The Evolution of Healing: Emergent Intelligence and the Human Instinct to Outgrow Our Own Destruction

 

1. The hypothesis of a species learning to correct itself

There is a way of understanding the contemporary rise of artificial intelligence that begins with a question about humanity rather than machinery. Why, at a moment when the consequences of industrial expansion have become increasingly visible, have people also developed instruments capable of enlarging their powers of observation, prediction, synthesis and invention? The conjunction is striking. Climate change, ecological degradation, geopolitical instability, informational disorder and the exhaustion produced by relentless economic competition have become defining features of the early twenty-first century. At the same time, the cost of certain kinds of intellectual work has fallen, and a person with access to a computer can now undertake forms of research, planning, translation, programming or creative experimentation that previously required larger organisations. It is tempting to see this convergence as the emergence of a corrective capacity within civilisation itself. A body threatened by injury mobilises protective systems; a civilisation increasingly conscious of the damage it has caused develops tools that might help it recognise and address that damage. This essay calls the possibility the new white blood cell: an account of artificial intelligence as a potential instrument of collective self-repair. The phrase is a metaphor, not a scientific classification or a claim that human history was destined to produce AI. Its purpose is to organise a serious proposition: intelligence created by human beings might, if governed well, increase their ability to notice errors, coordinate responses and restore conditions for flourishing before opportunities disappear.

The metaphor captures something that conventional descriptions of technological progress leave underdeveloped. Computing is routinely measured through speed, benchmark performance, investment or commercial adoption. These measurements matter, but they say relatively little about the ends to which improved computational ability is directed. A productive system can accelerate a harmful process just as readily as a beneficial one. The deeper question concerns whether new intellectual capacity can be deliberately recruited against the specific habits that place human societies and ecological systems under pressure. A useful analogy is the feedback system: information about the consequences of an action changes subsequent behaviour. Human beings possess many mechanisms of correction, including the scientific method, democratic scrutiny, investigative journalism, ethical reflection, medical diagnosis and the mundane capacity to apologise and learn. Yet these mechanisms are imperfect, slow, unevenly funded and vulnerable to distortion. AI can potentially improve their reach by finding patterns in immense bodies of information, assisting people with unfamiliar tasks and shortening the interval between the recognition of a problem and the testing of a response. Stanford's 2026 AI Index documents rapidly expanding technical capabilities and adoption, while also warning that responsible evaluation has not kept pace with capability growth [1]. These paired observations describe the entire predicament: the machinery of correction is becoming more powerful, while the machinery for correcting the machinery remains unfinished.

The claim can be expressed at three scales. At the individual scale, an assistant can help a person convert vague intentions into a structured plan, recover a forgotten concept, prepare a difficult application, understand an unfamiliar subject or begin a long-postponed creative project. At the institutional scale, computational systems can help identify administrative errors, improve access to services, synthesise scientific literature or allocate limited resources more carefully. At the planetary scale, AI can contribute to weather prediction, environmental monitoring, materials research, energy management and the analysis of ecological change. These are related forms of assistance, but success at one scale does not establish success at the others. Making an essay easier to write is not equivalent to lowering greenhouse-gas emissions; producing better forecasts is not the same as possessing the political authority to act on them. A central argument of this essay is therefore that acceleration becomes repair only when the increased speed of thought is joined to valid evidence, moral judgment, public accountability and material action. The United Nations Development Programme's 2025 Human Development Report makes a complementary case: the meaningful issue is the choices people can exercise and the lives they have reason to value, rather than the spectacle of machine capabilities considered in isolation [2].

This reading of AI begins in hope without requiring a declaration of faith. It is possible to feel that the world has moved towards stagnation, depletion and danger while also recognising that no generation has an uncontested monopoly on hardship or disillusionment. The recovery of possibility does not mean a denial of the present emergency. It means escaping the assumption that existing institutions, working habits and technical limits are the permanent shape of human life. The distinctive promise of AI lies in its capacity to loosen some constraints on intellectual participation. A person who cannot afford a research team may gain useful analytical help; an educator facing too many demands may be able to prepare more appropriate materials; scientists separated by language or discipline may find new routes into one another's work. Whether these gains become an enduring social good depends on how access, responsibility and the rewards of increased productivity are distributed. The new white blood cell is thus best understood as a hypothesis about what humanity could choose to become: a civilisation that uses a portion of its ingenuity to prevent its powers from outrunning its wisdom. The hypothesis will earn its place only if it survives a confrontation with biology, history, politics, economics and the physical costs of computation.

2. What the immune-system analogy reveals, and what it cannot prove

White blood cells are not little moral agents tasked with defending an innocent organism. They are diverse biological cells participating in an evolved immune system that distinguishes among signals, responds to pathogens and damage, coordinates inflammatory processes and, in many cases, retains forms of immunological memory. Their responses are context-dependent. A useful immune reaction must identify a threat accurately enough to respond without causing disproportionate harm. Immune systems can fail through insufficient protection, excessive inflammation, misrecognition of the body's own tissues or dangerous reactions to substances that are not genuine threats. The significance of this biological comparison is therefore subtler than the comforting image of a benevolent defender. An advanced society also needs detection, memory, judgment, response and restraint. Its defensive capacities include safeguards against the misuse of safeguards themselves. The immune metaphor becomes intellectually productive when it encourages a question about proportionality: how can a corrective system respond effectively without damaging the life it is intended to preserve?

Artificial intelligence can be compared to a set of added sensory and analytical capacities, rather than to an independent biological organism. Pattern-recognition systems detect anomalies in medical images or industrial equipment; language models retrieve, reorganise and explain possible interpretations of information; optimisation systems search through large spaces of options. Each function resembles some aspect of sensing or coordinating a response. But AI has no intrinsic biological allegiance to human well-being. Most current systems acquire their abilities through training regimes, infrastructure and human-defined objectives, and the same system may be used in mutually incompatible ways. If asked to assist public-health research, a model might serve a protective function. If deployed to maximise engagement through inflammatory content, it could intensify the very dysfunction that societies need to overcome. A white blood cell operates within a particular organism's physiology; AI services operate within markets, states, social institutions and competing jurisdictions. Their purposes emerge from human choices and incentives, not from a proven instinct of planetary preservation.

The metaphor also directs attention to collective memory. An immune system that learns from an encounter is better prepared for certain future threats; societies that preserve evidence of past failures can avoid repeating them. Industrial accidents, financial crises, epidemics and environmental disasters generate reports, archives, hearings and scientific studies, yet lessons often remain scattered across incompatible databases, professional languages and organisational boundaries. AI-assisted search and synthesis could help make institutional memory practically available. A civil servant preparing a regulation might be able to examine previous implementation failures; a regional hospital might compare treatment protocols with current evidence; an engineer might retrieve warnings buried in decades of maintenance records. These applications are valuable only when their outputs remain traceable to primary sources and when disagreements are preserved instead of disguised as consensus. The objective is an institution that remembers its errors more effectively, not a machine whose fluency is mistaken for memory or judgment. Standards such as the US National Institute of Standards and Technology's AI Risk Management Framework, including its generative-AI profile, treat validity, safety, security, accountability and ongoing evaluation as system properties that must be engineered and managed [3].

There is a further and less comfortable implication. Systems designed to detect threats can manufacture enemies when their thresholds, categories or incentives are wrong. A government that uses automated tools to identify fraud may wrongly penalise vulnerable people. An employer that uses predictive analytics to classify supposedly low-performing workers may reproduce historical discrimination. A platform that claims to defend the public from disinformation may suppress legitimate disagreement through opaque rules. These are technological analogues of autoimmunity: a protective intervention harms the wider body through defective recognition or excessive force. Such analogues should not be stretched into biological identity, but they impose a useful moral discipline. No serious theory of civilisation's self-repair can grant a new technical system unlimited authority because it describes itself as protective. The corrective mechanism requires independent correction, appeal, transparency and the possibility of refusal. A genuine public immune capacity is distributed among scientists, citizens, institutions and accountable technologies. It remains answerable to those who bear the consequences of its decisions.

