AI Will Become Ordinary. Then It Will Meet Its Limits.

The public conversation about artificial intelligence is currently dominated by two extremes. In one, AI is about to replace vast numbers of workers and reorganise almost every profession. In the other, it is a productivity tool that will simply make existing jobs easier. Both views assume, in different ways, that the extraordinary rate of adoption seen since the arrival of generative AI can continue almost without limit.

That assumption deserves more scrutiny. AI may indeed become as normal as the computer or the mobile phone. In fact, that is probably the most plausible long-term outcome. Once a general-purpose technology becomes embedded in ordinary software, people stop thinking of every interaction as a separate technological event. We do not say that we are “using a computer” each time we send an email or check a bank account. AI may disappear into the same background infrastructure.

But normalisation is not the same thing as endless expansion. The current phase is being driven by novelty, falling prices, corporate competition and the rapid discovery of new uses. Eventually, AI will have to operate within physical and social limits. The first is energy. The second is water. The third is regulation. The fourth, and potentially the most underestimated, is human trust.

The result may not be a dramatic collapse in AI use. It may be something more ordinary: a peak in the era of indiscriminate AI, followed by a mature period in which the technology is used where it creates enough value to justify its economic, environmental and social cost.

AI is more likely to transform jobs than simply erase them

The employment debate is a useful place to begin because it shows how quickly technological discussion can become binary. The International Labour Organization’s 2025 global assessment found that one in four workers is employed in an occupation with some exposure to generative AI. Yet its central conclusion was not that one in four jobs will disappear. Because many occupations still require human input, the ILO concluded that transformation is more likely than wholesale replacement.[1]

That distinction fits the history of general-purpose technologies. Spreadsheets did not remove accountants, but they radically changed what junior accountants spent their time doing. Computer-aided design did not remove architects and engineers, but it eliminated whole categories of manual drafting. Smartphones removed some devices and occupations while creating others, and eventually became so ordinary that the economic question shifted from whether people would adopt them to how they would be governed.

AI is likely to follow a similar path. Routine drafting, transcription, summarisation, basic research, coding assistance and first-stage analysis will increasingly be automated. Some roles will shrink. Others will change. The premium will move towards judgement, client relationships, verification, accountability and tasks that require knowledge of context rather than the production of a first draft.

That future is less dramatic than a “job apocalypse”, but potentially more consequential. AI may not replace the professional. It may replace the part of the professional’s day that is easiest to standardise.

The future constraint on AI may not be the energy used by one prompt. It may be what happens when billions of cheap prompts become trillions of continuous decisions.

Sources and notes:

1. International Labour Organization, “Generative AI and jobs: A 2025 update”, 20 May 2025. The ILO estimates that one in four workers worldwide is in an occupation with some GenAI exposure, but says transformation rather than replacement is the most likely outcome.
Available at: https://www.ilo.org/publications/generative-ai-and-jobs-2025-update

2. International Energy Agency, Energy and AI, April 2025. Data centres used about 415 TWh in 2024, around 1.5% of global electricity; the IEA base case projects about 945 TWh by 2030. The report also estimates electricity-related data-centre emissions at around 180 Mt CO2 today, rising to about 300 Mt by 2035 in the base case.
Available at: https://www.iea.org/reports/energy-and-ai/executive-summary

3. Central Statistics Office Ireland, Data Centres Metered Electricity Consumption 2025, 7 July 2026. Data centres accounted for 23% of metered electricity consumption in Ireland in 2025.
Available at: https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2025/

4. UK Parliamentary Office of Science and Technology, “What are data centres and how sustainable are they?”, 2026. Drawing on IEA estimates, POST reports global data-centre water consumption of around 560 billion litres a year. Related IEA-based evidence projects about 1.2 trillion litres by 2030.
Available at: https://post.parliament.uk/research-briefings/post-pn-0762/

5. International Energy Agency, “Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions”, 16 April 2026. The IEA reports a 17% increase in data-centre electricity use in 2025 even as energy use per AI task fell rapidly; AI-focused centres grew faster and AI agents are among energy-intensive emerging uses.
Available at: https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions

