Issue 8 · October 2026
A plainer question first
AI has somehow become a subject on which moderation is difficult. Depending on whom you listen to, it will either transform human civilisation beyond recognition or bring about its destruction. Every new model release brings another round of predictions about which jobs will vanish and whether people will lose control of what they have built.
The enthusiasm is understandable, and so is some of the apprehension. These systems can now read and write, interpret images, process enormous quantities of information and handle tasks that once needed considerable human expertise. I still find the certainty with which people describe our collective future puzzling.
If AI is as capable as we are told, what should we use it for?
The Question Behind the Argument
People have always turned useful discoveries to purposes their inventors never intended.
Another powerful human invention
Human beings have been inventing things for thousands of years. Some inventions made life easier and some transformed civilisation. Others allowed destruction on a scale earlier generations could scarcely imagine, and many did several of these at once.
One invention, two outcomes
AI belongs in that history. It will be useful, and it will be abused, misunderstood and oversold. Companies will make fortunes while others lose them chasing opportunities that never materialise. Some occupations will change substantially, governments will regulate after the technology has moved on, and criminals will find uses for it as they have for almost every technology before it.
None of that requires AI to become humanity’s salvation or its end. It requires us to see it as another powerful invention, shaped by the people who build it, the institutions that deploy it and the purposes it serves.
The limits of our imagination
Much of today’s AI investment begins with familiar business questions: how to raise productivity, cut costs, automate customer service, produce more content or operate with fewer people. They are legitimate questions, and I run a business myself. Reducing repetitive work and making information easier to reach has real value.
For something repeatedly called transformational, though, our collective ambition looks curiously small. We have systems that read enormous volumes of material, translate languages, interpret images and explain unfamiliar concepts. Meanwhile many people are held back by a lack of access to information and expertise.
Four ordinary situations
- A farmer has no practical way to consult a specialist when a crop begins to fail.
- A child struggling with mathematics never gets the individual attention needed to understand it.
- A small business owner signs an agreement without grasping its implications because legal advice is expensive.
- A patient travels several hours to see a doctor and goes home having understood only part of the advice.
For a technology that promises to make knowledge more accessible, this ought to be fertile ground.
The same capability, pointed at different things
Conceptual diagram. No data.
Where It Could Help
Places where people are short of knowledge, access or time.
Healthcare should begin before the hospital
Much of healthcare happens too late. Someone ignores a symptom because the nearest doctor is far away. A parent cannot tell whether a child’s fever needs attention. An elderly person struggles with a prescription, and a patient realises after the consultation that several important questions were never asked.
Three points in a patient’s journey
Routine examinations, laboratory tests and histories could help qualified professionals spot patterns that would otherwise go unnoticed, and properly validated systems could point to people who need earlier intervention.
Healthcare punishes confident mistakes. A plausible but incorrect explanation can delay treatment or lead someone to a dangerous decision, so clinical validation, privacy protection and clear escalation to a person are essential. The World Health Organization has examined both the opportunities and the risks, and the technology should be tested where it will actually be used, with clinicians and patients involved in its design.
Improving access to healthcare does not require an artificial doctor. Helping people make better use of the expertise that already exists would be a substantial achievement.
Expertise should not depend on geography or income
Older people living independently could use help with appointments, medication schedules, correspondence and everyday administration. Carefully designed systems might also help carers notice changes that deserve attention. Such systems should support independence without encouraging isolation, and the people they serve should keep real control over their own information and decisions. Helping an elderly person live independently and with dignity for longer seems an excellent use of ingenuity.
Healthcare is one example of a wider problem. Legal advice, tax guidance, accounting, financial planning and technical support are services many people need and cannot afford. Large organisations keep departments or advisers for these matters, while individuals and small enterprises work things out alone.
Someone starting a small business has to understand costs, pricing, cash flow, contracts, taxation and regulation, and may be excellent at making something valuable while knowing little about running an enterprise. An assistant with reliable information could explain unfamiliar concepts, test possible scenarios and prepare questions for an accountant or lawyer. An employee could get a plain explanation of an employment agreement, and a citizen could find which government programme applies to them without working through dozens of websites.
None of this removes the need for professional judgement. Complex legal matters and financial decisions still call for accountable expertise. People then know what they need and which questions to ask, and a lawyer spends less time on routine terminology and more on the client’s actual circumstances.
A classroom with room for every learner
We have organised education for centuries around groups of students moving through a curriculum at about the same pace. Teachers do extraordinary work within that limit, with classrooms of widely different abilities and circumstances.
One lesson, four students
- One understands the idea at once.