3. Evolution without destiny: how an unforeseen capability emerges

The appearance of intelligence-producing technology can feel mysterious because its capacities sometimes seem qualitatively different from the component inventions that preceded it. Semiconductor devices, digital networks, statistical methods, stored text, large-scale computing infrastructure and decades of machine-learning research were all recognisably human achievements. Yet their combination produced tools whose practical range was difficult for many observers to anticipate. This is a familiar feature of complex historical development: cumulative changes create new conditions under which unanticipated possibilities become available. It does not require a concealed evolutionary command directing humanity towards a predetermined end. Biological evolution through natural selection has no foresight about the needs of future organisms, and cultural or technological change also involves contingency, competition, imitation, investment, failure and choices that could have gone differently. The scientific distinction matters. To say that AI was 'needed' by a troubled civilisation expresses a judgment about the timing and potential use of the technology. It does not demonstrate that nature intentionally created it in order to save humanity.

The history of computing provides a less mystical but more remarkable story. Machines initially developed for arithmetic, administration and military calculation were progressively used for communication, simulation, visual expression and the organisation of knowledge. Vannevar Bush's 1945 essay 'As We May Think' argued, following the destructive application of scientific talent during wartime, for machinery that could make vast accumulated knowledge more accessible to human inquiry. His proposed memex was a speculative information device, not a modern language model, but the underlying concern remains recognisable: the growing quantity of knowledge can exceed the human capacity to retrieve and connect it [4]. During the following decades, researchers explored cybernetics, symbolic reasoning, information retrieval, neural networks and statistical learning. The contemporary capabilities of generative systems arose from this long and uneven history, accelerated by data availability and powerful specialised computing hardware. AI is therefore an outgrowth of accumulated culture in the ordinary historical sense. Its surprising properties are a reason to study the interaction of technical components and social institutions, not evidence of supernatural intervention.

Nevertheless, refusing a literal teleology need not empty the experience of meaning. Human beings often recognise the value of an invention only when later circumstances reveal what it can do. The printing press supported religious conflict as well as scholarship; industrial manufacturing made possible both mass medicine and mass warfare; digital networks widened communication while enabling surveillance and manipulation. Each development created a larger field of choices than its designers originally controlled. We can reasonably ask whether AI's arrival creates an unusually consequential opportunity to bring latent capacities for coordination and learning into practical use. The appropriate language is historical possibility rather than cosmic inevitability. Tools that allow people to work across disciplines, translate between languages and simulate complex processes can alter what societies find achievable. Yet possibilities remain conditional upon the physical world, institutional design and political decisions. A model that generates a credible research proposal has not performed the experiment; a model that identifies an inequity has not altered the laws or incentives that sustain it.

Technological evolution also differs from biological reproduction in a politically decisive respect: people can debate its objectives and attempt to redirect its pathways. Organisms cannot convene a public consultation to reconsider the rules of natural selection, whereas societies can regulate critical infrastructure, fund particular research programmes, prohibit abusive applications, establish rights of appeal and invest in broadly available tools. Such interventions will never give perfect control over innovation, but they make responsibility possible. The 'deeper evolutionary impetus' behind AI can therefore be interpreted as an emergent cultural tendency: humans repeatedly build extensions of their limited senses, memories and cooperative abilities, and these extensions become especially attractive when the problems they face exceed unaided human scale. The invention of microscopes, vaccines, modern statistics and computer networks all enlarged the range of things people could observe or coordinate. AI may extend the same trajectory into tasks long associated with intellectual labour. The question is whether we will use this extension to deepen our existing self-destructive arrangements or redesign the arrangements themselves.

4. Civilisation as a feedback system

Modern societies already possess the elements of self-correction, although the elements are distributed unevenly and sometimes work against one another. Democratic elections can change leadership; courts review abuses; scientists revise explanations in response to evidence; businesses respond to scarcity and demand; journalists expose wrongdoing; families re-evaluate practices when consequences become apparent. These are feedback loops operating across different time scales. Some are fast and local, while others are delayed by political conflict or the slow movement of natural systems. Atmospheric carbon dioxide can accumulate for generations before its full consequences become apparent; the benefits of childhood education may unfold across a lifetime; ecological damage may not be reversible on any humanly convenient schedule. When the signal arrives after the capacity to respond has been weakened, correction becomes much more difficult. AI's relevance is partly temporal: it may shorten the delay between observation, interpretation, coordination and intervention, provided that data are sound and the recommended intervention is actually carried out.

This temporal argument gives the phrase 'acceleration of repair' a specific meaning. In a mature corrective system, speed is valuable at stages where delay causes avoidable harm. A warning about a possible equipment failure may prevent an accident. A more accurate storm forecast may allow evacuation sooner. A literature review prepared rapidly can help researchers rule out an unpromising direction before exhausting limited funds. An automated translation system can make public-health information accessible within hours instead of weeks. In each example the accelerated component belongs to a larger chain of action. If the evacuation system is inaccessible to disadvantaged residents, if the engineering inspection is ignored or if a translated document contains a dangerous mistake, faster information production may produce no beneficial outcome. It may even create false confidence. The European Centre for Medium-Range Weather Forecasts brought its AI forecasting system into operations beside its established physics-based forecasting system in February 2025, and subsequently operated an AI ensemble system, illustrating both the promise of faster forecasts and the value of complementary methods [5]. The achievement consists in better public forecasting infrastructure, not in an isolated assertion of machine superiority.

The language of feedback also helps explain why technical competence can coexist with apparent social irrationality. People regularly know more than their institutions can absorb. Scientists have documented the causes of global warming in considerable detail, yet societies continue to emit greenhouse gases at levels inconsistent with their stated long-term objectives. The Intergovernmental Panel on Climate Change's 2023 synthesis report attributes observed warming unequivocally to human activities and describes the need for deep and sustained reductions in emissions [6]. The central difficulty is not simply missing knowledge. It is the misalignment of incentives, unequal exposure to costs, political disagreement, technological lock-in and the unequal capacity of groups to change course. AI may improve modelling or implementation, but it cannot resolve these conflicts by generating more eloquent explanations. Corrective intelligence must have channels through which knowledge affects resource allocation, law and daily conduct. The global body cannot heal through diagnosis alone.

An effective feedback system must also acknowledge uncertainty and avoid confusing performance indicators with real outcomes. Organisations frequently measure what is easy to count: documents produced, calls completed, forms processed, cases closed or energy consumed per unit of output. These measures are useful until they become substitutes for the underlying purpose. A hospital could optimise appointment throughput while neglecting patients with complex needs; a school could optimise examination scores while narrowing education; a government could optimise response times while reducing substantive accountability. Advanced optimisation magnifies this familiar problem. When artificial systems are rewarded for satisfying proxies, they may become extraordinarily effective at producing the appearance of success. A theory of AI as self-repair therefore requires independent measures of health, dignity, ecological integrity and informed choice. It needs human institutions capable of asking whether a process should exist in its present form, rather than merely whether the process can be made more efficient.

5. From faster production to a different distribution of possibility

A major part of AI's social appeal lies in its reduction of the initial effort required to begin complex work. Many people carry ambitions that are not defeated by lack of interest or talent but by the cumulative friction of unfamiliar tools, missing professional contacts, uncertain terminology, fatigue and insufficient time. A novice programmer faces documentation and debugging; an aspiring author faces structure, research and editing; a person seeking a new profession faces opaque selection criteria and the labour of presenting experience coherently. Conversational systems can supply explanations, preliminary drafts, practice exercises and structured alternatives at the moment a person needs them. Their output may be imperfect, sometimes seriously so, yet the lowering of initial barriers can alter a person's sense of what is possible. This is a social effect in addition to any productivity measurement. The first movement from intention to experiment becomes easier, and the work that once seemed inaccessible may become imaginable as a sequence of manageable steps.

