In June 2023, a lawyer in New York faced sanctions by a US federal judge after submitting a legal brief containing fabricated case citations generated by ChatGPT.
The lawyer had not invented the cases; he had simply trusted the machine that did. Though initially treated as a curiosity and an early cautionary tale from the frontier of generative Artificial Intelligence (AI), the episode has since become a preview of a much larger problem.
AI promised to make work faster, and, in many respects, had delivered on that promise. Yet, like a double-edged sword, these productivities come with a hidden cost: the human labour of checking, correcting and second-guessing AI outputs.
Race against invisibility
Speaking at Kaspersky’s Cybersecurity Weekend 2026, the company’s Managing Director for Asia Pacific, Adrian Hia, said organisations across the region are grappling with a “visibility gap” as AI reshapes both business operations and the threat landscape.
Hia noted that cybersecurity has moved from a race against time to a race against invisibility, as AI-driven speed and connectivity widen blind spots across IT and operational technology environments.
Kaspersky said that it detected and blocked around half a million unique malicious files daily in 2025, a 7 per cent rise from 2024. The firm also identified more than 15,000 malware samples disguised as agentic AI software this year alone.
Sojun Ryu, Senior Security Researcher at Kaspersky’s Global Research and Analysis Team, warned that as AI systems increasingly plan, deploy tools and act independently, the decision to extend trust still rests with humans. When speed outpaces verification, he said, that speed also amplifies risk. In other words, “trust but verify” remains as relevant as ever.
The scale of the problem is significant. Kaspersky’s Compromise Assessment division found that 31 per cent of analysed security incidents were active for more than three months before discovery, while 52 per cent of high-severity compromises went unnoticed for over 90 days. One incident, the company said, had gone undetected for four years.
Botsitting: New category of work
The verification burden is not confined to security teams. A recent study found employees save roughly 11 hours a week using AI tools, yet spend more than six hours checking outputs, correcting errors and filling in missing context. The practice has acquired an informal name: botsitting.
Research from BetterUp Labs and Stanford found that 41 per cent of workers had encountered low-quality, AI-generated content that required rework, contributing to slower output and eroding trust between colleagues. As a result, productivity gains are often offset — the devil, it seems, is in the details.
A separate industry report, State of AI in Business 2025, found that despite investment of between 30 and 40 billion US dollars in enterprise investment in generative AI, 95 per cent of organisations were seeing no measurable return.
When the machine gets it wrong
The consequences of insufficient oversight are already visible. AI-detection firm GPTZero identified fabricated claims and false footnotes in four reports produced by professional services firm PwC, and warned that popular AI chatbots had begun repeating the false information because it carried a recognised industry name, according to a report by Reuters news agency.
GPTZero co-founder Alex Cui put it bluntly: Such errors reflected a lack of the human verification typically expected from major professional services firms. “Such errors betray a lack of care and human verification that many come to expect from work produced by the Big Four.”
Deloitte Australia was required to partially refund the 440,000 Australian dollars ($290,000) paid by the Australian government for a report that was littered with apparent AI-generated errors, including a fabricated quote from a federal court judgment and references to non-existent academic research papers, according to the Associated Press.
The financial services firm’s report to the Department of Employment and Workplace Relations was originally published on the department’s website in July 2025. A revised version was published on October 4, 2025, after Chris Rudge, a Sydney University researcher specialising in health and welfare law, said he alerted the media that the report was “full of fabricated references.”
Deloitte had reviewed the 237-page report and “confirmed some footnotes and references were incorrect,” the department said in a statement.
Cui noted that similar reviews of EY and KPMG reports had previously led to retractions, costing staff time and damaging client confidence. The incidents underscore the risks of treating AI output as something to rubber-stamp rather than rigorously verify.
The money question
Bob Michaels, a partner at CrossCountry Consulting, said corporate AI spending is escalating largely unchecked, and that because accounting rules generally do not allow companies to spread software costs over time, much of it hits earnings immediately.
He said many companies bought AI tools first and are only now trying to work out how to measure their value, leaving finance teams struggling to retrofit return-on-investment calculations onto money already spent.
Also Read: World Bank Urges Developing Countries to Embrace AI or Be Left Behind
Michaels said genuine returns should be measured through metrics such as time saved, shortened processes and more efficient output, rather than being assumed. He cautioned against stripping AI costs out of adjusted earnings to flatter results, arguing the spending is permanent, and investors will not accept a “one-off cost” explanation indefinitely. Even mid-sized companies, he said, are now spending between$5 and 10 million a year on AI tools and infrastructure.
He added that employees are often using AI daily with little training and no consistent standard for what should be double-checked, a gap that, in his view, is likely to create new job categories built around scrutinising AI-generated work rather than eliminating existing ones. Instead of replacing people outright, AI may simply shift where human effort is required, ensuring there is always a “human in the loop.”
Brain fry: The human toll
There is also a cognitive cost. Researchers have begun describing a form of mental fatigue specific to human-AI collaboration as “AI brain fry”, distinct from conventional burnout in that it stems from cognitive overload rather than emotional strain.
Early findings suggest around 14 per cent of AI users report symptoms including headaches, difficulty concentrating and slower decision-making.
Workers reporting these symptoms also recorded 33 per cent higher decision fatigue and made 39 per cent more major errors on average, according to the same research. They were also 39 per cent more likely to consider leaving their jobs.
The underlying explanation draws on cognitive load theory: human working memory is finite, and AI can generate information faster than people can evaluate it, echoing the “Zoom fatigue” first documented during the Covid-19 pandemic.
Across cybersecurity, corporate finance and individual workplaces, a consistent pattern emerges. AI increases the volume and speed of output, but verifying that output remains a human task, and that task is expensive, whether measured in security incidents left undetected for months, in accounting write-offs, or in worker fatigue. As organisations race to embrace AI, they may ultimately discover that someone still has to pay the piper. The technology can accelerate work, but trust cannot be automated.
