Do you trust AI? Here’s why half of users don’t

A recent, global KPMG and University of Melbourne study revealed that half of those surveyed said they don’t trust AI to provide them with accurate responses.

Titled “Trust, attitudes and use of artificial intelligence,” the study involved a survey 48,340 people in 47 countries to explore the use of and views on AI.

The survey revealed that 54% of respondents are “wary,” especially about the safety and societal impact of AI.

Despite mixed feelings, 72% accept AI as a useful technical tool. Trust and acceptance are lower in advanced economies (39% trust, 65% accept) vs. emerging economies (57% trust, 84% accept).

Part of the unease with AI appears to stem from a lack of training. Only 39% of survey respondents report having some form of AI training, whether at work, school, or independently. Not surprisingly, nearly half of respondents (48%) say they have little knowledge or understanding of AI. Those with AI training see more efficiency (76% vs. 56%) and revenue gains (55% vs. 34%), and managers benefit more than other roles in both areas.

The KPMG/Melbourne University study also found:

Most people (70%) support AI regulation, and only 43% think current laws are adequate. There’s strong demand for international regulation (76%), national regulation (69%), and co-regulation by industry, government, and other regulators (71%), with 88% of respondents saying laws are needed to combat AI-driven misinformation.

At work, 58% of employees regularly use AI, mostly free generative AI (genAI) tools. Over half of respondents report performance gains, but many see negative effects on workload, teamwork, and compliance. Misuse and lack of oversight are common, with governance and training lagging behind adoption.

In education, 83% of students use AI for efficiency and stress reduction. However, misuse is widespread, raising concerns about over-reliance and fairness. Only half say their schools offer proper policies or training for responsible AI use.

The findings were similar in Hitachi Vantara’s State of Data Infrastructure report released last year that identified “a critical” AI trust gap: just 36% of IT leaders regularly trust AI outputs, while only 38% of organizations are improving their training data quality.

There’s good reason for this mistrust, as these systems are prone to errors and hallucinations (things the models make up but present as facts). Recent testing of genAI models shows they’re even willing to override human instructions and then lie about it.

Hallucinations getting worse

Jason Hardy, CTO at Hitachi Vantara, called the trust gap “The AI Paradox.” As AI grows more advanced, its reliability can drop. He warned that without quality training data and strong safeguards, such as protocols for verifying outputs, AI systems risk producing inaccurate results.

“A key part of understanding the increasing prevalence of AI hallucinations lies in being able to trace the system’s behavior back to the original training data, making data quality and context paramount to avoid a ‘hallucination domino’ effect,” Hardy said in an email reply to Computerworld.

AI models often struggle with multi-step, technical problems, where small errors can snowball into major inaccuracies — a growing issue in newer systems, according to Hardy.

With original training data running low, models now rely on new, often lower-quality sources. Treating all data as equally valuable worsens the problem, making it harder to trace and fix AI hallucinations. As global AI development accelerates, inconsistent data quality standards pose a major challenge. While some systems prioritize cost, others recognize that strong quality control is key to reducing errors and hallucinations long-term, he said.

In a concerning trend, recent tests show that hallucinations are on the rise in newer AI reasoning systems, spiking as high as 79% in one test, according to The New York Times.

In fact, the Artificial Intelligence Commission (AIC) — a Washington, DC-based organization dedicated to promoting responsible AI development and deployment — recently reported that AI hallucinations are getting worse, not better.

Tests by OpenAI creator ChatGPT revealed that its newest o3 and o4-mini reasoning models hallucinated much of the time. The company found the o3 model hallucinated 33% of the time during its PersonQA tests, in which the bot is asked questions about public figures. The o3 model hallucinated 51% of the time on SimpleQA tests, which ask short fact-based questions.

The smaller, faster o4-mini model did worse, hallucinating 41% of the time on PersonQA and 79% on SimpleQA. The newer GPT-4.5 model, released in February, performed better, with a 37.1% hallucination rate on SimpleQA. OpenAI publishes the latest results of these and other tests on its Safety evaluations hub.

“The increase in hallucinations by reasoning models may very well be due to AI overthinking,” said Brandon Purcell, a vice president and principal analyst at Forrester Research.

Forrester’s research data aligns with other scrutiny into AI trust: Over half of business leaders worry about generative AI, slowing adoption and limiting its value, according to the research firm. To close this trust gap, companies should demand transparency, invest in explainable and traceable AI, and monitor performance in real time, Purcell said.

In fact, hallucinations are “a feature of large language models, not a bug,” Purcell said.

“While we don’t know exactly how LLMs work, chances are the training data itself is not stored in the model. The model is just a representation of the statistical patterns in the training data,” Purcell said. “If you want to reduce hallucinations, you need to ground a model in a correct and current canonical data set using retrieval augmented generation or another technique that finds the answer from a source that is external to the model.”

The problem is that large language reasoning models follow multi-step processes, so small early errors can lead to hallucinations. As questions are repeated, the hallucinations can become even worse and more bizarre. LLMs, Purcell argued, are best used for reasoning, while smaller models are better suited for fact-based Q&A.

That’s why many believe the future of AI is small, not large models.

SLMs to the rescue

In 2025, small language models will likely come into their own, as enterprises increasingly deploy them to address specific tasks without overburdening data center processing and power. In the coming year, SLM integration could surge by as much as 60%, according to a Forrester report.

A recent Capital One survey of 4,000 business leaders and technical practitioners across industries found that while 87% believe their data ecosystem is ready for AI at scale, 70% of technologists spend hours daily fixing data issues.

Three out of four (75%) IT-decision makers believe SLMs outperform LLMs in speed, cost, accuracy and ROI, according to a Harris Poll of more than 500 users commissioned by the startup Hyperscience.

As Hitachi’s Hardy noted, the quality of data fed into AI models is also key to their accuracy.

“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,” said Andrew Joiner, CEO of Hyperscience, which develops AI-based office work automation tools. “The true potential…lies in adopting tailored SLMs, which can transform document processing and enhance operational efficiency.”

Forrester’s Purcell also recommends that businesses “thoroughly test AI” before, during, and after deployment — using humans or AI for red teaming. High-stakes systems, like medical AI, should first be validated in simulations, similar to how “autonomous vehicles are tested,” he said.