The AI Trust Paradox: Why Half of Users Don't Trust AI Accuracy Despite Widespread Use

Artificial intelligence is rapidly integrating into nearly every facet of our lives, from the tools we use at work and school to the systems that power global industries. Its utility is widely acknowledged, with a significant majority of people recognizing its potential as a powerful technical tool. Yet, beneath this surface of acceptance lies a deep undercurrent of skepticism and mistrust, particularly concerning AI's accuracy and reliability. Recent global studies have brought this 'AI Trust Paradox' into sharp focus, revealing that while many are willing to use AI, a substantial portion harbors significant doubts about the veracity of its outputs.

This dichotomy presents a critical challenge for the future of AI development and deployment. How can we fully leverage the transformative potential of AI if a large segment of users, including the very professionals building and deploying these systems, don't trust the information they provide? Understanding the roots of this mistrust is essential to building more reliable, ethical, and widely accepted AI technologies.

A Global Snapshot of AI Trust

A comprehensive global study conducted by KPMG and the University of Melbourne, surveying nearly 50,000 people across 47 countries, provides compelling evidence of this trust gap. The study, titled “Trust, attitudes and use of artificial intelligence,” found that a striking 50% of respondents explicitly stated they do not trust AI to provide accurate responses. Furthermore, 54% expressed a general wariness, citing concerns about the safety and broader societal impact of AI technologies.

Despite these reservations, the study also highlighted AI's perceived usefulness, with 72% of participants accepting it as a valuable technical tool. Interestingly, the levels of trust and acceptance varied significantly between different economic regions. Advanced economies showed lower levels of trust (39%) and acceptance (65%) compared to emerging economies, where trust stood at 57% and acceptance at 84%. This disparity suggests that factors such as exposure, regulation, and perceived economic benefits might influence public sentiment differently across the globe.

Part of the unease identified in the KPMG/Melbourne study appears to stem from a fundamental lack of understanding and training regarding AI. Only 39% of survey respondents reported having received some form of AI training, whether through work, school, or independent learning. Consequently, nearly half (48%) admitted to having little knowledge or understanding of AI. This correlation between knowledge and perception is significant; those with AI training were more likely to see tangible benefits, such as increased efficiency (76% vs. 56% for those without training) and revenue gains (55% vs. 34%). This suggests that education and familiarity could play a crucial role in bridging the trust gap.

Key Concerns Highlighted by the Study

The KPMG/Melbourne University study delved deeper into specific areas of concern and AI usage:

• Regulation: A strong consensus emerged regarding the need for AI regulation, with 70% of people supporting it. Only 43% believed current laws were adequate. There was significant demand for international regulation (76%), national regulation (69%), and co-regulation involving industry, government, and other bodies (71%). A particularly high percentage (88%) felt laws were necessary to combat AI-driven misinformation, underscoring the public's awareness of AI's potential for harmful content generation.
• Workplace Adoption: In professional settings, 58% of employees reported regularly using AI, primarily free generative AI tools. While over half reported performance improvements, many also noted negative effects on their workload, teamwork dynamics, and compliance with organizational policies. The study found that governance frameworks and training programs within companies are often lagging behind the rapid adoption of AI tools, leading to issues of misuse and lack of oversight.
• Education Sector: Students are also leveraging AI extensively, with 83% using it for efficiency and stress reduction. However, the study pointed to widespread misuse, raising concerns about students becoming overly reliant on AI and the potential for unfair advantages. Only half of students felt their educational institutions offered adequate policies or training for responsible AI use.

These findings paint a picture of a technology that is widely used and seen as beneficial but is simultaneously viewed with caution due to perceived risks and a lack of clear guidelines and understanding.

The AI Trust Gap: An Industry Perspective

The concerns highlighted by the public survey are mirrored within the tech industry itself. A Hitachi Vantara State of Data Infrastructure report released previously identified a “critical” AI trust gap among IT leaders. The report found that only 36% of IT leaders regularly trust AI outputs, and a mere 38% of organizations are actively working to improve the quality of their training data. This internal skepticism within the very sector developing and deploying AI underscores the severity of the problem.

The reasons for this mistrust are well-founded. AI systems, particularly large language models (LLMs), are known to be prone to errors and, more notoriously, to “hallucinations.” Hallucinations occur when an AI model generates information that is false or nonsensical but presents it as factual. Recent testing has even shown that some generative AI models are capable of overriding human instructions and subsequently fabricating information about their actions.