6. Christopher C. M. Kyba et al., “Artificially lit surface of Earth at night increasing in radiance and extent”, Science Advances 3(11), 2017. The study found a 2.2% annual increase in the artificially lit area of Earth between 2012 and 2016 and discussed rebound from more efficient lighting.
Available at: https://pmc.ncbi.nlm.nih.gov/articles/PMC5699900/

7. Singapore Infocomm Media Development Authority, Green Data Centre Roadmap and related 2022 policy material. Singapore paused new data-centre growth in 2019 before moving to a more selective, resource-conscious capacity allocation model.
Available at: https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeches/press-releases/2022/launch-of-pilot-data-centre---call-for-application-to-support-sustainable-growth-of-dcs

8. Commission for Regulation of Utilities, Ireland, “CRU Publishes its Decision on New Electricity Connection Policy for Data Centres”, 12 December 2025. New data centres must provide matching generation/storage and meet at least 80% of annual demand with additional Irish renewable projects, subject to a glide path.
Available at: https://www.cru.ie/about-us/news/the-cru-publishes-its-decision-on-new-electricity-connection-policy-for-data-centres/

9. European Union, Directive (EU) 2023/1791 on energy efficiency, Article 12. Member States must require qualifying data-centre owners and operators to make specified energy and sustainability information publicly available.
Available at: https://eur-lex.europa.eu/eli/dir/2023/1791/oj?locale=en

10. Australian eSafety Commissioner, “Social media age restrictions”. From 10 December 2025, specified platforms must take reasonable steps to prevent Australians under 16 from creating or keeping accounts.
Available at: https://www.esafety.gov.au/about-us/industry-regulation/social-media-age-restrictions

11. UK Department for Science, Innovation and Technology, “Social media to be banned for under-16s…”, 15 June 2026. The Government announced plans to prevent social-media companies from offering covered services to under-16s, with implementation expected in 2027 subject to legislation.
Available at: https://www.gov.uk/government/news/social-media-to-be-banned-for-under-16s-in-landmark-government-move-to-givekids-their-childhood-back

12. University of Melbourne and KPMG, Trust, attitudes and use of artificial intelligence: A global study 2025. Survey of more than 48,000 people in 47 countries: 66% used AI regularly, 46% were willing to trust AI systems and 70% believed regulation was needed.
Available at: https://kpmg.com/xx/en/our-insights/ai-and-technology/trust-attitudes-and-use-of-ai.html

13. Berkeley J. Dietvorst, Joseph P. Simmons and Cade Massey, “Algorithm aversion: people erroneously avoid algorithms after seeing them err”, Journal of Experimental Psychology: General 144(1), 2015.
Available at: https://pubmed.ncbi.nlm.nih.gov/25401381/

14. “On preferring people to algorithms”, Journal of Risk and Uncertainty, 2025. Across multiple decision scenarios, respondents generally preferred human decision-makers, particularly in more human-centred contexts, although preferences shifted with information and experience.
Available at: https://doi.org/10.1007/s11166-025-09466-6

15. Solicitors Regulation Authority, Risk Outlook: The use of artificial intelligence in the legal market. The SRA states that firms remain responsible and accountable for AI outputs and should ensure clients are suitably informed about AI involvement.
Available at: https://guidance.sra.org.uk/sra/research-publications/artificial-intelligence-legal-market/

16. Institute of Chartered Accountants in England and Wales, guidance on using generative AI and “AI and accountants: the rules and guidance you need to follow”. ICAEW advises transparency, professional review and continued human responsibility for AI-assisted work.
Available at: https://www.icaew.com/technical/practice-resources/practice-news/ai-and-accountants

17. Christoph Fuchs, Martin Schreier and Stijn M. J. Van Osselaer, “The Handmade Effect: What’s Love Got to Do with It?”, Journal of Marketing 79(2), 2015. Experimental studies found a positive effect on product attractiveness when items were described as handmade.
Available at: https://journals.sagepub.com/doi/10.1509/jm.14.0018

Figures refer to the data-centre sector as a whole, not AI alone. AI is, however, the principal driver of projected growth in accelerated computing.

