In practice, these integrations manifest across several key areas: 

• Content Creation: Many creators now use generative AI to accelerate brainstorming, drafting, editing, and multimedia production (Adobe, 2025). 

• Creative Tools: Platforms such as Adobe Creative Cloud integrate AI-powered features for image editing, content generation, object removal, and workflow automation while allowing creators to maintain artistic control (Adobe, 2025). 

• Business Growth: Surveys indicate that many creative professionals report increased productivity and business growth after incorporating generative AI into their workflows (Adobe, 2025). 

The rapid adoption of generative AI has also introduced concerns regarding copyright, intellectual property, originality, and the ethical use of creative works for AI model training. As these technologies continue to evolve, maintaining transparency, proper attribution, and respect for creators' rights will remain essential (Adobe, 2025).

Measuring and Calibrating the ETG: Frameworks and Benchmarks

Trust Calibration: Concepts and Metrics

Trust calibration is the process of aligning a user's level of trust with the actual reliability and capabilities of an AI system. Proper trust calibration is essential for effective human–AI collaboration because both over-trust and under-trust can negatively affect decision-making.

Given these limitations, trust calibration is essential for effective human-AI collaboration.

When trust is too high, users may rely on AI without sufficient verification, increasing the likelihood of errors and safety risks (Gosmar & Dahl, 2025; Pulkundwar et al., 2025). Conversely, when trust is too low, individuals may ignore valuable AI-assisted insights and fail to realize productivity gains.

Achieving appropriate trust calibration enables users to leverage AI effectively while maintaining critical judgment and independent decision-making.

Researchers evaluate trust calibration using several established metrics and methods:

• Expected Calibration Error (ECE): Measures how closely an AI system's confidence aligns with its actual accuracy. Lower values indicate better calibration between predicted confidence and real-world performance. 

• Trust Calibration Error (TCE): Measures the difference between a user's reported trust in an AI system and the system's objectively measured reliability. 

• Behavioral Indicators: Researchers observe behaviors such as whether users accept or reject AI recommendations, switch between manual and automated modes, or modify AI-generated outputs before using them. 

• Self-Report and Physiological Measures: Surveys, interviews, eye-tracking, and biometric data can provide additional insight into how users perceive and trust AI systems during decision-making. 

Together, these approaches provide a more comprehensive understanding of how individuals interact with AI and whether their level of trust appropriately reflects the system's capabilities (Okamura & Yamada, 2020; Trust Calibration Maturity Model, 2025).

The Trust Calibration Maturity Model (TCMM)

Developed as a standardized evaluation framework in recent AI trust research published on arXiv, the Trust Calibration Maturity Model (TCMM) was introduced to help organizations measure and align user trust with actual system performance (Trust Calibration Maturity Model, 2025). The model assesses AI systems across five key dimensions that influence user perception, decision-making, and overall system reliability (Trust Calibration Maturity Model, 2025).

Dimension	Description
Performance Characterization	Measures how accurately the system's performance, reliability, uncertainty, and limitations are evaluated and communicated.
Bias and Robustness	Assesses the presence of systematic errors and evaluates how consistently the system performs across different users and environments.
Transparency	Examines how easily users can understand, interpret, and anticipate the system's behavior and decision-making processes.
Safety and Security	Evaluates safeguards against misuse, adversarial attacks, privacy risks, and system vulnerabilities.
Usability	Measures the quality of the user experience, including ease of use, clarity, accessibility, and the likelihood of user error.

Higher maturity levels require continuous evaluation, user-centered design, transparent communication, and ongoing improvements as AI systems evolve. The TCMM provides organizations with a structured framework for assessing trustworthiness, communicating system limitations, and identifying opportunities for improvement.

Sentiment-Driven Framework for ETG Calibration

A sentiment-driven framework combines large-scale sentiment analysis with trust calibration metrics to monitor and improve public understanding of artificial intelligence. Rather than measuring technical performance alone, this approach also evaluates how people perceive, interpret, and respond to AI technologies (Okamura & Yamada, 2020).

The framework consists of five primary components:

1. Continuous Sentiment Monitoring

Natural language processing techniques are used to analyze public discussions about AI across surveys, news articles, online communities, and social media platforms. Monitoring sentiment over time allows researchers to identify changing patterns in public trust, fear, and expectations (Okamura & Yamada, 2020).

