Most public conversation about artificial intelligence swings between two poles. At one end, the machines will end us. At the other, they will save us, and quickly. Both are dramatic stories that leave us with little responsibility for what happens next. If the outcome is apocalypse or utopia, there is little for the rest of us to do but watch.
I find a more immediate question in the evidence: how do we bring a powerful technology into our work and institutions while preserving the judgement and expertise they depend on?
That question takes me back to change, where I have spent most of my working life.
The gap between capability and value
The technology has become much more capable and affordable. Stanford's 2025 AI Index found that the cost of running a model at the level of early ChatGPT fell more than 280-fold in roughly two years. Intelligence, or something that looks enough like it to be useful, is becoming cheap.
Value is a different matter. MIT's NANDA initiative reported in 2025 that, across an estimated US$30 to 40 billion of enterprise spending on generative AI, about 95 per cent of organisations were seeing no measurable return. The researchers did not blame the models so much as the fit: tools that did not learn, did not remember, and did not slot into the way work actually flows.
Economists have been making a similar point from further away. Daron Acemoglu's Simple Macroeconomics of AI estimated a total factor productivity gain of well under one per cent over ten years, because only a fraction of tasks are affected and the gains on each are modest. Closer to home, the Productivity Commission's interim report on data and digital technology is more optimistic, suggesting AI could add more than $116 billion to Australian economic activity over the next decade. Those estimates leave considerable room for uncertainty. What organisations and governments do with the technology will help determine the gains.
I have seen this pattern with every wave of technology I have worked through. A capable tool attracts a flurry of pilots, then runs into established processes, incentives and habits. Progress stalls when using it requires changes to how decisions are made. In my experience, that is usually where the harder work begins.
What happens to our judgement
The second part of the conundrum is more personal, and harder to measure.
In mid-2025, the research group METR ran a randomised trial with experienced open-source developers working in codebases they knew well. With AI tools, they took 19 per cent longer to finish their tasks. Afterwards, they believed the tools had made them about 20 per cent faster. METR is careful to call this a snapshot of early-2025 tools in one setting, and later studies may show something different. What stays with me is how far their experience diverged from the measured result. They felt faster even though the tasks took longer.
A Microsoft Research and Carnegie Mellon survey of 319 knowledge workers points in the same direction. The more people trusted the AI to do a task, the less critical thinking they reported applying to it. The more they trusted themselves, the more they questioned the output.
Using these tools changes the work. People spend less effort producing a draft and more checking whether it is sound. That requires skill and practice, yet organisations can buy licences without giving much thought to how people will develop either.
Work can feel easier without getting better. We still need to judge the quality of what we produce.
The work, the grid and the institutions
The third part of the conundrum is structural.
On work, the most careful evidence does not describe mass replacement. The International Labour Organization's 2025 global index found that about one in four jobs worldwide is exposed to generative AI, and concluded that few jobs consist of tasks that are fully automatable. Changes to existing jobs are the more likely effect, with some workers much more exposed than others. In high-income countries, 9.6 per cent of jobs held by women sit in the highest-exposure category, compared with 3.2 per cent for men, largely because of the concentration of women in clerical and administrative roles. That leaves a practical question about who will carry the cost of the transition and what support they will have.
On energy, the International Energy Agency's Energy and AI report projects that data centre electricity use will more than double, from around 415 terawatt-hours in 2024 to around 945 by 2030, slightly more than Japan consumes today. The same report notes AI could help run energy systems more efficiently. Those potential efficiencies need to be considered alongside the electricity demand. AI runs on physical infrastructure, and decisions about how to power it are already being made.
On institutions, the challenge is pace. Regulation, education, professional standards and procurement rules move in years. AI capabilities can change substantially in months. That difference in pace gives those deploying the technology considerable influence over its use, often before the people affected have had much say.
What the new world asks of us
How well we redesign work around these tools will shape much of what comes next.
Buying the tools is only the beginning. We need to decide which judgements people should retain and give them the practice to keep those skills sharp. Workers in exposed roles need support before their jobs change, and energy demand belongs in the planning from the outset. I also want small organisations, schools and community groups to have access to the capability that large enterprises take for granted. That is much of why I do the work I do at Company31.
The fear that AI will end us can make the danger feel distant from everyday decisions. The quieter conundrum is already here, in the choices we make about training, budgets, working practices and who gets a say. Absorbing this technology well will take sustained attention to those choices.
That is work we have to take responsibility for ourselves.