Promise, power and the reality behind the hype

Artificial intelligence is becoming a general-purpose layer in modern society, but it should not be confused with human-like intelligence. Most current AI systems are statistical models trained to recognise patterns, predict likely outputs and transform information. They can write, classify, translate, generate images, analyse data and assist with software development at remarkable speed. They do not possess stable human judgment, lived experience or a dependable understanding of truth.

The progress is nevertheless substantial. AI is improving medical imaging, scientific research, accessibility tools, logistics, education and creative production. Model performance has advanced rapidly, smaller systems have become more capable, and the cost of obtaining a given level of performance has fallen sharply. Yet benchmark scores are not the same as general intelligence. Systems that perform impressively in controlled evaluations can still invent facts, misunderstand ambiguous instructions, lose track of long contexts or fail unpredictably when connected to real organisations and infrastructure. Artificial general intelligence remains an undefined and unachieved objective, not an established product with a reliable delivery date.

Economic promise versus realised value

The economic story requires similar caution. Frequently repeated claims that AI is “already adding trillions” confuse forecasts with realised value. McKinsey estimated that generative AI could potentially add the equivalent of $2.6 trillion to $4.4 trillion annually across selected use cases. That was a conditional estimate based on successful deployment, organisational change and productivity gains—not a measurement of wealth already created. Surveys show widespread experimentation, but the financial benefits reported by individual business functions have often remained modest. Adoption is real; economy-wide transformation is not yet proven.

At the same time, AI has become deeply entangled with the wider economy. Technology companies are spending hundreds of billions on processors, data centres, networking, cooling and power generation. The International Energy Agency reported that capital expenditure by five large technology companies exceeded $400 billion in 2025 and was expected to increase by a further 75% in 2026. Financing is spreading beyond the technology sector into banks, private-credit funds, infrastructure investors, insurers and pension capital.

Infrastructure spending and circular demand

Some investment relationships are partly circular. Chip manufacturers invest in model companies that purchase their processors. Cloud providers finance AI laboratories that commit to buying cloud capacity. Data-centre developers secure funding against long-term contracts with the same companies driving demand. These arrangements can accelerate construction and may be commercially rational, but they can also make demand appear more independent than it is. If AI revenue fails to catch up with infrastructure spending, the consequences could extend beyond speculative startups into energy, property, credit and financial markets. This does not prove that AI is a bubble, but it does mean that the industry cannot be evaluated by model capability alone.

Token prices and the true cost of applications

Token prices illustrate another contradiction. The price of running models with an older level of capability has fallen dramatically, making useful AI available to more people and smaller organisations. Stanford’s AI Index found that the inference cost of obtaining approximately GPT-3.5-level benchmark performance fell more than 280-fold between late 2022 and late 2024. However, cheaper individual tokens do not automatically produce cheaper applications.

Frontier models still charge considerably more for generated output than for ordinary input. Agentic systems may make numerous model calls, resend growing conversation histories, call external tools, retry failed operations and use additional models to inspect the first model’s work. A task that looks inexpensive in a demonstration can become costly when performed thousands of times with security, monitoring, logging, evaluation, compliance and human review. Caching, smaller models and strict computational budgets can reduce expenditure, but the advertised token rate is only one component of the total cost of ownership.

AI-generated software and the production gap

Similar realism is needed around AI-generated software. A person can now describe an application and receive a plausible interface or working prototype within minutes. This is a genuine shift in access to programming. It is not the same as independently producing a secure, maintainable and commercially reliable product.

Real applications require requirements analysis, architecture, authentication, databases, error handling, testing, deployment, privacy controls, security updates and long-term maintenance. A 2026 evaluation of prompt-to-application platforms found visually convincing results but a steep production-readiness gap. In its limited sample, no evaluated platform exceeded 60% on engineering quality or 65% on security. The study requires wider replication, but its observed failure modes are familiar: attractive front ends can conceal missing back-end functions, weak infrastructure and insecure implementation.

