Clearly, last year’s fears of a stockmarket “AI bubble” have long gone. AI firms and related stocks are now being treated as the backbone of American prosperity, and market saviours. At the same time, public perceptions of AI are heading in the opposite direction. Despite the daily miracles of AI wizardry, the technology is increasingly being seen as:
(a) the potential destroyer of white-collar job security
(b) the lawless plunderer of written knowledge and creativity
(c) the enabler of surveillance, privacy intrusions and other forms of state control on a massive scale.
Then there are the environmental harms. To cool down their componentry, data centres are draining ancient aquifers and siphoning off oceanic amounts of the world’s scarce water reserves that communities depend on to live, and to grow their food. Hovering over all of this is the existential threat that quasi-sentient, ultra-logical AI agents may pose to humanity itself.
This doesn’t mean that AI opponents are Luddites. As others have said, it seems entirely reasonable to be concerned about a process whereby more and more workers are having to act like machines, in order that machines can learn to act more like people.
Regardless, otherwise sensible people seem to be convinced that AI could be the answer to New Zealand’s chronic low rates of workplace productivity. Allegedly, it is only if we lift our r&d game and invest in new technology – and what’s newer than AI? – can the nation hope to become more productive, richer, happier and healthier.
Before New Zealand heads on down that path, a few cautionary points bear consideration. For instance: the links between AI and productivity are proving to be more complex, and less conclusive than initially thought. Early expectations that AI will improve work outputs by stripping out the humdrum detail (from writing emails to debugging coding, to doing basic legal work and journalism) have proved to be only one dimension of the workplace story.
In tech industries and law firms for example, AI doesn’t seem to be freeing up human labour at all. If and when AI enables work to be done faster, this merely fosters higher expectations of output. Only an economist would believe that there are no personal or social harms involved in AI enabling the workplace mouse-wheel to be spun even faster. Evidence is accumulating that AI-assisted workers are at higher risk of developing addictive, workaholic behaviours, and burnout.
Even the speed gains in task completion can be somewhat illusory. Any initial gains in efficiency and speed are being tempered by the realisation that AI tools create substantial new demands for managing and supervising their work.
Meaning: instead of freeing up human labour in the utopian way envisaged long ago by the economist John Maynard Keynes – who once famously predicted that technological advances would soon enable people to work for only 15 hours a week – AI appears to intensify the nature of work. Given AI’s inherent error rates, AI tools cannot be relied on, much less left to their own devices, particularly in those areas of work where accuracy is important. For example: a recent empirical study has shown that all of the three leading legal system AI research tools hallucinate between 17% and 33% of the time.
Similarly within the tech industry, senior software developers are reportedly spending a significant part of their time as “AI babysitters.” In mid 2025, research carried out by the San Francisco cloud computing company Fastly mirrored the mixed results :
“AI will bench test code and find errors much faster than a human, repairing them seamlessly. This has been the case many times,” one senior developer said. A junior respondent noted the trade-offs: “It’s always hard when AI assumes what I’m doing and that’s not the case, so I have to go back and redo it myself.”
Moreover, Fastly found, senior developers were more likely to say they were investing time into fixing AI-generated code:
Just under 30% of seniors reported editing AI output enough to offset most of the time savings, compared to 17% of juniors. Even so, 59% of seniors say AI tools help them ship faster overall, compared to 49% of juniors.
So, in some contexts AI tools do increase output and accelerate the rate at which projects can be completed. But those net efficiency gains occur among only 59% of the senior staff and among less than half of the junior staff. This is hardly a ringing endorsement, especially if the sizeable installation and running costs of the machines are included in the equation.
Moreover, the gains rarely “free up” anyone. As mentioned, AI is creating a new requirement to review and correct what it has done; and nearly one third of senior developers in the Fastly study found this requirement entirely cancelled out the efficiency gains.
Also as mentioned, where AI use was successful, this has been fed into heightened expectations of the outputs that AI-assisted workers can produce, and the fostering of the addictive, workaholic impulses required to achieve them. This appears to be increasing the workplace incidence of what researchers are calling “AI brain-fry.” In the coding context, this can take the form of “vibe coding paralysis.” For those interested, there’s a good account of it available here. After portraying the symptoms, the author concluded:
Vibe coding paralysis is the syndrome of wanting to do so much — and being able to do so much — that you end up finishing nothing… The paradox: the more capability you have, the more you feel compelled to use it. The more you use it, the more fragmented your attention becomes. The more fragmented your attention, the less you actually ship. It’s not burnout in the traditional sense. It’s something weirder — a kind of cognitive overload, masked as productivity.
In another example that went viral last year. Microsoft software engineer Victor Dibia reported on his experience of being an early adopter of Anthropic’s Claude. Like many of his peers, Dibia had assumed AI would free up his labour to a significant extent, first by taking over mundane tasks (with Claude serving much like a junior lawyer in a law firm) and then by freeing him up to be more productive on more important tasks.
What Dibia found instead was that while productivity did increase, so did (a) his workload, largely thanks to his new tasks of managing/correcting the errors in the AI outputs and (b) so did the gradient of the work expected of him. This isn’t new. In economics, it is called the Jevons Paradox.
Back in 1865, William Stanley Jevons first identified the paradox that steam engines that were more efficient were not reducing the world’s dependency on coal, but were actually increasing it, by boosting demand.
He maintained that more efficient steam engines would not decrease the use of coal in British factories but would actually increase it. As the fossil fuel became cheaper, demand for the resource would grow, leading to the construction of more engines.