- Another needs it explained differently.
- A third missed an earlier lesson that this one depends on.
- A fourth is learning in a second language and is afraid of looking foolish by asking.
An experienced teacher can spot many of these difficulties, but continuous individual attention is rarely possible. A well-designed AI system could notice where a learner is stuck, offer a simpler example, check whether the concept has landed and move on to harder problems. The same idea might be explained through cricket to one child and music to another. A student could ask the same question repeatedly without embarrassment, and someone could get help in their strongest language while building proficiency in another. That matters in a country as linguistically varied as India.
Helping somebody learn differs from completing their work. A student who submits a polished essay without understanding the subject has gained very little. Useful educational AI should push students to think, attempt, notice mistakes and improve, with teachers involved in deciding whether it actually improves learning. It could also take over preparing materials, translating resources and administrative tasks, and give working adults a way to learn new skills at a pace that fits their lives. We talk a great deal about AI displacing workers. We should give at least as much attention to helping people adapt.
Accessibility, language and the freedom to participate
Technology has already contributed enormously to the lives of people with disabilities, though access remains uneven. Many digital systems still assume that users can read a screen, hear instructions or operate a keyboard, and those who cannot often depend on specialised software or help from others.
One document, four ways in
- Described aloud. The contents of a document or the objects around someone with impaired vision.
- Spoken, not typed. Applications operated by voice for someone with limited mobility.
- Captioned live. Conversations and meetings made accessible to people with hearing impairments.
- Simplified. Complex writing explained in plainer language or presented as audio.
Speech synthesis is opening possibilities for people who have lost the ability to speak. Restoring someone’s capacity to express themselves in a familiar voice is hard to measure in productivity statistics. These tools need care where errors could endanger someone or biometric data is involved.
Language is a related problem. Despite the number of languages spoken around the world, the digital economy is dominated by a small group of them, and people fluent in regional or indigenous languages often find the services they need unavailable in a form they can use. Translation, speech recognition and local-language interfaces could help, especially when built with local knowledge. UNESCO has flagged linguistic diversity as a significant concern, since most of the world’s languages are poorly represented in current systems.
Languages also carry history, humour and technical knowledge, and some are fading as younger generations move to more widely used ones. AI could help communities document oral histories. Communities should keep control over how their language and culture are recorded, distributed and used.
Agriculture, food and the people who produce it
Farming depends on knowledge that changes with the location. Soil, rainfall, temperature, pests, disease, water and market conditions all shape a farmer’s decisions. Large enterprises can afford specialists and sophisticated systems, and small farmers rarely can. A farmer who notices unusual damage may need advice immediately, because the wrong treatment wastes money and can make the problem worse.
A system that examines a photograph, combines it with local weather and crop information and explains the likely causes in the farmer’s language could be genuinely useful, including by telling the farmer when to seek a specialist. This is starting to take shape in India.
In August 2026, the World Bank described initiatives in Kerala and Maharashtra. Kerala’s KATHIR platform combines agricultural data, remote sensing and advisory services, and Maharashtra’s MahaVISTAAR uses conversational AI to help farmers get information from agricultural sources. The programmes are at different stages. Their approach rests on information that reflects local conditions, comes from trustworthy sources and reaches farmers on devices they already use. Making that work with weak connectivity takes more effort than producing an impressive demonstration in a conference hall.
World Bank, August 2026
Food is lost between harvest and consumption through problems in storage, transport, refrigeration, processing and distribution. Better forecasting, quality assessment and logistics could help spot spoilage, anticipate storage needs and match supply to demand.
When infrastructure begins to fail
Some of the most valuable uses of AI may be almost invisible to the public. Electricity networks, water systems, railways, factories, hospitals and telecommunications all depend on equipment that needs constant monitoring. Machines often show signs of deterioration before they fail, in vibration, temperature, pressure, power consumption and behaviour. Predictive maintenance has existed for decades, and more sophisticated models can find complex patterns across far more operating data.
Equipment signals trouble before it fails
Illustrative only. Not measured data.
An engineer could be alerted to abnormal behaviour in a pump before it interrupts water supply. A railway team could investigate a developing fault before it disrupts service. These systems have to work alongside established engineering controls, and the measure is whether they warn early enough, and accurately enough, to justify acting. A hospital generator that fails in an emergency can cost far more than the equipment itself.
Water deserves particular attention. AI and conventional analysis could help find leaks, examine consumption, schedule irrigation and run treatment plants more efficiently. No single technology resolves water scarcity, but cutting avoidable losses in systems we already run would be worthwhile. The International Energy Agency has examined similar uses in electricity networks and industry, with benefits that depend on implementation.