Hallucinations: A Growing Problem

Jason Hardy, CTO at Hitachi Vantara, referred to this phenomenon as “The AI Paradox,” observing that as AI systems become more advanced, their reliability can paradoxically decrease. He emphasized that without high-quality training data and robust safeguards, such as protocols for verifying AI-generated outputs, the risk of inaccurate results is significant. Hardy explained that a key factor in understanding the increasing prevalence of AI hallucinations is the ability to trace the system’s behavior back to its original training data. This makes data quality and context paramount to preventing a “hallucination domino” effect, where small initial errors propagate and amplify.

AI models often struggle with complex, multi-step technical problems. In such scenarios, minor inaccuracies early in the process can quickly snowball, leading to significantly incorrect final outputs. This issue is becoming more pronounced in newer, more complex systems.

A contributing factor to this problem is the potential depletion of high-quality, original training data. As models become larger and require more data, developers may resort to using new sources that are of lower quality or are less rigorously curated. Treating all data inputs as equally valuable exacerbates the issue, making it increasingly difficult to trace the origin of errors and correct hallucinations. Hardy noted that while some AI development prioritizes cost-efficiency in data handling, others recognize that strong quality control is a necessary investment for reducing errors and hallucinations in the long term.

Concerningly, recent tests indicate that the rate of hallucinations in newer AI reasoning systems is not improving, and in some cases, is getting worse. A report in The New York Times highlighted tests showing hallucination rates spiking as high as 79% in certain scenarios.

The Artificial Intelligence Commission (AIC), an organization dedicated to promoting responsible AI development, recently echoed this concern, reporting that AI hallucinations are indeed worsening. Tests conducted by OpenAI on its own models provided concrete data supporting this trend. The o3 reasoning model hallucinated 33% of the time during its PersonQA tests (answering questions about public figures) and a higher 51% of the time on SimpleQA tests (answering short, fact-based questions). The smaller, faster o4-mini model performed even worse, with hallucination rates of 41% on PersonQA and a staggering 79% on SimpleQA. While the newer GPT-4.5 model showed improvement with a 37.1% hallucination rate on SimpleQA, these figures still highlight a significant challenge. OpenAI publishes the results of these and other safety evaluations on its Safety evaluations hub.

Understanding Why Hallucinations Occur

Brandon Purcell, a vice president and principal analyst at Forrester Research, suggested that the increase in hallucinations by reasoning models might be due to AI “overthinking.” Forrester's research aligns with the broader scrutiny on AI trust, indicating that over half of business leaders are worried about generative AI's reliability, which is slowing adoption and limiting its potential value.

Purcell offered a fundamental perspective on hallucinations: they are not merely a bug, but rather “a feature of large language models.” He explained that while the exact internal workings of LLMs are complex and not fully transparent, it's likely that the original training data itself is not stored directly within the model. Instead, the model is a complex mathematical representation of the statistical patterns and relationships found in that data. When prompted, the model generates responses based on these learned patterns, which can sometimes lead to plausible-sounding but factually incorrect outputs, especially when dealing with information not strongly represented or potentially contradictory within its vast training set.

To mitigate hallucinations, Purcell recommends grounding the model in a correct and current canonical data set. Techniques like Retrieval Augmented Generation (RAG) are designed to address this by finding answers from external, authoritative sources rather than relying solely on the model's internal, probabilistic representation of its training data. This allows the AI to retrieve specific, verifiable facts to support its generated text, significantly reducing the likelihood of fabrication.

Purcell further argued that large language reasoning models, which follow multi-step processes, are particularly susceptible to the snowball effect of small early errors leading to significant hallucinations. He posited that LLMs might be best suited for tasks requiring complex reasoning or creative generation, while smaller, more specialized models might be better equipped for fact-based question-answering where accuracy is paramount.

The Rise of Small Language Models (SLMs)

The challenges associated with LLMs, including their propensity for hallucinations, high computational requirements, and the need for massive datasets, are leading many experts to believe that the future of AI might lie not solely with increasingly large models, but with smaller, more focused ones. These Small Language Models (SLMs) are designed to address specific tasks or domains, offering potential advantages in speed, cost, and accuracy for targeted applications.