Evidence on coding productivity is mixed rather than universally positive or negative. A 2025 randomised study found that experienced open-source developers working in repositories they knew well took 19% longer when using early-2025 AI tools, despite believing that AI had made them faster. A 2026 follow-up using newer tools found signs of improvement, but its estimates remained uncertain. Other research reports productivity gains in different tasks and development environments. The defensible conclusion is that AI changes software work more reliably than it eliminates software expertise. It shifts effort from manually writing every line towards specifying, reviewing, testing, integrating and correcting machine-generated output.

Creating a prototype with AI alone is increasingly possible. Creating a dependable business, healthcare, educational or financial application without experienced human oversight remains far more difficult. Software is not merely code generation. It is also the management of changing requirements, conflicting interests, security risks, unusual users, unreliable external services and consequences that emerge months after launch.

Where AI projects succeed—and fail

AI projects tend to succeed where tasks are bounded, repetitive, measurable and supported by suitable data: document extraction, search, fraud triage, routine customer support, code assistance and narrow workflow automation. They struggle when goals are ambiguous, consequences are serious, processes are long, data is inconsistent or success cannot be clearly measured. Gartner has predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. This does not establish that AI is useless. It shows that autonomy is frequently marketed before the surrounding engineering, economics and governance are ready.

Power, transparency and environmental cost

Social questions are equally important. AI may increase productivity while redistributing bargaining power away from workers. It can broaden education and accessibility while also enabling surveillance, automated discrimination, persuasive misinformation and industrial-scale synthetic media. Training data raises unresolved questions about consent, copyright and compensation. Dependence on a small number of model, processor and cloud providers creates strategic concentration, while limited transparency makes independent evaluation difficult. Stanford researchers found that transparency among major foundation-model developers declined during 2025 even as deployment expanded.

Environmental costs are also material. The IEA projects that global data-centre electricity consumption could roughly double from 485 terawatt-hours in 2025 to around 950 terawatt-hours in 2030, accounting for approximately 3% of global electricity demand. The effects will be concentrated around particular grids and communities, where data centres may compete with homes and industries for electricity, water and infrastructure. AI can also help optimise grids, discover materials and improve energy efficiency, but these possible benefits do not erase the immediate physical cost of operating the technology.

Geopolitical competition is not a simple race

The geopolitical picture is more complicated than a simple race in which the United States and China occupy fixed positions. The United States leads in private investment and the development of many notable frontier models, while China leads in parts of AI research and has rapidly narrowed the model-performance gap. Open-source communities, European regulation, sovereign-computing programmes and specialised laboratories also shape the field. Leadership may ultimately depend less on announcing the largest model than on securing energy, semiconductors, talent, data, public legitimacy and economically sustainable applications.

A credible future for artificial intelligence

Artificial intelligence is therefore neither a magic solution nor an empty fraud. It is a powerful but probabilistic infrastructure whose value depends on how it is embedded in human systems. Responsible progress requires honest cost accounting, independent evaluation, clear limits on autonomy, human accountability and a willingness to reject applications that create more risk than value.

The most credible future is not one in which AI replaces human intelligence wholesale. It is one in which carefully designed systems expand what people can do without disguising uncertainty, concentrating power without scrutiny or confusing technological possibility with social progress.

Watching AI critically

The documentaries in the companion player approach AI from different editorial positions. Some emphasise economic and scientific opportunity; others foreground systemic risk, labour disruption, surveillance or loss of institutional control. None should be treated as a substitute for current technical evidence. Dates matter particularly in AI: model capability, commercial incentives and regulation can change faster than a documentary can be produced.

When assessing a film or interview, distinguish demonstrations from independently reproduced results, forecasts from measured outcomes, and a speaker’s institutional or commercial interests from the evidence they cite. The strongest understanding comes from comparing competing accounts rather than selecting the most reassuring or alarming one.

Artificial Intelligence