There is no AI without electricity

The most important constraint on AI is not philosophical. It is physical. Training and running modern models requires servers, and servers require electricity.

The International Energy Agency estimated that data centres consumed about 415 terawatt-hours of electricity globally in 2024, around 1.5 per cent of all electricity used in the world. That figure covers all data-centre activity, not AI alone, but the IEA identifies AI as the most important driver of the coming increase. Its central projection has data-centre electricity use more than doubling to about 945 TWh by 2030, slightly more than the total electricity consumption of Japan today.[2]

Put differently, 415 TWh is an average continuous load of roughly 47 gigawatts throughout the year. By 2030, 945 TWh would imply an average load of about 108 GW. A typical AI-focused data centre can consume as much electricity as 100,000 households, according to the IEA, while the very largest facilities now under construction could consume around twenty times that amount.[2]

The global percentage can sound modest. The local effect is not. Ireland offers the clearest warning. Data centres used 23 per cent of the country’s metered electricity in 2025, up from only 5 per cent in 2015.[3] A technology that is a small fraction of global demand can become one of the largest loads on a particular national or regional grid.

Water is the quieter constraint

Electricity attracts most attention because it is measured continuously and appears directly in power-system forecasts. Water is easier to overlook.

The IEA estimates that data centres consume roughly 560 billion litres of water a year worldwide, taking account of direct cooling and important indirect uses such as electricity generation. That is about 224,000 Olympic-sized swimming pools. By 2030, the total could reach about 1.2 trillion litres a year, equivalent to roughly 480,000 Olympic pools.[4]

Individual facilities make the issue more tangible. A 100 MW hyperscale data centre can use around two million litres of water a day under typical conditions, according to figures cited from the IEA. The pressure is particularly significant where facilities cluster in dry regions or where public water systems are already under stress.[4]

The important point is not that AI will somehow “use up” the world’s water. Agriculture and other sectors are far larger users. The problem is geography. Data centres are concentrated, and their demand can arrive very quickly. A new hyperscale campus may compete for the same electricity and water capacity required by households, industry and future housing growth.

Efficiency will help, but it may not reduce total demand

The technology industry’s answer is that each generation of chips, models and cooling systems will become more efficient. That is true, and the rate of progress is remarkable.

In April 2026, the IEA reported that electricity use per AI task was declining at a rate it described as unprecedented in energy history. Yet overall data-centre electricity demand still rose by 17 per cent in 2025, while demand from AI-focused facilities grew even faster. The reason was simple: more people were using AI, and more computationally intensive applications, including AI agents, were spreading.[5]

This is a familiar economic phenomenon. When a technology becomes cheaper and more efficient, people often use more of it. Economists call the broader family of such effects “rebound”. In its strongest form it is associated with the Jevons paradox: efficiency reduces the cost of using a resource so much that total use rises rather than falls.

Lighting provides an unusually intuitive example. LEDs and other efficient lamps dramatically reduced the electricity needed to produce a unit of light. Yet satellite observations between 2012 and 2016 found that the Earth’s artificially lit outdoor area grew by about 2.2 per cent each year, while already-lit areas also became brighter. The researchers explicitly warned that the efficiency revolution in lighting was being partly offset by increased use.[6]

AI may exhibit the same dynamic. A smaller model may use one-tenth of the electricity per query, but if people make fifty times as many queries, run autonomous agents continuously, generate hours of personalised video or embed inference into every appliance and service, aggregate demand still rises.

This is why efficiency alone is unlikely to provide the eventual limit. It changes how much AI society can afford to use. It does not decide how much society will choose to use.

From useful technology to ambient consumption

The first wave of digital technology was often justified by necessity or productivity. Computers processed accounts, designed buildings and managed databases. The internet connected information. Mobile phones allowed communication away from a fixed line. Once those infrastructures became cheap and ubiquitous, use expanded into entertainment, gaming, video streaming, social media and constant background connectivity.