AI itself uses substantial electricity and water, and its hardware carries environmental cost. A system burning considerable computing for a marginal gain is hard to justify, while a smaller model that prevents repeated equipment failure could be worth far more.
Making disaster response more effective
We collect enormous quantities of weather data, satellite imagery and environmental observations, and still struggle to get useful information to the people who need it. A cyclone forecast or flood warning only helps when people understand what it means for them and what to do. Their immediate questions are practical: will my neighbourhood flood, should I leave, which roads are open, where is help?
AI could help turn verified warnings into local languages and formats suited to different communities, and help authorities organise information from several sources into guidance people can follow. Accuracy matters more in an emergency than anywhere else.
Afterwards, emergency services need to know how bad the damage is, which communities are affected and where to send limited people and supplies. The World Food Programme’s SKAI project, developed with Google Research, uses machine learning to identify building damage from satellite imagery after disasters and has supported assessments in affected regions. Images still have to be combined with field reports and local knowledge. A damaged building may be empty, while a community that looks intact may be short of food, medicine or clean water.
World Food Programme, with Google Research
Scientific discovery, and the world we are changing
Researchers spend years examining published studies, interpreting results and comparing observations. AlphaFold, developed by DeepMind, predicted protein structures with remarkable accuracy, and with the European Bioinformatics Institute it made predictions for more than 200 million proteins public in 2022. Protein structures help scientists investigate disease, drug development and other biology. The predictions are not experimental proof of every function, but access to them changes how researchers approach their work. Similar possibilities exist in materials science, chemistry and medicine.
Organising scientific knowledge could help as well. Relevant findings may sit in papers from neighbouring disciplines or in data nobody can easily search. Failed experiments hold information too, since knowing why an approach did not work prevents repetition and suggests better directions.
Better questions and more possibilities examined, with evidence still deciding the answer.
Environmental protection is a related field. Conservation teams collect satellite images, acoustic recordings and photographs. AI could help identify species in camera-trap images, recognise calls and highlight changes in vegetation, forests, coastlines and land use. Data quality matters. An animal missing from a photograph has not necessarily left the area, and a confident classification does not prove a species was present, so scientists and local communities remain essential to interpreting the results.
Communities also hold generations of knowledge about seasons, plants and local conditions that formal databases poorly represent. Technology could help document some of it, provided the communities decide what is recorded and how it is used.
Making public institutions easier to use
Public services are often organised around administrative requirements instead of the experience of the people using them. Citizens must understand eligibility criteria, find forms, gather documents and move between departments, which is hardest for elderly people, people with disabilities and those unfamiliar with the institution’s terminology. AI could explain procedures, identify the relevant service, list the documents required and help prepare applications. It could help officials locate information, organise case records and answer routine enquiries. Poorly designed processes should be simplified, not hidden behind a conversational interface.
| AI may help | A person decides |
|---|---|
| Explaining procedures and services | Eligibility for assistance |
| Preparing applications | Rejection of an application |
| Organising evidence and case records | Any decision that affects someone’s rights |
| Transcribing proceedings, translating documents | Guilt and liability |
| Searching legal material, finding relevant cases | Final accountability for the outcome |
People must be able to understand decisions, challenge errors and obtain human review. The justice system raises the same point. Organising evidence and searching law consume professional time, particularly where backlogs are large. Judicial decisions involve legal reasoning, procedural safeguards and responsibilities that cannot be handed to a model because it writes convincing legal prose.
Choosing Well
What happens to the people whose work changes, why so few of these problems attract investment, and where an organisation might begin.
Freeing people to do better work
A remarkable amount of human effort goes to work that adds little to understanding, creativity or professional satisfaction. Copying information between systems, searching folders, formatting documents, preparing routine reports, transcribing meetings and reconciling repetitive records fill large parts of many working days. AI could help where the information is unstructured, provided it performs reliably and its output can be checked.
Where the time could go
- A nurse spends less time on administrative records and more with patients.
- A teacher gives more attention to students.
- An engineer investigates a hard problem, not years of maintenance reports.
- A security analyst examines an unusual event rather than assembling information from disconnected systems.
If a process takes half as long, output expectations tend to double. Sometimes that is justified. Other times the benefit should be better service, fewer errors or better working conditions.
An organisation that uses AI to produce twice as many documents has not become twice as effective. It may have created twice as much for somebody else to read.
The employment effects are real. Some roles will change substantially and some will disappear, with real consequences for individuals and families. AI can assist with training, but helping people adapt takes more than a learning application. Employers, educational institutions and governments all share responsibility for making sure change does not leave large groups without viable opportunities.