2. Benchmarking and Evaluation

Standardized evaluation benchmarks—including HELM, HarmBench, and TruthfulQA—can be used to assess AI systems for safety, factual accuracy, robustness, and reliability. Comparing public perceptions with benchmark performance helps identify where trust is appropriately calibrated and where misconceptions exist (Okamura & Yamada, 2020).

3. Feedback Loops

Providing users with transparent information about AI performance, confidence levels, uncertainty estimates, and known limitations encourages more informed decision-making. Continuous feedback helps users develop more realistic expectations regarding AI capabilities (Okamura & Yamada, 2020).

4. Targeted Trust Interventions

When signs of over-trust or under-trust are detected, systems can provide carefully designed prompts, warnings, or explanatory messages that encourage users to reconsider their level of reliance on AI-generated outputs (Okamura & Yamada, 2020).

5. Education and AI Literacy

Long-term improvements in trust calibration require strong AI literacy programs that teach users how AI systems function, where they perform well, where they struggle, and how to evaluate AI-generated information critically (Okamura & Yamada, 2020).

Together, these five components provide a practical framework for reducing the Epistemic Trust Gap by aligning public perception with the demonstrated capabilities of modern AI systems (Okamura & Yamada, 2020).

AI Literacy: The Foundation for Closing the ETG

What Is AI Literacy?

AI literacy refers to the knowledge, skills, and attitudes required to understand, evaluate, and use artificial intelligence responsibly. It extends beyond simply knowing how to operate AI tools and emphasizes the ability to critically assess AI-generated information and make informed decisions (Bechtold, 2025; OECD, 2025). At its core, AI literacy encompasses understanding model capabilities and limitations, critically evaluating generated outputs for bias and uncertainty, using tools ethically, and questioning pseudo-confident responses (Bechtold, 2025). Developing these skills enables individuals to use AI as a tool for informed decision-making rather than relying on it unquestioningly.

Without widespread AI literacy, the Epistemic Trust Gap will continue to grow, increasing the likelihood of misuse, missed opportunities, and poor decision-making. Reducing this gap requires shared commitment from educators, employers, governments, and technology developers.

Education

AI literacy is rapidly becoming an essential educational competency. Organizations such as the OECD and the European Commission have developed frameworks that encourage integrating AI education into school curricula from an early age. Teaching students how AI works—and how to evaluate its outputs critically—helps prepare them for an increasingly AI-driven world (Bechtold, 2025; Organisation for Economic Co-operation and Development, 2025).

Workforce Readiness

As AI transforms workplaces across industries, employees must continuously develop new technical and critical thinking skills. Organizations that invest in AI literacy often experience greater innovation, improved productivity, and higher employee confidence when adopting AI technologies (Kouloukoui et al., 2025).

Public Policy

According to Bechtold (2025), governments also play a critical role in promoting AI literacy by developing policies that address digital citizenship, privacy, ethics, workforce development, and responsible AI adoption. Ensuring that citizens understand both the benefits and limitations of AI helps support informed participation in an increasingly technology-driven society.

Responsible AI Governance and Communication

Building Trust Through Transparency

Reducing the Epistemic Trust Gap requires more than improving AI systems themselves—it also requires improving how those systems communicate with users. Transparency enables individuals to make informed decisions by helping them understand what AI can and cannot reliably accomplish (National Institute of Standards and Technology, 2026; Okamura & Yamada, 2020).

Organizations can strengthen trust by:

• Clearly communicating the intended purpose of an AI system. 

• Explaining known limitations and potential sources of error. 

• Providing confidence estimates where appropriate. 

• Identifying situations that require human review or oversight. 

When users understand an AI system's strengths and limitations, they are better equipped to calibrate their trust appropriately (National Institute of Standards and Technology, 2026; Okamura & Yamada, 2020).

Explainability and Human Oversight

Explainable AI (XAI) has become an important area of research because users are more likely to trust systems whose recommendations can be understood and evaluated (European Union, 2024; NIST, 2026).

Although not every AI model can fully explain its internal reasoning, organizations can improve transparency by providing:

• Clear explanations of how recommendations are generated. 

• Descriptions of the data used to train the model. 

• Information about known biases and limitations. 

• Documentation describing intended use cases and situations where the system may perform poorly. 

Equally important is maintaining meaningful human oversight. AI should support but not replace human judgment, particularly in high-risk domains such as healthcare, finance, education, and law. Human review remains essential for ensuring accountability, identifying errors, and addressing situations that require ethical reasoning or contextual understanding (European Union, 2024; NIST, 2026).