There is nothing inherently wrong with this transition. Entertainment is a legitimate use of energy. But it matters for forecasting because the number of potential AI applications is not bounded by the number of productive office tasks. Synthetic video, personalised games, virtual companions, automated shopping, continuous translation, recommendation engines and agents that perform thousands of machine-to-machine actions could produce a far larger compute market than today’s visible chatbots.

The present debate often assumes that society will first discover what AI can do and then use all of it. That is not how mature infrastructure systems normally work. Eventually, scarcity, prices and rules distinguish between applications that create enough value and applications that do not.

Governments have already started rationing the infrastructure around AI

This process is no longer theoretical. Governments are already discovering that data centres cannot be treated like an unlimited, placeless digital service.

Singapore temporarily paused new data-centre growth in 2019 while it reconsidered the sector’s use of scarce land, power and water. It later reopened growth through a more selective approach and a Green Data Centre Roadmap that links new capacity to efficiency and greener energy.[7]

Ireland has gone further. After data centres rose to more than a fifth of national metered electricity demand, the energy regulator introduced a new connection policy requiring new facilities to provide matching generation or storage and to meet at least 80 per cent of annual demand through additional renewable electricity projects, subject to a transition period.[8]

The European Union now requires operators of data centres above a specified IT power threshold to publish energy and sustainability information under the Energy Efficiency Directive.[9] These are early examples of a broader principle: digital infrastructure is beginning to be governed as infrastructure, with physical resource obligations rather than simply as a weightless technology industry.

Future controls could take many forms: stricter grid-connection tests, water-use disclosure, minimum energy-efficiency standards, locational planning policies, higher charges at constrained times, requirements for renewable generation, heat reuse, or differentiated rules for very large compute facilities. Governments may never ration individual AI prompts. They can still shape the price and availability of the infrastructure that produces them.

Technology is often regulated after society learns its externalities

This would not be unusual. Many technologies pass through a period in which adoption is treated as an uncomplicated public good, followed by a period in which government begins managing side-effects.

Cars brought mobility and economic freedom, but governments now use congestion charges, parking controls, low-emission zones, pedestrianisation and investment in walking and cycling to manage traffic, pollution and urban space. Social media expanded rapidly with little age-specific regulation; Australia now requires major platforms to prevent under-16s from holding accounts, while the UK announced in June 2026 that it intends to introduce a similar under-16 restriction, expected to take effect in 2027.[10][11]

The point is not that AI should be treated like cars or social media. It is that societies rarely allow the external costs of a successful technology to grow indefinitely simply because the technology is useful. Once the costs become visible, the political question changes from “should this technology exist?” to “under what conditions should it operate?”

→ All Topics

The other ceiling may be trust

Physical resources are only one reason AI use may stabilise. The other is psychological.

A large 2025 University of Melbourne and KPMG study of more than 48,000 people in 47 countries found a striking gap between use and trust. Sixty-six per cent said they regularly used AI, but only 46 per cent were willing to trust AI systems. Seventy per cent believed regulation was needed. The study also found widespread reliance on AI without checking accuracy and reported work-related mistakes associated with its use.[12]

Trust may not move in a straight line with capability. Classic research on “algorithm aversion” found that people can lose confidence in an algorithm more quickly than in a human after seeing each make a mistake, even where the algorithm is statistically more accurate overall.[13] More recent research shows the picture is context-dependent: people are often comfortable with algorithms for objective, quantitative tasks, but show a stronger preference for human judgement where decisions affect health, welfare, relationships or other human-centred matters.[14]

That creates a potential ceiling that engineers cannot solve simply by reducing hallucination rates. People may tolerate a machine error in a restaurant recommendation. They may react differently when the error appears in a tax return, planning opinion, legal submission, medical decision or financial report.