Why are we building so little of this?
What strikes me about these possibilities is how few are futuristic. Many are feasible today at least in part, and some are already deployed. Others need better data, specialised models or more careful integration with existing systems. We have plenty of useful problems. The incentives are uneven. A product for a well-funded enterprise with reliable infrastructure and a purchasing budget is easier to build than a service for a rural health centre, a small farmer or a public institution short of money, and the people who would benefit most may be the least able to pay.
None of this began with AI. Human beings have always mixed invention with ambition, curiosity, generosity and self-interest. Commercial success and social value also need not compete, and a company that helps farmers cut losses or makes industrial infrastructure more reliable can create considerable economic value. Private investment, public research, universities, open technology and community-led development each suit different problems.
Open-weight models and smaller specialised systems may suit places where affordability, language support, privacy or poor connectivity make dependence on large remote services impractical. Not every useful application needs the most sophisticated frontier model. Sometimes a small adapted model is right, sometimes conventional automation is enough.
Over the life of the system
- Who will maintain it?
- Who is responsible when its advice is wrong?
- What happens when the service is unavailable?
- Can the intended users afford it?
- Does it still work under the conditions they actually live in?
Choosing problems worth solving
I suspect we spend too much time debating what AI may eventually become and too little on what we could accomplish with what already exists. Long-term research and genuine risks both deserve attention. But the practical possibilities are substantial, and many do not require machines far more intelligent than today’s.
| Application | What it needs more than brilliance |
|---|---|
| Farm adviser | Excellent local information, in the language the farmer speaks |
| Learning assistant | Patience, reliable subject knowledge and a sound grasp of how people learn |
| Medical support | Accurate records, careful validation and integration with clinical services |
| Industrial monitor | Dependable sensors and well-tested analytical models |
| Public service | Administrative reform as much as software |
These problems are hard because they involve the world as it is. People have different needs, information is incomplete, institutions have established processes, resources are limited and mistakes have consequences. Solving them takes people who understand the technology working with people who understand where it will be used. That is rarely as exciting to watch as an autonomous agent completing tasks on a screen, and harder to demonstrate during a product launch.
Begin where the need is
I would like to see more organisations begin their AI strategy by looking at where a lack of knowledge, access, resources or attention is preventing a worthwhile outcome. Once those problems are understood, we can decide whether AI offers a useful answer. That is a more productive starting point than deciding that every organisation needs AI and then searching for something to make it do.
Four places to look
- Where are people struggling because the information they need is hard to obtain?
- Where are qualified professionals spending time on work that stops them using their expertise?
- Where do failures occur because warning signs are buried in large volumes of data?
- Where could better communication, translation or analysis materially improve someone’s circumstances?
The direction of technology follows the choices we make about what deserves investment, attention and effort.
I am not suggesting that every company drop its commercial objectives in pursuit of a vaguely defined social good. Businesses need to survive, investors need returns and individuals are entitled to their ambitions. I would simply like us to be more ambitious about the kind of problems worth solving.
The closing signal
I do not believe AI is going to solve humanity. Humanity was never a problem with a software fix. I am equally unconvinced by the confidence of those who say we are approaching extinction. We have lived with technologies capable of extraordinary harm before, and we have shown both remarkable ingenuity and considerable carelessness in using them. We have let commercial interests shape inventions in ways we later regretted, regulated too late and persisted with ideas long after their limits were clear. We have also used technology to cure diseases, raise food production, connect continents and open knowledge to people who would once have been shut out.
AI will probably follow a similarly untidy path, with environmental costs, employment disruption, speculation, fraud and applications that should never have been built. There will also be discoveries whose value takes years to appreciate, and some will come from people working on modest problems that matter enormously to the communities they serve. For decades we have built systems that make it easier to click, buy, scroll and consume. We now have technology capable of more demanding work: making expertise accessible, improving essential services, preserving knowledge, reducing waste and supporting discovery.
The closing signal
What could we build with AI if the first measure of success was that somebody’s life became meaningfully better?
Sources and references
- World Health Organization. Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models. Geneva: WHO, 2024.
- World Bank. Small AI Transforms Farming in India. 27 August 2026.
- World Food Programme Innovation Accelerator. SKAI. Developed with Google Research.
- EMBL-EBI and Google DeepMind. AlphaFold Protein Structure Database: About.
- EMBL-EBI. New AlphaFold DB release with 200+ million predicted structures. 28 July 2022.
- International Energy Agency. Energy and AI: AI for energy optimisation and innovation. April 2025.