We tend to think about infrastructure as something physical. Roads allow people and goods to move, electricity networks distribute energy, telecommunications systems connect households and businesses, while schools, libraries and universities provide part of the institutional infrastructure through which knowledge is created and shared.
Yet modern economies also depend on less visible forms of infrastructure. Statistical literacy allows citizens to interpret claims about unemployment, inflation, health and risk. Financial literacy helps people participate in increasingly complex financial systems. Digital literacy became more important as access to employment, banking, education and government services moved online.
Artificial intelligence is creating a similar requirement.
As AI becomes embedded in search, education, workplaces, public administration, professional services and consumer products, knowing how to use these systems becomes valuable. More importantly, knowing how to question them becomes necessary. People increasingly need to understand what an AI system can reasonably establish, where its information comes from, why a fluent answer may still be wrong, what happens to information they provide, and when a machine-generated recommendation warrants further scrutiny.
These capabilities are often discussed as individual skills. Increasingly, they should also be understood as part of the infrastructure required for people and institutions to function effectively in an AI-mediated economy.
Literacy is more than knowing how to prompt
Much of the early discussion about using generative AI concentrated on prompting. Better instructions produced better answers, so learning to communicate effectively with a model became an obvious practical skill.
That remains useful, but it is a narrow conception of what people now need to understand.
A person can be highly proficient at prompting while having little understanding of whether the resulting answer deserves confidence. They may know how to generate a sophisticated report without recognising an invented citation, ask for a financial analysis without checking the assumptions behind the calculation, or provide commercially sensitive information without understanding how the service handles that data.
The deeper capability involves understanding the relationship between the user, the model, the evidence and the surrounding system.
That includes knowing that a language model is not simply retrieving a stored answer from a database, that generated confidence is not the same as evidentiary confidence, and that a model connected to search or organisational data operates differently from one answering solely from its existing parameters and supplied context. It includes understanding that AI can be useful for interpretation, comparison and generation while remaining unreliable for particular factual claims, and that an agent able to take actions creates different risks from a chatbot that can only suggest them.
This is closer to critical reasoning than software training.
The objective is not to teach everyone how to build a neural network. It is to give people enough understanding of the system to make sensible decisions about when, where and how it should be used.
The analogy with statistical literacy is useful
Most people do not need to be statisticians, but modern life routinely requires statistical judgement.
A news story reports that the risk of an outcome has doubled. A reader benefits from knowing whether the underlying probability increased from 1 per cent to 2 per cent or from 20 per cent to 40 per cent. An opinion poll reports a difference between two groups, and understanding sample size and uncertainty helps determine how much confidence should be placed in it. A medical test produces a positive result, but its meaning depends partly on the prevalence of the condition being tested.
We do not expect every citizen to calculate confidence intervals or derive Bayes’ theorem before participating in society. We do, however, benefit when people possess enough statistical understanding to recognise that numbers require context.
AI requires a similar form of practical competence.
People do not need to understand the mathematics of transformer architectures before using a language model, but they should understand that plausible language is not evidence of truth. They do not need to know how embedding models are trained, but it can be useful to understand why semantic retrieval can return information that is conceptually related rather than based on exact keyword matches. They do not need to become cybersecurity specialists, but they should recognise that information supplied to an AI service has entered a technical and contractual system whose data practices matter.
The objective is functional understanding rather than technical mastery.
AI changes the information environment
The need becomes more significant because generative AI does not simply provide another source of information. It changes the economics of producing information itself.
Text, images, audio, software and increasingly video can be generated at very low marginal cost. Organisations can produce more communications, marketers can create more advertising variations, individuals can publish more material and automated systems can generate content at volumes that would have been prohibitively expensive when every item required substantial human labour.
The supply of plausible information can therefore increase much faster than the supply of human attention available to evaluate it.
This creates a different information problem.
For much of the internet era, access was the constraint. Search engines became valuable because they helped people find information within an expanding digital corpus. In an environment of abundant generated material, finding information remains important, but establishing provenance, reliability and relevance becomes increasingly valuable.
The scarce capability moves towards evaluation.
Who produced this? On what evidence? Is the source independent? Has the information been generated, retrieved or inferred? Does the cited document actually support the claim? Is this a real image, a synthetic one, or a mixture of both? Has the information been selected because it is accurate, because it is engaging, or because an optimisation system predicts that I am likely to respond to it?
These questions are not solely questions about AI systems. They concern the information environment those systems help create.
The costs of weak understanding will not be distributed evenly
New technologies rarely affect everyone in the same way, and differences in the ability to use AI effectively may become another source of economic inequality.
A professional who understands how to use AI for research, analysis, drafting and criticism may become considerably more productive. A small business owner who can automate administrative processes may reduce costs previously affordable only to larger firms. A student who knows how to use AI as a tutor, critic and research assistant may gain access to forms of individualised support that were previously expensive.
But access to the same model does not imply access to the same benefit.