Model Cards and AI Documentation

One practical approach to improving transparency is the use of Model Cards and similar documentation standards (NIST, 2026; Stanford Center for Research on Foundation Models, 2026). These documents provide users with accessible information about an AI system, including:

• Intended applications. 

• Performance across different tasks. 

• Known limitations. 

• Evaluation procedures. 

• Ethical considerations. 

• Potential risks and mitigation strategies. 

Providing standardized documentation encourages responsible AI deployment and helps users develop realistic expectations regarding system performance (NIST, 2026; Stanford Center for Research on Foundation Models, 2026).

Recommendations for Reducing the Epistemic Trust Gap

Successfully reducing the Epistemic Trust Gap requires coordinated efforts from individuals, educators, organizations, governments, and AI developers.

For Individuals

• Verify important AI-generated information using reliable sources. 

• Treat AI as a decision-support tool rather than an unquestionable authority. 

• Maintain critical thinking when evaluating AI recommendations. 

• Understand that confidence in an AI response does not necessarily indicate accuracy. 

For Educators

• Integrate AI literacy into educational curricula. 

• Teach students how AI systems generate information and where they may fail. 

• Encourage critical evaluation of AI-generated content. 

• Promote ethical and responsible AI use across disciplines. 

For Organizations

• Develop clear governance policies for AI deployment. 

• Provide employees with ongoing AI literacy and professional development programs. 

• Establish procedures for validating AI-generated outputs before implementation. 

• Maintain appropriate levels of human oversight for high-impact decisions. 

For Governments and Policymakers

• Develop evidence-based AI regulations that balance innovation with public safety. 

• Promote transparency standards for AI developers. 

• Support research on AI trust, fairness, and explainability. 

• Encourage public education initiatives focused on AI literacy and responsible technology use. 

For AI Developers

• Prioritize transparency, safety, and robustness during system development. 

• Design interfaces that communicate uncertainty and confidence appropriately. 

• Reduce hallucinations through improved model evaluation and testing. 

• Continue developing explainable AI techniques that help users better understand system behavior. 

Looking Ahead: The Future of Human–AI Collaboration

Artificial intelligence will continue to evolve rapidly, becoming increasingly integrated into nearly every aspect of society. As AI capabilities expand, the challenge will not simply be building more powerful systems but ensuring that people understand how to use them responsibly.

The future of AI should not be measured solely by technological progress. Instead, success will depend on how effectively humans and AI collaborate, complement one another, and build relationships based on appropriate levels of trust.

Closing the Epistemic Trust Gap will require continued advances in AI literacy, transparent system design, responsible governance, and ongoing research into human-AI interaction. These efforts will help individuals make informed decisions while allowing society to realize the benefits of AI without placing unwarranted trust in its capabilities.

Conclusion

Artificial intelligence is transforming the way people learn, work, create, and solve problems. Yet its rapid adoption has also highlighted an important challenge: public perception often fails to align with AI's actual capabilities. This mismatch, described in this article as the Epistemic Trust Gap (ETG), influences how individuals interact with AI, how organizations deploy it, and how society responds to technological change.

Throughout this article, the psychological foundations of trust, the influence of media and culture, the limitations of modern AI systems, and the growing importance of AI literacy were all examined. The practical strategies for improving trust calibration through transparency, explainability, governance, education, and responsible AI development were also explored.

Current evidence suggests that artificial intelligence is far more likely to augment human abilities than replace them entirely. The greatest opportunities will emerge not from choosing between humans and AI but from developing systems that combine computational efficiency with uniquely human qualities such as creativity, empathy, ethical reasoning, and critical judgment.

Ultimately, reducing the Epistemic Trust Gap is not simply about improving artificial intelligence—it is about improving human understanding of artificial intelligence. By fostering evidence-based expectations, encouraging critical thinking, and promoting responsible AI use, society can better harness the benefits of these technologies while minimizing unnecessary fear, misplaced confidence, and avoidable risks.

As artificial intelligence continues to evolve, perhaps the most important question is not whether AI will become more intelligent, but whether we will become more informed, thoughtful, and responsible in the way we choose to trust it.