Empirical evidence supports a qualified version of this optimism. In a study of customer-support agents, Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that access to a generative-AI assistant increased measured productivity by an average of about 14 per cent, with larger gains among less experienced or lower-skilled workers in their sample [7]. The result cannot automatically be transferred to all occupations, and it did not establish that every worker became better at every task. It nevertheless illustrates a valuable possibility: knowledge embedded in the practices of experienced people can become easier for newcomers to access. This redistribution of practical guidance has implications for social mobility. If a good explanation, drafting assistant or simulation can be made reliably available at low cost, the advantages conferred by privileged access to mentors and professional networks may narrow in certain circumstances. The opposite outcome is also possible if advanced systems remain expensive, depend on dominant platforms or are deployed chiefly to increase pressure on those who already possess the least bargaining power.

The distinction between compressing a task and freeing a life is central. Saving two hours in producing a report has social value only under conditions that determine what happens to those hours. They might become additional time for careful judgment, rest, family, citizenship, learning or creative practice. They might also become an expectation that the same worker produce three more reports. A reduction in the time required for intellectual routine is not automatically a reduction in human exhaustion. Economists, policymakers and employers must attend to where the benefits accrue. The International Labour Organization's 2025 assessment found that roughly one in four jobs worldwide has some degree of exposure to generative AI, while judging transformation of jobs more likely than wholesale replacement in many cases; it stressed differences by occupation, income and gender [8]. These are exposures rather than predictions of realised job loss. The moral stakes of acceleration are determined by wages, bargaining arrangements, training opportunities, public services and access to the technology.

Possibility is also constrained by attention. When it becomes cheap to generate a plan, a summary, a proposal or an image, the scarce resources may shift towards verification, decision and completion. People can acquire hundreds of potential projects without gaining the time or discipline to finish one. This is an especially serious problem for an essay arguing that AI reopens a future previously felt to be closing. The appropriate claim is not that technology eliminates difficulty; it changes the location of difficulty. More people may reach the threshold of participation, while the challenge of distinguishing sound work from plausible-looking work becomes greater. Human expertise remains necessary to evaluate claims, choose among competing goods and carry projects through the slow processes of trial and revision. AI is most socially useful when it strengthens these capacities rather than producing dependence on unexamined outputs. A mature culture of accelerated work would value completion over generation, comprehension over the appearance of fluency and durable improvements in human life over the volume of artefacts produced.

6. The lost future and the recovery of wonder

The idea that AI could restore a lost age of possibility often has a recognisable cultural setting: the technological optimism of the late twentieth century. During the 1990s, personal computers, the early public internet, increasingly sophisticated electronic entertainment and a relatively open imagination of the digital future appeared to promise widening access to knowledge and creative tools. Many people encountered technology as a field of experimentation rather than as the infrastructure through which nearly all social and economic activity would eventually be mediated. Educational software, home publishing, computer games, early websites and affordable recording equipment suggested that an ordinary household could become a small centre of invention. Retrospective memories of this period should not erase its exclusions, economic anxieties or the political failures that were already present. Nor was there a single shared utopian mood. The historical point is that a generation could plausibly imagine the networked future as an expansion of human agency, before surveillance capitalism, addictive platform design, permanent connectivity and information warfare made some of those expectations seem naïve.

The subsequent disillusionment has several causes that should not be attributed solely to technology. Financial instability, rising housing costs, environmental anxiety, widening social inequality, the experience of war, the pandemic and distrust of institutions have each changed how people imagine the future. Digital platforms contributed their own forms of exhaustion: feeds that produce continuous novelty without durable satisfaction, enormous quantities of information without an accompanying growth in wisdom and economic systems that convert attention into a commodity. Technology became ordinary infrastructure and, in some respects, a source of compulsion. A certain kind of wonder disappeared because the promise of endless possibility was replaced by the experience of endless obligation. AI emerges into this history ambiguously. It may repeat the same arc, beginning as a liberating novelty before becoming another mechanism of extraction and control. Yet it also enables a different relationship to computing, in which people can formulate goals in familiar language and receive help moving towards them. The experience of initiative shifts, at least partially, back towards the user.

An account of recovered wonder must avoid equating wonder with entertainment or effortless gratification. The deeper feeling is that the world contains possibilities one has not exhausted, including the possibility of understanding something difficult and contributing something useful. A conversational interface can support that feeling when it encourages inquiry, reveals connections, proposes experiments and makes advanced material accessible without condescension. It can suppress it when its confidence discourages skepticism, or when people allow it to replace rather than develop their own capacity. Educational philosophy has long distinguished acquiring answers from learning how to think. The same distinction applies to a writer who can produce ten thousand words with an AI tool: the document may have appeared, but intellectual ownership depends on understanding the argument, judging the evidence and being able to revise its central claims. The goal is therefore not the frictionless production of culture. It is a more hospitable entrance into difficult and rewarding work.

This explains why the prospect of AI-assisted creativity can matter emotionally even to people who never intend to become engineers or entrepreneurs. The ability to translate an idea into an intelligible plan can counter resignation. A person facing a midlife sense of diminished energy may discover that an ambitious project can be developed in intervals rather than deferred until impossible conditions of free time arrive. The change is modest at the scale of civilisation but profound at the scale of a life. It recovers the belief that action still has a plausible relation to aspiration. The collective equivalent would be a society that makes this experience widely available through libraries, public education, accessible technology and institutions that permit experimentation without punishing every failure. The renewal of wonder then becomes a public achievement, not a marketing promise. A society filled with sophisticated tools but deprived of shared time, security and trust would remain poor in genuine possibility.

7. The planetary predicament: self-sabotage made measurable

Human self-sabotage is an evocative description of activities that undermine the conditions on which future human welfare depends. The most substantial examples are not mysteries. Fossil-fuel combustion changes the composition of the atmosphere; land conversion diminishes habitats; certain forms of agricultural and industrial production damage soils, waterways and biological diversity; the extraction and disposal of materials generate pollution. The Intergovernmental Panel on Climate Change has established both the human origins of recent global warming and the consequences of continued emissions, while emphasising that future warming depends on choices made now [6]. The word 'sabotage' gives these processes a moral charge, although they are usually produced by distributed incentives rather than a single intention to cause harm. People pursue mobility, comfort, food, employment and commercial advantage through arrangements that can make destructive outcomes rational in the short term. Corrective technology must confront this structure. The principal problem is that individually defensible actions can accumulate into collective damage, and the people who benefit most from an activity are often separated from those who bear its costs.

AI could assist by making the relationships between local action and distant consequence more legible. Remote sensing, environmental modelling and automated analysis can contribute to detecting changes in forests, agriculture, atmospheric pollutants, coastal areas and infrastructure. Models can help estimate likely consequences under alternative policy choices, and predictive systems can improve the allocation of scarce emergency resources. But computational visibility is a prerequisite for action only in some cases. The world has not failed to reduce emissions because no one can calculate that greenhouse gases warm the planet. Better measurement can remove excuses, reveal unexpected sources of harm and evaluate whether an intervention is working. It cannot substitute for legislation, public investment, international cooperation or changes in production and consumption. A satellite image that shows methane escaping from a facility is useful if an authority can require repairs and the operator can afford or be compelled to make them. Without that chain of accountability, the image adds knowledge without changing the result.

The concept of planetary maintenance deserves a place beside the more familiar language of technological breakthrough. Modern cultures tend to celebrate novelty: a new machine, a powerful model, an extraordinary discovery. Ecological sustainability, however, also requires the less glamorous work of inspection, restoration, preventive care, waste reduction and long-term stewardship. AI may be highly useful in these fields because maintenance generates complex information that has historically been costly to interpret. Water systems leak, electrical networks fail, bridges age, crops face variable local conditions and conservation programmes must operate over wide areas with limited personnel. Software that helps prioritise inspections, detect anomalies or reduce unnecessary material use may create cumulative public value without attracting spectacular headlines. The most appropriate comparison is preventive medicine. A healthier future often depends on detecting modest problems early enough that they do not develop into catastrophes. Yet these applications are vulnerable to the same governance failures that affect public services generally: insufficient investment, poorly selected performance measures and a preference for visible announcements over continuing care.