One user may treat the first generated answer as authoritative. Another may compare alternatives, request evidence, provide better context, identify uncertainty and use specialist tools when required. Both technically have access to AI, yet their effective capability is very different.
This is an important distinction because digital inequality has never been solely about whether a household possesses an internet connection. Differences in devices, skills, confidence, support and the quality of access affect what people can actually do with that connection.
AI is likely to produce a similar distinction between nominal access and effective use.
If the economic returns to effective AI use become substantial, differences in capability may compound existing inequalities in education, employment and business productivity.
Workplaces will become important sites of learning
A large share of practical AI capability will be developed at work.
This is already occurring informally as employees experiment with AI assistants, coding tools, transcription systems and other applications. Some learn quickly through repeated use, while others have limited opportunities to experiment or work in organisations where access is restricted without an alternative pathway for developing competence.
The result can be uneven capability within the same workforce.
Organisations have an incentive to address this because the value of an AI system depends partly on whether employees understand how to use it appropriately. Purchasing licences without developing user capability can produce an unusual combination of low adoption among some employees and inappropriate overreliance among others.
Training therefore needs to move beyond demonstrations of product features.
Employees need to know what kinds of tasks are suitable for AI, what information can be shared with approved systems, how outputs should be checked, when external evidence is required and when a decision should remain with a person. As organisations introduce agents and automated workflows, employees will also need some understanding of what those systems can access and what actions they are authorised to take.
This is not simply technology training. It is becoming part of ordinary professional competence.
Education faces a more fundamental question
Schools and universities face an even more difficult challenge because AI affects not only how students learn, but what it is useful for them to learn.
The first institutional response to generative AI understandably concentrated on assessment integrity. If a model can produce an essay, solve a problem or write code, educators need to know whether submitted work demonstrates the student’s own capability.
That problem is real, but it is not the whole educational problem.
Students entering the workforce over the coming decade are likely to operate in environments where AI systems are routine. Preventing students from using those systems during their education without also teaching them how to use them critically risks producing graduates who have demonstrated competence under increasingly artificial conditions.
The opposite approach is equally problematic. Allowing AI to perform intellectual work before students have developed enough knowledge to evaluate its output can weaken the very expertise required to use the technology well.
Education therefore has to manage two objectives simultaneously: developing underlying human capability and developing the capacity to work effectively with increasingly capable machines.
Those objectives are not inherently contradictory, but they require more thoughtful assessment design than either unrestricted use or blanket prohibition.
A student may need to demonstrate that they can solve a problem independently, then demonstrate that they can use AI to extend the analysis, interrogate the result and identify weaknesses. In another context, assessment might concentrate on the student’s ability to verify sources, defend methodological choices or explain why an AI-generated recommendation should be rejected.
The question gradually moves from whether AI was used towards what intellectual work the student remains responsible for.
Public servants and public institutions need the same capability
The case for broad capability becomes stronger when AI enters government.
Public institutions increasingly have opportunities to use AI for administration, document processing, service delivery, policy analysis and communication. These applications can improve efficiency and make large bodies of information easier to work with, but the consequences of inappropriate use can also extend beyond the organisation itself.
A public servant using AI to summarise a lengthy consultation document needs to understand what may be lost through summarisation. An analyst using AI to investigate policy options needs to distinguish retrieved evidence from generated interpretation. A service team using automated classification needs to know whether particular groups experience systematically different error rates.
Managers commissioning these systems need enough understanding to ask sensible questions of technical suppliers, while senior decision-makers need enough understanding to interpret assurances about accuracy, safety and risk.
This illustrates why AI capability cannot be confined to technical specialists.
A society in which a small group understands the technology while everyone else simply receives its outputs creates a substantial asymmetry of information and authority. The people affected by AI-mediated decisions need some capacity to understand what has occurred, and the people responsible for institutions need enough capability to govern the systems being used on their behalf.
Consumers increasingly interact with AI even when they did not choose to
Another reason to treat AI understanding as broadly enabling capability is that people will not always make an explicit decision to use artificial intelligence.
AI can operate behind recommendation systems, customer service interfaces, fraud detection, insurance processes, hiring platforms, search tools, pricing systems and other digital services. A person may encounter the consequences of an AI system without ever opening an AI application themselves.
This changes the meaning of literacy.
Knowing how to write a prompt helps when you are operating the system. It does little when the system is operating somewhere else and making a recommendation about you.
People also need some understanding of their position when interacting with AI-mediated institutions. Was this decision automated? Can it be reviewed? What information was used? Is there a way to correct inaccurate data? Does a human have meaningful authority to reconsider the outcome?
These questions connect individual capability with regulation and institutional design. Literacy cannot substitute for legal rights or organisational accountability, but rights are more useful when people understand when they may need to exercise them.
Organisations need literacy at several levels
The phrase “AI training” can obscure the fact that different people require different forms of competence.