References

Adobe. (2025). Adobe delivers new AI innovations, assistants and models across Creative Cloud   to empower creative professionals. https://news.adobe.com/news/2025/10/adobe-max-2025-creative-cloud

Adobe. (2025). Inaugural Adobe Creators' Toolkit Report. https://news.adobe.com/news/downloads/pdfs/2025/10/10282025-inauguraladobecreatorstoolkitreport.pdf

Armstrong, A. (2023). Science fiction media's influence on public perceptions of AI and technology. Dakota State University. https://scholar.dsu.edu/honors/11/

Bechtold, L. (2025). Why AI literacy is crucial for responsible AI transformation. World Economic Forum. https://www.weforum.org/stories/emerging-technologies/ai-literacy-and-strategic-transformation/

European Union. (2024). Artificial Intelligence Act. https://artificialintelligenceact.eu/the-act/

Georgia State University. (2024). National Institute for Student Success awarded $7.6 million to study benefits of AI-enhanced classroom chatbots. https://news.gsu.edu/2024/01/11/national-institute-for-student-success-awarded-7-6-million-grant-by-u-s-department-of-education/

Gosmar, D., & Dahl, D. A. (2025). Hallucination mitigation using agentic AI natural language-based frameworks. arXiv. https://arxiv.org/abs/2501.13946

Harvard Business School. (2026). Enhance or eliminate? How AI will likely change these jobs. https://www.library.hbs.edu/working-knowledge/enhance-or-eliminate-how-ai-will-likely-change-these-jobs

Kouloukoui, D., de Marcellis-Warin, N., & Warin, T. (2025). Balancing risks and benefits: Public perceptions of AI through traditional surveys and social media analysis. Harvard Dataverse.

Kurmanbayeva, L. T., Tanabayeva, A. S., Doszhanova, A. I., Olzhashov, A. A., Bakarassov, D., & Bisenbaev, A. K. (2026). The pseudo-confidence paradox: The epistemic gap in everyday AI use. Philosophies, 11(3), 97. https://doi.org/10.3390/philosophies11030097

Mastercard. (2026). Decision Intelligence for fraud and risk management. https://www.mastercard.com/us/en/business/cybersecurity-fraud-prevention/risk-decisioning/decision-intelligence.html

MIT Sloan School of Management. (2025). New MIT Sloan research suggests that AI is more likely to complement, not replace, human workers. https://mitsloan.mit.edu/press/new-mit-sloan-research-suggests-ai-more-likely-to-complement-not-replace-human-workers

National Institute of Standards and Technology. (2026). AI Risk Management Framework (AI RMF). https://www.nist.gov/itl/ai-risk-management-framework

Okamura, K., & Yamada, S. (2020). Adaptive trust calibration for human–AI collaboration. PLOS ONE, 15(2), e0229132. https://doi.org/10.1371/journal.pone.0229132

Organisation for Economic Co-operation and Development. (2025). Empowering learners for the age of AI. https://www.oecd.org/en/publications/empowering-learners-for-the-age-of-ai_65cd27d4-en.html

Pessianzadeh, A., et al. (2026). In generative AI we (dis)trust? Computational analysis of trust and distrust in Reddit discussions. arXiv. https://arxiv.org/abs/2510.16173

Pew Research Center. (2025). How the U.S. public and AI experts view artificial intelligence. https://www.pewresearch.org/internet/2025/04/03/how-the-us-public-and-ai-experts-view-artificial-intelligence/

Pew Research Center. (2025). How Americans view AI and its impact on people and society. https://www.pewresearch.org/science/2025/09/17/how-americans-view-ai-and-its-impact-on-people-and-society/

Pulkundwar, P., et al. (2025). A concise review of hallucinations in large language models and their mitigation. arXiv. https://arxiv.org/abs/2512.02527

Stanford Center for Research on Foundation Models. (2026). Holistic Evaluation of Language Models (HELM): Safety Benchmark. https://crfm.stanford.edu/helm/

Stanford Human-Centered Artificial Intelligence. (2025). AI Index Report 2025. https://hai.stanford.edu/ai-index/2025-ai-index-report

Stanford Human-Centered Artificial Intelligence. (2026). AI Index Report 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report

Stan, A. (2026). False Cognitive Power Transfer (FCPT): From Individual Atrophy to Collective Demoralization in the Age of AI. Zenodo. https://doi.org/10.5281/zenodo.18110162

Trust Calibration Maturity Model. (2025). arXiv. https://arxiv.org/abs/2503.15511