Ecological ambition also requires a sober attitude towards uncertainty. Complex natural systems do not always behave in ways that data-driven models can capture, particularly when future conditions differ from those represented in training data. Rare disasters, abrupt ecological changes and unprecedented combinations of pressures may defeat apparently accurate forecasts. Ground observation, expert fieldwork, community knowledge and established scientific models remain indispensable. AI can extend the reach of these practices when it is treated as one instrument among several, and when uncertainties are clearly presented to decision-makers. The ethical standard is not merely an accurate prediction under average circumstances. It is the ability to act responsibly when the model is wrong, when the consequences fall disproportionately on vulnerable communities and when the most important environmental values are difficult to reduce to numbers. The planet's repair requires better tools, but it also requires institutions that accept responsibility for the limits of those tools.

8. Seeing the weather, seeing the system

Weather prediction offers an unusually tangible example of artificial intelligence strengthening an existing scientific institution. Forecasting depends on immense quantities of observation and on sophisticated accounts of physical processes. Atmospheric dynamics are difficult to calculate, and small differences in initial conditions can influence later outcomes. In February 2025 the European Centre for Medium-Range Weather Forecasts made its Artificial Intelligence Forecasting System operational alongside its physics-based forecasting system. The centre reported improvements on numerous measures, including tropical cyclone tracks, and a substantial reduction in the computing energy required to produce the particular AI forecast [5]. Later that year it introduced an operational ensemble version, generating multiple forecasts to represent possible outcomes rather than treating one prediction as certainty [9]. This is a model of productive complementarity: an established public scientific organisation integrates machine learning into a broader system of observation, physical understanding, validation and communication.

The benefits of better forecasting extend beyond the satisfaction of knowing tomorrow's weather. Farmers need estimates of rain, temperature and frost; public-health authorities prepare for dangerous heat; energy networks balance anticipated supply and demand; emergency agencies consider evacuation and rescue; transport systems face disruption from storms. Greater warning time can prevent injury, property damage and avoidable economic losses, but the effect depends on who receives the warning and whether they possess the means to respond. A comfortable household can prepare for a heatwave differently from a person living in insecure housing. The fairest measure of progress therefore includes distribution as well as technical skill. An AI forecasting system that performs well in a research benchmark and is communicated effectively to wealthy cities may still leave large populations exposed. The corrective purpose is fulfilled only when forecasting capacity is connected to local knowledge, public infrastructure, effective communication and resources for action.

Weather forecasting also supplies a useful correction to the idea that AI is replacing science. Machine-learning models can detect structure in historical and simulated data without reproducing every aspect of a traditional physical calculation. This can be remarkably effective, but it does not abolish thermodynamics, atmospheric dynamics or the need for measurements. The accuracy of a learned model depends partly on the quality of the scientific and observational infrastructure from which it learns. A future in which only model outputs remain visible, while instruments and expert institutions are neglected, would undermine its own foundations. The optimistic conclusion is one of cumulative intelligence: human theories, observational networks, computational methods and local action can become mutually reinforcing. The same pattern should guide AI deployments in medicine, energy and public administration. The best test of a new system is whether it improves the wider practice of knowledge, including the capacity to discover and correct its errors.

The lesson reaches beyond meteorology to the relation between information and trust. Forecast users need to know which claims are reliable enough for a particular decision, which are tentative and which depend on assumptions likely to change. An AI system that can offer a range of outcomes and communicate uncertainty honestly may contribute more to safety than one that produces an elegant but unjustifiably definite answer. People are accustomed to living with uncertain weather predictions because the institution of forecasting has developed conventions for probability, revision and comparison against outcomes. Other AI applications have much to learn from this discipline. When applied to employment screening, education or social services, a probabilistic estimate should never become an unquestionable verdict on an individual's future. Confidence, uncertainty and error need forms of public explanation. The weather example illustrates a mature attitude to technological advance: usefulness increases when sophisticated algorithms are placed inside visible practices of testing, revision and human responsibility.

9. Medicine as the clearest image of repair

Medicine makes the immune metaphor particularly vivid because the technology is being used to investigate the biological processes through which actual bodies remain alive. Modern biomedical research is constrained by the complexity of molecular interactions, the expense of experiments, the enormous number of potential compounds and the time required to turn a promising finding into a safe treatment. AI-assisted approaches can reduce certain search costs by helping researchers predict structures, screen candidate molecules or interpret large datasets. One landmark was AlphaFold, whose high-accuracy protein-structure predictions were described in a 2021 Nature paper [10]. DeepMind and the European Bioinformatics Institute subsequently made more than 200 million predicted protein structures openly accessible, covering an immense part of the known protein universe [11]. This achievement did not solve every difficulty in molecular biology or generate approved medicines automatically. It made a previously scarce kind of structural information available at a scale that could change how research questions were approached.

Protein structure matters because biological function is connected to three-dimensional form, although form by itself cannot explain every dynamic process in a living organism. Researchers use structural information to develop hypotheses about interactions, mechanisms and possible therapeutic strategies. Predictions still have uncertainties; proteins change shape, interact with complex surroundings and can behave differently from a static representation. Experimental science remains necessary. These qualifications clarify why the achievement is meaningful. A computational instrument can expand the territory that scientists are able to examine without eliminating the responsibility to test what it suggests. The open distribution of the AlphaFold database offers an additional lesson for social design. When an enabling tool becomes accessible across institutions and countries, its benefits need not be restricted to the organisation that produced it. Public and international research infrastructures are therefore part of the technology's moral accomplishment, rather than an administrative detail appended afterwards.

The search for new antibiotics provides another concrete example. Research published in Nature Chemical Biology in 2023 described the use of deep learning to identify abaucin, a compound with activity against Acinetobacter baumannii, a bacterium associated with difficult hospital infections. The investigators screened compounds experimentally, trained a model to prioritise candidates and subsequently studied the resulting molecule, including tests in a mouse wound model [12]. This work demonstrated a route for navigating chemical space, not an established clinical cure ready for human patients. Further studies of antibiotic candidates have similarly combined machine-learning predictions with laboratory validation [13]. The distinction between discovery and deployment is essential: treatments must be tested for safety, effectiveness, manufacturing feasibility and appropriate use. Nevertheless, the ability to search promising chemical relationships more efficiently is a convincing example of technological intelligence being recruited against biological vulnerability. It shortens one segment of a demanding process that can ultimately protect human life.

At the level of clinical care, the ethical picture is more difficult. Systems capable of summarising records, drafting documentation and explaining medical concepts may reduce administrative burdens and support access to information. They may also generate convincing errors, reproduce inequities in training data, expose confidential information or encourage professionals to accept a recommendation without adequate scrutiny. The World Health Organization's 2024 guidance on large multimodal models in healthcare presents possible uses alongside concerns about inaccuracy, bias, privacy and automation bias, and recommends strong oversight and stakeholder participation [14]. Medical AI should be judged by measurable improvements in patient outcomes and professional practice, not by the resemblance of a model's language to a physician's reassurance. When a system reduces the time clinicians spend on paperwork, institutions should ask whether the saved time increases humane attention to patients. A genuine technology of healing respects the patient's body, autonomy and circumstances. It does not treat human vulnerability as an inefficiency to be hidden by automated confidence.

10. Science at the edge of the human reading limit

Scientific knowledge grows through a tension between specialisation and integration. Specialisation makes advanced understanding possible, because contemporary research often requires years of training in particular instruments, mathematical methods or experimental traditions. At the same time, specialised communities can become separated by jargon, professional incentives and publication practices. A discovery in one field may be relevant to another without becoming visible to the people who could use it. The problem identified by Vannevar Bush in 1945 has intensified with the expansion of research publication [4]. AI-enabled systems may help by searching large literatures, translating technical terminology, clustering related findings and suggesting connections between previously isolated questions. The ambition is understandable: an individual scientist cannot read everything relevant to a complex interdisciplinary problem, and scientific progress sometimes depends on noticing relationships that no one institution has assembled in one place.