A general employee using an approved AI assistant may need to understand data handling, appropriate task selection, prompting, verification and organisational policy. A manager responsible for a team using AI needs additional capability around workflow design, performance measurement and the allocation of human oversight.
A procurement specialist needs to understand enough about AI systems to question suppliers about data practices, evaluation, security and model updates. A risk professional needs to understand where probabilistic systems differ from conventional software. Senior executives and boards need sufficient knowledge to make decisions about investment, risk appetite and accountability without depending entirely on either vendors or technical specialists to frame the issue for them.
Technical teams require deeper capability again.
The objective should not be to make everyone equally expert. It should be to make the distribution of knowledge appropriate to the distribution of responsibility.
This is how organisations already approach financial capability, cybersecurity and workplace safety. Different roles require different depths of knowledge, but a baseline level of understanding is widely distributed because the underlying issue touches many parts of the organisation.
AI is moving in the same direction.
Public infrastructure does not have to mean a government programme
Describing AI literacy as public infrastructure does not imply that governments should become the sole providers of AI education.
Infrastructure can be supplied through many institutions.
Schools can establish foundational understanding. Universities and vocational institutions can integrate AI into professional education. Employers can develop task-specific capability. Libraries can provide community access and support. Professional associations can establish expectations appropriate to particular occupations. Governments can provide trusted resources and incorporate capability into public services and workforce programmes.
Technology companies also have responsibilities because interface design can either support or undermine informed use. Systems can make sources visible, distinguish retrieved information from generated content, communicate uncertainty more effectively and provide users with meaningful controls over data and automated actions.
The objective is an ecosystem in which people can acquire capability through the institutions they already encounter.
This matters because AI systems will continue to change. Training everyone on the interface of a particular product produces knowledge with a short half-life. Teaching people how to think about evidence, uncertainty, data, automation and machine-generated recommendations produces capability that transfers across systems.
The infrastructure needs to support adaptation rather than familiarity with one generation of tools.
There is an economic case for treating it as infrastructure
Markets will supply a great deal of AI training because individuals and organisations have obvious incentives to acquire valuable skills. That does not necessarily mean the socially desirable level of capability will arise automatically.
Some of the benefits of a more informed population accrue beyond the individual receiving the training. Employees who recognise security risks protect organisational information. Consumers who can identify unreliable generated material reduce opportunities for deception. Managers who understand AI procurement create stronger incentives for suppliers to provide evidence about their systems. Public servants who can interrogate automated recommendations improve institutional decision-making.
These are spillovers.
There is also a coordination problem. Employers may hesitate to invest substantially in transferable skills when employees can leave, individuals may struggle to identify which capabilities will remain useful, and educational institutions can take time to adapt curricula to rapidly changing technology.
These are familiar reasons why societies invest collectively in education and other forms of enabling capability.
The argument is not that every citizen requires an intensive course in artificial intelligence. It is that a baseline capacity to interact critically with AI increasingly produces benefits that extend beyond the person who possesses it.
The benchmark should be agency, not technical knowledge
There is a danger that efforts to improve AI understanding become another exercise in teaching terminology.
People can learn definitions of large language models, neural networks, tokens and embeddings without becoming materially better equipped to make decisions involving AI. Technical vocabulary is useful when it clarifies how a system behaves, but vocabulary should not become the objective.
A more meaningful test is whether people retain agency.
Can they recognise when an AI answer needs verification? Can they distinguish a generated explanation from evidence? Do they understand what they are sharing with a system? Can they decide when automation is appropriate and when human judgement matters? Can they challenge an AI-mediated decision rather than assuming that computational output is inherently objective? Can they use AI to extend their own capability without becoming unable to work when the system is wrong?
Those are practical outcomes rather than technical facts.
They also suggest that AI literacy should not be framed principally as learning to keep up with machines. The objective is to equip people to participate effectively in institutions and markets in which machines play a larger role.
The infrastructure we cannot see
The most important infrastructure is often noticed only when it is absent.
We rarely think about basic numeracy when reading a mortgage rate, digital literacy when submitting an online form or information literacy when deciding whether a source deserves confidence. These capabilities sit quietly behind participation in modern economic and civic life.
AI is beginning to require another layer.
As models become more capable, the need for human understanding does not disappear. In some respects it becomes more important because the outputs become more persuasive, the systems become more embedded and the boundary between recommendation and action becomes less visible.
A society that has widespread access to AI but weak capacity to interrogate it may become more productive in some activities while also becoming more susceptible to poor decisions, information asymmetries and excessive dependence on systems that relatively few people understand.
The alternative is not universal technical expertise. It is broad functional competence, supported by deeper expertise where responsibilities demand it.
If AI becomes part of the infrastructure through which people work, learn, obtain services and make decisions, then the ability to use and question it becomes part of that infrastructure too. The investment required is not simply in models, computing capacity and digital systems, but in the human capability needed to use them with judgement.
That may prove to be one of the less visible investments of the AI transition, but also one of the most important.