With enthusiasm, Microsoft CEO Satya Nadella has invoked Jevons Paradox with respect to the potential he sees for unbridled AI growth. (Nadella doesn’t mention the consequences for the workforce, or for the environment.) Further down the Microsoft pay ladder, Victor Dibia cited a cautionary example from the 1950s of the Jevons Paradox :
Sociologist Ruth Schwartz Cowan documented this extensively in her book More Work for Mother (1983), showing how “labour-saving” devices paradoxically maintained or even increased the housewife’s workload throughout the 20th century. The technology made each individual task easier, but social standards and expectations expanded to fill (or exceed) the time saved.
The incentives to ignore AI’s fallibility
As Gautam Mukunda of the Yale School of Management has wryly pointed out to Bloomberg News, AI machines are often wrong, but they are never uncertain. Reportedly, “I don’t know” responses result in the Large Language Models (LLMs) being marked down in market evaluations. As OpenAI itself recently admitted:
Our new research paper …. argues that language models hallucinate because standard training and evaluation procedures reward guessing over acknowledging uncertainty.
The desire to deny the existence of AI fallibility could also explain why OpenAI is also backing draft legislation in Illinois that will limit the liability of AI labs if and when AI-enabled results cause or contribute to mass numbers of deaths, or to a financial meltdown.
AI’s unfounded certainty should be problematic for the likes of law firms, where accuracy – not certainty, or speed – is supposedly crucial to the firm’s reputation. As Mukunda recently explained on Bloomberg News, there is a fundamental risk involved when AI interacts with the standard business model operated by many service firms:
…All professional service firms, even ones like Boston Consulting Group, Goldman Sachs Group Inc., and the Big Four accountants, have a similar business model – they have partners who oversee teams of associates and check their work. The more associates per partner, the more the partners earn. That’s what’s known in professional services as leverage, and it’s so important that law firms are actually ranked by profits per partner.
But this model means they’re all going to struggle to capture the economic gains artificial intelligence seems to promise, because partners’ ability to supervise and verify AI-enabled work will become the rate-limiting step to the firms’ growth.
Added to the pressure to put speed over accuracy is the cost of the AI tools, which companies have been sold on by the promise of heightened productivity. Yet as Mukunda went on to explain, there’s also another factor at work. In the kind of firms where people are paid via outputs measured in strictly measured time segments – e.g. law firms, or accountancy firms – the remuneration system does not reward (or promote) the staff who insist on checking the AI’s bold assertions, and getting it right:
Although productivity gains from AI show up immediately, the risks from missing a hallucination may not show up for years, if at all. This creates a principal-agent problem, because there will be a constant temptation for the partners setting policy to reap the gains and not pay enough attention to the risks.
Compensation committees reward visible outputs. A partner who catches three AI hallucinations before they reach a client has decreased their short-run productivity. A partner who pushes AI-assisted work through the pyramid faster looks better. The risk is that firms will select for the partners most enthusiastic about AI, not the ones best at managing it.
No doubt, AI is a shining example of human ingenuity. Quite accidentally though, it is also exposing the social, ethical and environmental limits of unfettered capitalism. AI’s advantages also come hand in hand with the risk of financial/reputational downsides for the companies that use the technology without having guard-rails in place, and – ultimately – without there being much in the way of oversight by the state. As Mukunda concludes:
Companies are companies. They will, eventually, be expected to turn a profit. Humanistic goals will become subsumed by data-driven metrics. The idea of doing good brings everyone together — but somehow, “good” ends up a conflicted border, with angry people on either side.
As mentioned at the outset, the public already senses the risks they are facing from the unregulated use of AI. If only the politicians recognised them, and were willing to take protective action.
Footnote One: Evidently, reliance on AI’s Large Language Models ( LLMs) does make you dumber. This MIT study measured the brain functions in three groups – an LLM -assisted group, a search engine assisted group, and a brain-only group – over the course of the writing of an essay. The results:
LLM [compared to] to-Brain participants showed weaker neural connectivity and under-engagement of alpha and beta networks; and the Brain-to-LLM participants demonstrated higher memory recall, and re-engagement of widespread occipito-parietal and prefrontal nodes, likely supporting the visual processing, similar to the one frequently perceived in the Search Engine group.
Moreover, the MIT researchers found :
The reported ownership of LLM group’s essays in the interviews was low. The Search Engine group had strong ownership, but lesser than the Brain-only group. The LLM group also fell behind in their ability to quote from the essays they wrote just minutes prior….Over the course of 4 sessions…the LLM group’s participants performed worse than their counterparts in the Brain-only group at all levels: neural, linguistic, scoring. We hope this study serves as a preliminary guide to understanding the cognitive and practical impacts of AI on learning environments.
Footnote Two: Using AI to tailor your CV and job applications should be a win/win for job seekers and employers alike, right? Instead, recent research by Forbes business magazine suggests that AI fabrications and embellishments of the work skills held by job applicants may be sabotaging the hiring process. Among the Forbes research findings :
*About 84% of HR leaders say their teams are experiencing heavier workloads due to the mass influx of AI-generated and AI-optimized applications.
*An estimated 65% say skills are harder to verify in AI-optimized resumes, while 67% say AI resumes slow down the hiring process.
Pointing the way
The lyrics of some pop songs offer good advice (Don’t fear the Reaper. Be kind to your four footed friend. Relax. Don’t wear sandals.) You can’t hurry love.) Some songs (eg “Que Sera Sera”) also offer really bad advice about parenting. A couple of years ago, the excellent Korean TV series Pachinko resurrected an all but forgotten mid-1960s hit by the Grass Roots. Good song, good advice:
Talking of good advice: From the classic reggae period of the early 1970s, “Ital Correction” by Niney the Observer offers a shopping guide for the healthy vegetables at your local market.