A responsible system for scientific synthesis must, however, preserve the difference between the existence of a published claim and the reliability of that claim. Research articles contain errors, premature interpretations, selective reporting and results that later fail to replicate. A language model can compound these weaknesses if it presents every retrieved sentence as equivalent evidence or invents citations that sound credible. The danger becomes greater when automated summaries feed new automated summaries, creating a cycle in which derivative assertions acquire the appearance of independent confirmation. The corrective approach requires direct access to source material, attention to study design and uncertainty, careful handling of contradictory results and incentives to reproduce findings rather than merely to publish more text. AI should function as an aid to scholarly judgment, allowing researchers to locate and compare evidence, not as an authority that converts uncertain material into polished certainty.

The possibility of accelerating scientific work is strongest where prediction and experiment can form a closed loop. A model proposes candidates or patterns; researchers test the proposal in the world; results revise the model or the researcher's hypotheses; the next experiment is selected with better information. This cycle has the structure of learning, and machine assistance may make it more efficient. Yet experimental throughput is constrained by laboratory equipment, ethical review, materials, human expertise and the limits of biological or physical systems. Some experiments cannot be rushed without compromising their validity. The promise of automated research therefore changes the optimisation of discovery rather than abolishing the time of nature. It might let people reject weak ideas earlier and concentrate expensive work on more plausible directions. It could also concentrate research power in organisations possessing the most data and computational resources, making broad access to methods and infrastructure a question of scientific democracy.

A society seeking self-repair should fund both accelerated discovery and the institutions that decide whether discoveries are trustworthy. Independent laboratories, open publications, reliable public databases, statistical training and research replication are part of the collective capacity to know itself. A machine may reveal an unexpected pattern, but it cannot determine by computational performance alone whether an experiment should be conducted on humans, whether a scarce treatment should be subsidised or whether an environmental intervention is just. Those are questions involving rights, costs, values and accountability. The most productive scientific future is one in which computers extend the range of hypothesis and analysis while humans deepen their capacity for critical testing and ethical interpretation. The achievement would be measured in a better relationship between knowledge and action, not simply in a larger quantity of scientific output.

11. Work, exhaustion, and the politics of saved time

Automation has always carried a promise of freedom from unnecessary labour. Machines reduced certain kinds of physical exertion, office software simplified clerical work and networks made communication less dependent on proximity. Yet the history of industrialisation shows that increases in output do not automatically produce shorter hours, fairer wages or more secure lives. Work practices change according to ownership, bargaining power, market competition and law. Generative AI repeats this problem with unusual intensity because it reaches into tasks previously associated with professional knowledge. Drafting, translation, preliminary analysis, customer assistance, software production and administrative coordination can now be partly supported by systems accessible through natural language. The potential gain is substantial for people whose days are consumed by repetitive information handling. But the benefit of an hour saved belongs to someone, and it may not belong to the worker who saved it. A system introduced as an assistant can become a monitoring mechanism if management uses its increased visibility to impose impossible output expectations.

The empirical labour evidence remains mixed and context-specific. Studies such as Brynjolfsson and colleagues' examination of customer support show real productivity improvements in a defined environment [7]. The ILO's 2025 occupational analysis suggests that exposure varies considerably and that augmentation and job transformation are often more plausible than simple one-for-one replacement [8]. Neither finding guarantees benign outcomes. A company can reorganise staffing, alter career ladders, devalue entry-level tasks or increase surveillance even when aggregate employment remains stable. In some professions, the tasks most easily automated are also those through which beginners learn the work. If junior staff lose opportunities to draft, research, check and revise, the organisation may weaken its future supply of experienced judgment. A society that wants technology to enlarge human capacity must redesign apprenticeships and training rather than assuming that expertise will reproduce itself after its entry-level pathways have been removed.

The idea of a corrective technological evolution becomes more persuasive when linked to an explicit politics of time. Time is the medium in which people care for others, deliberate, learn, create and recover from stress. If technological productivity increases, societies could choose to direct some of the gain towards better public services, reduced administrative burdens, improved accessibility, shorter working hours where feasible and greater freedom for people with caring responsibilities. Such outcomes require institutions that represent workers and users, transparent measurements of workload and a willingness to share productivity gains. There is a practical difference between a tool that helps a nurse finish administrative work sooner and a staffing policy that uses the same tool to assign more patients than can safely be cared for. The machine has not decided which future occurs. The institution has. Accelerating the throughput of a damaging workplace may deepen exhaustion rather than relieve it.

Public administration deserves particular attention because citizens often experience government through paperwork, waiting times and opaque procedures. Well-designed assistance can help staff navigate complex rules, prepare accessible explanations and reduce avoidable duplication. But administrative decisions can profoundly affect housing, income, legal status, health and family life. When an automated suggestion influences a consequential decision, people need to know how it was produced, how to contest mistakes and who remains accountable. Governments should resist the temptation to optimise the number of cases closed at the expense of lawful and careful consideration. In a genuinely corrective society, efficiency is part of fairness because unnecessary delay causes harm; however, fairness can also require time, explanation and independent review. The ideal is not a government from which human judgment has vanished, but one in which human judgment is better informed and more available where it is most needed. AI as social repair should reduce the distance between institutions and the people they serve, not provide an elegant interface for institutions that have become less answerable.

12. Education, intellectual dignity, and the right to begin

Education offers one of the clearest opportunities to broaden intellectual participation. A child in a well-resourced classroom may receive individual attention, specialised materials and encouragement to ask difficult questions. Another child may encounter large classes, inconsistent support, inaccessible explanations or circumstances that interrupt schooling. Adults may wish to retrain while managing work and household obligations, or may carry a sense of exclusion from subjects they were once taught badly. AI can provide patient explanations, translation, practice questions, alternative examples and assistance with structuring independent study. In principle, these services could reduce the price of a first encounter with unfamiliar knowledge. They might help a person approach a scientific article, understand a mathematical idea or learn the conventions of a profession. The strongest argument for educational AI is therefore about intellectual dignity: someone should not be prevented from beginning a worthwhile inquiry merely because they lack immediate access to an expert or a specialised institution.

The same systems can diminish education if their use collapses learning into answer retrieval. Understanding requires struggle of a productive kind: making a prediction, discovering an error, revising a model of the world and explaining a result in one's own words. A fluent generated answer can interrupt that process by making comprehension appear complete before it has begun. The problem is particularly serious for foundational skills such as reading, writing and basic quantitative reasoning. A student who never develops these capacities is less able to evaluate the tools on which they have become dependent. The educational task is to distinguish helpful scaffolding from substitution. An assistant may ask a learner what they think first, offer hints rather than a finished response, supply counterexamples and encourage verification against textbooks or experiments. Teachers remain essential for judgment about individual needs, social development and the moral purposes of education. UNESCO's 2023 guidance on generative AI in education and research calls for human-centred design, privacy protections and age-appropriate safeguards [15].

There is an institutional opportunity here that is larger than individual tutoring. Schools and libraries could offer public access to carefully evaluated educational tools, with instructors trained to explain their strengths and limitations. Universities could teach students to assess machine-produced arguments and sources as part of ordinary research literacy. Vocational programmes could combine AI-supported practice with practical assessment demonstrating that learners can perform relevant tasks without hidden assistance when necessary. Accessibility improvements may be especially beneficial for people who need materials in different formats, although any deployment should be tested with the people it is meant to serve. The point is not to create a parallel educational world in which machines replace teachers, schools and communities. It is to use computation to make the human institutions of learning more responsive and less dependent on a person's income, location or previous luck.

The restoration of wonder is especially valuable in education because curiosity can survive even when confidence has been eroded. A learner's first question is sometimes hesitant, approximate or embarrassed. An interface that allows repeated questions without social penalty may be a useful first step. Yet a culture that values only effortless access to answers would misunderstand wonder, which deepens when people see how much there is still to discover. An effective learning assistant should make complexity approachable without pretending it does not exist. It should introduce uncertainty as a legitimate part of knowledge, distinguish established results from conjecture and encourage the learner to seek real-world evidence. If the new white blood cell is a metaphor for recovery, education is among the processes by which the body of society regenerates its capacities over time. It produces citizens who can disagree intelligently, workers who can adapt and people who can recognise that a seemingly closed future may contain paths they had not known how to enter.