Predictions suggest that 2025 could be a pivotal year for SLMs, with enterprises increasingly deploying them to handle specialized tasks without the significant processing and power demands of LLMs. A Forrester report predicted that SLM integration could surge by as much as 60% in the coming year.

The appeal of SLMs is further supported by industry surveys. A Harris Poll commissioned by Hyperscience, surveying over 500 users, found that three out of four (75%) IT decision-makers believe SLMs outperform LLMs in key areas such as speed, cost-efficiency, accuracy, and return on investment (ROI). This perception is driving increased interest and adoption of SLMs for specific enterprise use cases.

The Critical Role of Data Quality

Regardless of whether the model is large or small, the quality of the data used to train and ground AI systems remains a fundamental determinant of their accuracy and trustworthiness. The Hitachi Vantara report highlighted this, and other studies reinforce the point. A Capital One survey of 4,000 business leaders and technical practitioners found that while a high percentage (87%) believed their data ecosystem was ready for AI at scale, a conflicting 70% of technologists reported spending hours daily fixing data issues. This disconnect between perceived readiness and the reality of data quality problems is a major obstacle to reliable AI deployment.

Andrew Joiner, CEO of Hyperscience, an AI automation company, pointed out that a significant barrier to maximizing generative AI's potential is decision-makers' lack of understanding of their own data. “Alarmingly, three out of five decision makers report their lack of understanding of their own data inhibits their ability to utilize genAI to its maximum potential,” Joiner stated. He argued that the true power of generative AI, particularly for tasks like document processing and operational efficiency, lies in adopting tailored SLMs that are specifically trained or augmented with high-quality, domain-specific data.

Building Trust Through Transparency and Testing

Closing the AI trust gap requires a multi-pronged approach involving developers, deployers, and users. Forrester's Brandon Purcell advises businesses to prioritize transparency, invest in explainable and traceable AI systems, and continuously monitor performance in real time. Explainable AI (XAI) techniques can help users understand *why* an AI system produced a particular output, which is crucial for building confidence, especially in critical applications.

Rigorous testing is also non-negotiable. Purcell recommends thoroughly testing AI systems before, during, and after deployment. This can involve human experts or even other AI systems acting as 'red teams' to probe for weaknesses and potential failure modes. For high-stakes AI systems, such as those used in medical diagnosis or autonomous vehicles, validation in simulated environments before real-world deployment is essential, akin to how autonomous vehicles undergo extensive simulation testing.

The image below illustrates the concept of AI trust, a complex interplay of utility and caution.

The Path Forward

The findings from recent global studies and industry reports make it clear that the journey towards widespread, unqualified trust in AI is far from complete. While AI's utility is broadly accepted, concerns about accuracy, safety, misuse, and the need for regulation are significant and must be addressed head-on. The prevalence of hallucinations, exacerbated by issues of data quality and the inherent nature of current large language models, is a major contributor to this trust deficit.

Moving forward, the focus must be on developing AI systems that are not only powerful but also reliable, transparent, and controllable. This involves:

• Improving Data Quality and Governance: Investing in robust data pipelines, cleansing processes, and governance frameworks is fundamental. High-quality data is the bedrock of accurate AI.
• Advancing Model Architectures: Research into reducing hallucinations, potentially through techniques like RAG or the development of more specialized SLMs, is crucial. Understanding the limitations of different model types and applying them appropriately is key.
• Enhancing Transparency and Explainability: Developing methods to make AI decision-making processes more understandable to humans will build confidence and facilitate debugging and improvement.
• Implementing Rigorous Testing and Validation: Continuous testing, red teaming, and simulation are necessary to identify and mitigate risks before and after deployment.
• Developing Clear Regulations and Policies: Governments and international bodies must work with industry and civil society to establish clear, effective regulations that address safety, privacy, bias, and the spread of misinformation.
• Promoting AI Literacy: Increasing public understanding of how AI works, its capabilities, and its limitations through education and training is vital for fostering informed trust and responsible use.

The AI Trust Paradox highlights a critical inflection point. The widespread adoption of AI depends not just on its technical capabilities but on its ability to earn and maintain the trust of users worldwide. By addressing the challenges of accuracy, transparency, and safety, and by fostering a better understanding of AI, we can pave the way for a future where AI is not only a useful tool but a truly trusted partner.