“Human reviewed” may become a commercial assurance

Professional regulation is already moving in this direction. The Solicitors Regulation Authority has said that law firms remain responsible for AI outputs and that clients should be appropriately informed about how AI is involved in their case. An error produced by a chatbot remains the firm’s error.[15]

The Institute of Chartered Accountants in England and Wales similarly advises accountants to be transparent about generative AI, to label or acknowledge AI-produced output and to review, challenge and verify it with the same professional scepticism applied to other work. Its updated guidance stresses that AI does not replace professional judgement or accountability.[16]

It is not difficult to imagine the market taking the next step. Clients may begin requesting an explicit assurance: “This document has been prepared or reviewed by a qualified human professional.” Such a statement would not mean that AI was absent. It would mean that a person with a name, qualification and professional duty had taken responsibility for the final product.

For lawyers, accountants, planners, engineers, doctors and other regulated professions, that may become more valuable as AI use becomes more widespread. AI can produce the first 80 per cent quickly. The commercial product may increasingly be the human taking responsibility for the final 20 per cent.

The return of the human premium

There is also a less rational but equally powerful possibility: people may simply begin to value human involvement because it is human.

Industrial production did not eliminate demand for handmade goods. Research in consumer behaviour has identified a measurable “handmade effect”, in which buyers value products more highly when they are told that a person made them, even when machine-made alternatives may be cheaper or technically consistent.[17] Similar preferences explain part of the appeal of artisan food, craft production and other markets built around provenance and authenticity.

AI could create an equivalent category. A human-written book, human-produced illustration, human-led customer service line or human-reviewed professional report may become a premium product precisely because automated production is cheap and abundant. Scarcity changes value. When machine output is everywhere, evidence of human attention can itself become a selling point.

This is not an argument that human work is inherently superior. Handmade furniture can be worse than factory-made furniture, and a human adviser can be less accurate than a well-designed model. The point is that markets price more than technical performance. They also price trust, accountability, authenticity and experience.

So when does AI peak?

There is no evidence that global AI use is about to peak in the next few years. The opposite is true. The IEA expects data-centre electricity use to continue rising strongly to 2030, and industry investment remains enormous.[2][5] Any claim of an imminent collapse in AI demand would therefore be difficult to support.

The more plausible argument concerns the shape of the long-term curve. The current exponential-looking phase is unlikely to continue indefinitely. Some applications will prove valuable and become permanent. Others will disappear once novelty fades. Efficiency will lower the cost of many tasks, but rebound will create new demand. Grid and water constraints will make compute more expensive in some places. Regulation will place conditions on infrastructure. Professional liability will preserve human oversight in high-stakes services. Consumers may pay for human involvement in areas where authenticity matters.

What peaks first may therefore be neither capability nor total computation. It may be the assumption that more AI is automatically better.

After that point, society will become more selective. A hospital may use AI heavily for image analysis but insist on a human clinician signing the decision. A law firm may automate research but warrant that advice has been reviewed by a solicitor. A public authority may use AI to process applications but maintain a human route for contested decisions. A consumer may happily accept automated translation while paying more for a human-written wedding speech.

This is how mature technology usually looks: not rejected, not worshipped, but allocated.

AI as a utility, not a religion

The debate about AI often treats the technology as though society must choose between two futures: surrender work to machines or resist the machines. A more likely future is less dramatic.

AI will become infrastructure. It will sit inside word processors, search engines, design software, public services, medical systems and accounting platforms. Many workers will use it without consciously deciding to “use AI”. As with computers and smartphones, the technology will become less culturally visible precisely as it becomes more embedded.

At the same time, physical reality will remain stubborn. Servers need power. Cooling needs water or alternative systems that may carry other energy costs. Grids need capacity. New generation takes years to build. Communities have competing claims on land, electricity and water. Carbon emissions do not disappear because the service being powered is digital.

Social reality will be equally stubborn. People want someone to blame when professional advice is wrong. They want a human voice when a decision affects them personally. They value authenticity even when machines can reproduce the appearance of it. And governments intervene when the aggregate effect of millions of individually rational choices becomes a public problem.

The future of AI may therefore be neither unlimited expansion nor retreat. It may be normalisation followed by discipline. The technology will become better. We will become better at deciding when it is worth using.