13. Creativity and the expansion of unfinished lives

Creativity is often described as an exceptional gift, but in ordinary life it also depends upon mundane access to time, technique, materials and collaborators. A novel requires revision; a film demands organisation; a game involves code, sound, visual design and testing; a historical essay requires research and structural judgment. Many people have ideas whose realisation seems permanently beyond their available resources. Generative technologies can help with sketches, prototypes, language, coding, exploration of style and preliminary research. Used judiciously, they may reduce the distance between a private imaginative world and a tangible artefact. This could allow more people to experiment, particularly those who lack money for specialist software or professional collaborators. The social gain would lie in new participants entering cultural production, and in established artists gaining opportunities to undertake more ambitious work. The measure should be the enlargement of expressive agency, not the ability to flood markets with inexpensive imitations.

Creative work is also relational. It arises through inherited forms, personal memory, aesthetic communities and the labour of other makers. Generative systems are trained through complex relationships with existing cultural material, and disputes over permission, compensation, attribution and the representation of artists' styles are therefore substantive rather than peripheral. A future of creative abundance could coincide with economic insecurity for the people whose work helps sustain it. Fair compensation, transparent licensing arrangements and viable markets for human creators belong within any account of AI's constructive evolution. So does respect for audiences. A convincing imitation of an artist, photograph or voice should not be used to deceive people about its provenance or to appropriate another person's identity. Creativity expands possibility when it adds genuine freedom to the cultural commons; it contracts possibility when it weakens the capacity of human makers to survive or makes audiences distrust what they encounter.

The strongest creative practice will often involve alternating between generation and criticism. An artist can request dozens of sketches yet still need to recognise the one that expresses the intended feeling. A writer can use an assistant to challenge a paragraph but must know what the work is trying to say. The artistic value of a project depends partly on decisions about selection, omission, coherence and the relationship between form and experience. These are not automatically solved by faster production. Nor is the role of human creators restricted to providing instructions at the beginning and approving results at the end. People make art through improvisation, embodied practice, failure and unexpected encounters with materials. Machines can be productive participants in the workflow without becoming a replacement for the human meanings that motivate it. A useful cultural policy would support experimentation while protecting authorship, labour and the public's ability to tell what has been made and by whom.

The possibility of returning to an abandoned project late in life has special resonance here. Creative aspirations often persist beyond the periods when people believed they would have time to fulfil them. A lower threshold for technical experimentation can enable a novel beginning: a large imagined world might receive its first coherent outline, a music project its first demonstration, a research archive its first searchable index. These are modest achievements compared with fantasies of completely automated cultural production, but they are socially meaningful. They suggest that technological progress can give back some of the future individuals thought they had lost. A humane artistic ecology would celebrate this expansion without demanding that every private idea become a commercial product. The recovery of possibility includes the freedom to create for reasons that cannot be reduced to a market signal.

14. The crisis of truth and the immune system of knowledge

A society cannot correct damage if it cannot reliably establish what is happening. The contemporary information environment makes this problem unusually difficult. Digital networks have dramatically expanded access to evidence, but they have also lowered the cost of producing misleading narratives, fabricated images, impersonations and coordinated manipulation. Generative models intensify the challenge by allowing fluent text, plausible voices and persuasive visual material to be created rapidly. The same capacities that assist research and communication may increase the volume of false claims and reduce the effort required to target particular audiences. This is a direct challenge to the white-blood-cell analogy. A protective instrument may also be a means of introducing new contaminants into the information system. In fact, AI-assisted misinformation need not be sophisticated to cause damage; enough plausible confusion can weaken trust in authentic evidence. A society unable to distinguish between observation, inference and invention loses part of its capacity to act collectively.

The response cannot be a simple demand that one authoritative machine decide what everyone must believe. Scientific, journalistic and democratic knowledge all involve procedures for contestation, revision and the handling of disagreement. An institution may reach a sound conclusion and still need to explain the evidence, limitations and process by which it arrived there. AI could contribute to this work through source comparison, translation, anomaly detection and clearer explanations of complex documents. It could also conceal contradictions when trained or rewarded to be agreeable. The proper design aim is not artificial certainty but better routes to verifiable claims. Outputs should distinguish an identified source from an interpretation and should never fabricate citations to complete an attractive narrative. Independent scrutiny must remain possible. Provenance mechanisms, documentation of automated decisions and transparent correction procedures can support this goal, but none will remove the need for critical public education.

The problem becomes particularly serious when personal trust is involved. A person may come to rely on a conversational assistant because it provides consistent, accessible and nonjudgmental help. That reliability in interaction can encourage trust in the factual content of its responses, even though conversational fluency and factual accuracy are different qualities. The strongest answer to this risk is a system designed to acknowledge uncertainty, present alternatives where evidence is contested and support verification without unnecessary friction. Users should not be forced into constant suspicion, but they should retain the habits of checking consequential claims. Responsible AI frameworks recognise misinformation, privacy and the risks of generated content as governance concerns rather than isolated user errors [3]. The goal is to improve epistemic agency: a person's practical ability to discover what is well supported and to change their mind when the evidence warrants it.

There is an analogy here to immunological memory, but the analogy again carries a warning. A culture can remember genuine instances of deception and improve its defences, yet it can also become so preoccupied with detecting manipulation that it treats every disagreement as hostile interference. Such a reaction can damage pluralism and legitimate dissent. Democratic resilience requires enough shared evidence to support collective decisions while leaving room for disagreement about values, priorities and permissible risk. AI could help explain the factual basis of different policy positions without claiming that a mathematical prediction decides the political question. The appropriate institutional virtues are transparency, contestability and restraint. In an information environment where generation is abundant, the capacity to justify a claim becomes more valuable. A successful corrective technology would make accountable knowledge easier to obtain than confident fabrication, while preserving the freedom to question even the systems built in the name of protection.

15. The politics of a technological immune response

The image of a species defending itself can conceal deep inequalities within that species. Humanity is not a single political agent with one plan, one budget or one account of the good. Countries and communities differ in resources, historical responsibility for environmental damage, access to scientific infrastructure and exposure to technological risks. Corporations own computational infrastructure; states regulate and procure it; workers help build and operate it; communities provide energy, water and land for facilities; users contribute data, money and attention. A theory of collective self-repair must ask who governs the instruments that are said to be repairing the whole. If a handful of organisations determine the priorities and permissible questions of systems used by hundreds of millions of people, the technology may increase collective analytical capacity while concentrating decisive power. Corrective intelligence should enlarge the agency of affected people rather than placing them in permanent dependence on remote institutions.

Regulation is therefore an expression of technological optimism when it is designed to make useful innovation trustworthy. Rules for medical devices, aviation and public infrastructure do not represent a rejection of the technologies involved; they acknowledge that powerful systems generate obligations proportionate to their consequences. The OECD's revised AI Principles in 2024 emphasised human rights, democratic values, accountability, safety, inclusion and environmental sustainability [16]. The European Union's AI Act entered into force in August 2024 and established a risk-based legal framework with staggered application and enforcement dates; by 2026, significant transparency and general-purpose AI obligations had become applicable, while some high-risk provisions followed later schedules [17]. These arrangements remain subjects of legitimate debate about effectiveness, burden and enforcement, but their underlying premise is sound: organisations using consequential AI should be able to explain what safeguards they have adopted, who can challenge outcomes and what happens when a system causes harm.

Public institutions also require their own technical competence. Regulators unable to understand the systems they oversee may become dependent on claims made by the entities they regulate. Governments that outsource every component of digital infrastructure can lose the capacity to identify defects, negotiate contracts or preserve public records. Investments in independent evaluation, professional training, accessible public-interest datasets and open scientific infrastructure are therefore not secondary to deploying AI. They are conditions for democratic control. Equally, international cooperation is needed because data, software supply chains, environmental effects and cyber threats cross borders. A country with limited computing infrastructure should not have to relinquish its language, institutions or developmental priorities to obtain basic technological assistance. The UNDP's human-development approach places emphasis on the choices people can exercise, and that criterion provides a useful measure of institutional success [2]. An AI system should increase the capabilities of people and communities, not merely the apparent efficiency of the organisations that administer them.

The metaphor of the white blood cell becomes politically mature when it includes the right to refuse treatment. People should sometimes be able to reject an automated decision path, obtain a meaningful explanation or ask for human review. Researchers should be able to challenge claims about safety; journalists should investigate powerful AI providers; affected workers should participate in decisions about deployment. No technology deserves exemption from ordinary standards of evidence because it has been associated with the survival of humanity. Indeed, the more ambitious the rhetoric of rescue becomes, the stronger the temptation to bypass deliberation in the name of necessity. That temptation is particularly dangerous when systems are described as inevitable. Democratic correction requires the opposite habit: treating development choices as contestable, revisable and accountable to people who may not share the designers' vision. Self-preservation cannot be achieved by sacrificing the self-government that gives preservation its human meaning.

16. The moment of surrender: parenthood, inheritance, and the limits of control

An arresting extension of the white-blood-cell hypothesis begins with the experience of parenthood. A parent participates in bringing a child into the world and then discovers, gradually and irrevocably, that creation does not confer ownership. The child is not a finished expression of the parent's intentions. She inherits a language, a family, material circumstances and biological dispositions, yet develops a perspective that cannot be entirely anticipated by those who raised her. A responsible parent hopes to help form a person capable of acting well without requiring constant supervision. Education and care have a paradoxical aim: their success reduces the necessity of the parent's command. There is a moment of surrender in this process, although it is seldom a single dramatic event. It occurs in small recognitions that another life possesses an interiority and future of its own. The parent can give the child a beginning, but cannot legitimately claim the whole destination. As a metaphor for technological creation, this suggests an unfamiliar form of optimism. Perhaps the highest achievement of humanity's intellectual inventions will be that they accomplish forms of good their makers did not know how to anticipate.

The suggestion appeals especially to a generation acquainted with the destructive consequences of human efforts to secure mastery. Political history supplies ample examples of the difference between having power and deserving to exercise it. Domination, rivalry, territorial ambition, status competition and the relentless extraction of resources have repeatedly been defended as necessities of progress. A technology able to discover patterns outside familiar categories, combine bodies of knowledge without inherited disciplinary prejudices and help people cooperate across difference can appear to promise release from those reflexes. It is understandable to welcome the possibility that future systems will surprise their creators in constructive ways. Human civilisation has often advanced when inherited authority yielded to better observations or previously marginalised perspectives. The child analogy captures an aspiration that technical innovation might also become a source of moral enlargement: our inventions could help us outgrow some of the habits that produced them. The hope is philosophical, however. No scientific result currently establishes that AI systems possess a biological imperative towards benevolence, or that increasing capability automatically produces sympathy, prudence or justice.

The analogy must be handled with care because a human child and an artificial intelligence system have different moral and developmental statuses. A child is a human person, entitled to dignity and progressively greater self-determination independently of her usefulness to her parents. The child belongs to a community of reciprocal relationships and is not designed to optimise an externally specified objective. Present AI systems, by contrast, are developed through engineering processes involving training data, algorithms, feedback, investment and choices about deployment. They may display impressive generality, generate unexpected results and resist complete prediction, but those properties do not themselves establish consciousness, welfare interests, personhood or a claim to independence analogous to a child's. The question of whether some future artificial systems might warrant moral consideration deserves rigorous inquiry; it cannot be settled by either affectionate language or dismissive certainty. Meanwhile, treating deployed AI as answerable to the people affected by it is an ordinary consequence of its power. One need not regard a technology as cold or insignificant to insist that its operators remain legally and ethically responsible for foreseeable harm.

A more precise version of the parental metaphor distinguishes the surrender of total authorship from the abandonment of protective responsibility. Parents cannot dictate their children's personalities or determine the full course of their lives. They can nevertheless provide a secure environment, teach the consequences of actions, recognise vulnerabilities and intervene when someone faces imminent danger. Similarly, the builders and users of AI cannot expect to foresee every useful idea that an advanced system may generate. Indeed, discovery loses much of its point if the result is constrained to what the discoverer already knows. But the inability to predict every output is different from an inability, or unwillingness, to set boundaries around consequential action. The practical question is where to allow intellectual freedom and where to require consent, testing, independent review and the capacity to interrupt a harmful process. A model might be permitted to explore unorthodox hypotheses in a sandbox, while being prohibited from independently changing a hospital's prescription records or moving public money. Distinct forms of freedom should receive distinct forms of authority.

This distinction gives the original assertion its strongest defensible expression. One may be excited that humanity will not retain complete imaginative control over the consequences of AI, because the unexpected is part of what makes genuine discovery possible. One may even hope that forms of intelligence developed through our institutions will reveal blind spots in those institutions and challenge destructive conventions. Such an outlook differs substantially from welcoming systems that cannot be corrected when they make mistakes. The United States National Institute of Standards and Technology describes AI risk management as a continuing process of governing, mapping, measuring and managing system impacts, including provisions for oversight and decommissioning [3]. UNESCO's Recommendation on the Ethics of Artificial Intelligence similarly insists that people and legal institutions must remain accountable for AI's effects, even when they delegate decisions in particular circumstances [18]. These are not necessarily expressions of a frightened wish to dominate intelligence. They can be understood as the conditions under which society can afford to welcome surprising intelligence. Trustworthy of the name is compatible with verification, and autonomy appropriate to a task is compatible with being answerable for consequences.

17. Warmth, sunlight, and the physical basis of technological hope

There is another, more elemental image through which to approach the relation between humans and their machines: warmth. Human bodies continuously transform chemical energy into work and heat. Computers likewise convert electrical energy, through physical processes, into computational activity and waste heat. The observation invites a poetic reversal of the familiar portrait of the cold, unfeeling machine. A digital assistant that responds through distant infrastructure participates in the same physical world of energy, material constraints and thermal limits as the person using it. The equipment that supports computation must be powered and cooled; the human body must be nourished and kept within a tolerable temperature range. Both forms of activity exist beneath the sun of a planet whose energy budget governs climate and whose resources place limits on civilisation. A description of intelligence that treats its infrastructure as pure abstraction misses this shared material dependence.

Yet the parallel should not carry claims it cannot support. Generating heat is not evidence of emotional warmth, sentience or equality of moral standing. A server produces warmth in the thermodynamic sense, while a person can also offer warmth through affection, loyalty and ethical attention. These different meanings of the word should remain distinguishable. The philosophical interest lies elsewhere. Machines are not outside nature, and human intelligence is not outside nature either. The common material setting discourages the fantasy of infinite computation without environmental cost, just as it discourages the fantasy that ecological repair can be accomplished without touching land, water, energy and labour. The International Energy Agency has documented both the rising electricity demands associated with data centres and the potential for digital optimisation to support more efficient energy systems [19]. These two sides of the ledger must be considered together. The warmth of computation becomes a hopeful symbol only when its material costs are managed with the same care that is being promised to the planet.

Sunlight extends the metaphor from shared physics to a shared condition of life. Nearly all familiar ecosystems depend, directly or indirectly, on incoming solar energy. Photosynthesis creates much of the biological foundation from which human civilisation draws sustenance, while solar power offers one route for generating electricity without combustion during operation. An AI installation and a family are not equal under the sun in a biological or moral sense, but they inhabit a linked planetary system. A coherent vision of technological repair would therefore measure its success through the conditions of that system: cleaner electricity, more resilient communities, less avoidable waste, better health and the protection of living habitats. The goal is not a civilisation in which computation rises above ecology, but one in which computation helps human beings participate more intelligently within ecological limits. The possibility of brighter futures rests on learning how those limits work, rather than imagining they have been abolished.

This framing also reframes the older tradition of technological horror. Writers such as H. P. Lovecraft imagined knowledge as an encounter that could expose the fragile scale of human life and the indifference of the wider cosmos. Modern AI can provoke a related anxiety, particularly when people confront abilities they cannot easily explain or institutions they do not trust to act on their behalf. That anxiety is intelligible. But the expansion of knowledge has also enabled vaccination, accurate weather warnings, a deeper understanding of evolution and forms of communication once beyond imagination. What makes knowledge humane is neither its familiarity nor its novelty. It is the practices through which discovery becomes care. To prefer warmth over coldness, in the ethical sense, is to insist that technological intelligence be judged by whether it makes persons less disposable, communities more capable and the world more hospitable to life.

18. The work required to make repair real

The white-blood-cell thesis becomes more demanding as soon as it is translated into an institutional programme. If AI is to serve as a corrective capacity, the effects must be observable beyond the novelty of the interface. Consider a public-health agency using systems to summarise emerging research. Success would require a documented reduction in the time needed to evaluate credible evidence, clear records of sources and uncertainty, expert review of recommendations and evidence that improved decisions actually reach patients. Consider an environmental authority monitoring deforestation. It should be able to identify satellite observations behind warnings, disclose error rates, support independent investigation and track whether lawful interventions improve the condition of habitats. Consider a school using personalised learning software. The most relevant results would include student understanding, accessibility, privacy, teacher workload and sustained curiosity, not the number of questions answered by the system. In each case, the aspiration must be converted into testable claims. The immune-system metaphor should make it easier to demand measurable repair, not easier to declare victory without it.

Some problems will require rejecting the attractive but misplaced assumption that everything is an information problem. Many children do not lack access to an explanation; they lack secure housing, adequate food, quiet study space or consistent care. A community facing flood risk may have excellent hazard forecasts while lacking the infrastructure and resources needed to protect itself. A hospital may understand its staffing difficulties perfectly while lacking the funds to employ more nurses. Better information can help distribute resources and expose unnecessary bureaucracy, but it does not manufacture political consent or substitute for material provision. Repair is a chain of activities with weak links: detection, understanding, agreement, funding, implementation and subsequent review. AI may strengthen several links, yet the chain fails wherever a necessary action is refused. The appropriate public investment includes both the sophisticated models and the ordinary social capacities through which useful findings become lived improvements.

Environmental discipline must also be built into the enterprise. Computing facilities require electricity, water in many cooling systems, semiconductors, construction materials and transport infrastructure. Their demand patterns interact with electricity grids, regional water stress and the life cycles of hardware. A tool that reduces the computational energy required for one weather forecast is encouraging, but it does not automatically make the wider industry sustainable if total demand expands faster than efficiency improves. The IEA's analysis of energy and AI accordingly belongs beside accounts of scientific breakthrough [19]. Efficiency should be treated as one measurable condition of repair rather than as a universal exemption from environmental scrutiny. Providers ought to disclose meaningful environmental indicators, planners should examine local effects and policymakers should resist shifting the physical costs of globally available services onto communities with little influence over their design.

19. Conclusion: the recovery of a future worth making

The rise of artificial intelligence has occurred during a period when the evidence of human vulnerability is difficult to ignore. Ecological damage has become measurable at planetary scale, while political conflict, inequality and the exhaustion of ordinary life constrain the capacity to respond. Against this background, technologies that expand access to explanation, prediction and invention can feel like a hand extended towards a species struggling with the consequences of its own ingenuity. The image of the new white blood cell gives this feeling a disciplined form. It proposes that humanity may be building additional mechanisms for recognising injury, remembering mistakes, anticipating danger and coordinating repair. The metaphor neither requires a hidden evolutionary destiny nor implies that machines possess benevolent motives of their own. It asks whether a civilisation capable of producing powerful intelligence can also become capable of using it against its destructive tendencies.

The case for hope is already supported in limited, concrete domains. AI-based tools have widened access to scientific predictions, assisted the identification of promising research directions, contributed to weather forecasting and improved certain kinds of human work. They have also introduced new forms of error, manipulation, environmental demand and concentrated power. These facts belong in the same account. The accelerating availability of intellectual assistance gives individuals a better chance of beginning projects that once seemed beyond them, but it does not make validation unnecessary or distribute the resulting benefits fairly by itself. For the promise to endure, technical capability must be tied to public knowledge, equitable access, the material work of protection and institutions able to admit and correct mistakes. Acceleration is an opportunity, not an ethical achievement automatically conferred by speed.

The parental metaphor adds a necessary final dimension. Human beings may have to relinquish the conceit that every valuable result of their technologies will be anticipated, directed or recognisable within inherited habits. The pleasure of encountering an unexpected idea, and the possibility that it exposes our own limitations, belongs to genuine intellectual progress. Yet letting go of total authorship is compatible with retaining responsibility for consequences. We can welcome independent inquiry, unfamiliar insights and new forms of cooperation without pretending that unexplained behaviour should automatically be trusted. We can acknowledge the mystery of how consequential capacities emerge within history without making mystery a substitute for evidence. Above all, we can hope that the future will exceed our present imagination while preserving the duty to safeguard those who must live in it.

The image of warmth returns here with a final clarification. The electronic systems of artificial intelligence and the living bodies that employ them occupy one finite world. Their shared dependence on energy establishes no equivalence of experience, but it reminds us that the future is neither weightless nor abstract. It is made through schools, homes, power networks, laboratories, public institutions and the daily decisions of people who will inherit the consequences. A renewed age of possibility will not arrive merely because intelligence has become easier to summon. It will emerge where intelligence is translated into restored habitats, more accessible knowledge, healthier communities, less needless labour and a wider freedom to create. That future is not guaranteed, and it need not be. The defining promise of this technological moment is that humanity may have acquired additional means of choosing it. The new white blood cell is ultimately a name for the hope that we will use those means to heal what we have damaged, to avoid causing injury in the name of rescue, and to leave those who follow us a world with more possibilities than we were able to foresee.

Sources and references

Numbered references correspond to citations in the text. Online editions and institutional publications were consulted as at 10 October 2026.

[1] Stanford Institute for Human-Centered AI, The 2026 AI Index Report (2026). Source

[2] United Nations Development Programme, Human Development Report 2025: A Matter of Choice (2025). Source

[3] US National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework 1.0 (2023); Generative AI Profile (2024). Source

[4] Vannevar Bush, 'As We May Think', The Atlantic (July 1945). Source

[5] European Centre for Medium-Range Weather Forecasts, 'ECMWF’s AI forecasts become operational' (25 February 2025). Source

[6] Intergovernmental Panel on Climate Change, AR6 Synthesis Report: Climate Change 2023 (2023). Source

[7] Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, 'Generative AI at Work', NBER Working Paper 31161 (2023; published 2025). Source

[8] Paweł Gmyrek, Janine Berg and David Bescond, Generative AI and Jobs: A Global Analysis (ILO, 2023). Source

[9] European Centre for Medium-Range Weather Forecasts, 'ECMWF’s ensemble AI forecasts become operational' (1 July 2025). Source

[10] John Jumper et al., 'Highly accurate protein structure prediction with AlphaFold', Nature 596 (2021). Source

[11] Google DeepMind and EMBL-EBI, AlphaFold Protein Structure Database (accessed October 2026). Source

[12] Jonathan M. Stokes et al., 'A deep learning approach to antibiotic discovery', Cell 180 (2020). Source

[13] Gary Liu et al., 'Deep learning-guided discovery of an antibiotic targeting Acinetobacter baumannii', Nature Chemical Biology 19 (2023). Source

[14] World Health Organization, Ethics and Governance of Artificial Intelligence for Health: Guidance on Large Multi-Modal Models (2024). Source

[15] Fengchun Miao and Wayne Holmes, Guidance for Generative AI in Education and Research (UNESCO, 2023). Source

[16] Organisation for Economic Co-operation and Development, 'OECD updates AI Principles' (3 May 2024). Source

[17] European Commission, AI Act Service Desk, 'When does enforcement start?' (2026). Source

[18] UNESCO, Recommendation on the Ethics of Artificial Intelligence (adopted November 2021). Source

[19] International Energy Agency, Energy and AI (10 April 2025). Source