Abundant Intelligence

Introduction

Can AI make intelligence abundant? In a limited but already visible sense, yes: it can make some kinds of expert-like help cheaper, faster, and available to far more people. The strongest evidence is not that chatbots are flawless “digital geniuses”, but that AI can spread useful cognitive labour: explaining, drafting, coding, searching, tutoring, triaging, translating, modelling, and helping people navigate complex information. That matters because scarce expertise is a bottleneck in science, medicine, education, software, public services, and everyday life.

Overview image for Abundant Intelligence But intelligence abundance is not the same as wisdom, truth, or fair access. A model that gives plausible but wrong advice can make knowledge feel abundant while actually spreading error. A tool available only in rich countries, dominant languages, or well-resourced workplaces can widen gaps rather than close them. The bloom case is therefore conditional: AI could multiply human capability if it becomes reliable, affordable, well-governed, and broadly usable, not merely powerful. Stanford’s 2026 AI Index captures both sides of the moment: AI adoption is spreading rapidly, but the benefits are uneven and the most capable systems are often the least transparent. [Stanford HAI]hai.stanford.edu2026 ai index report2026 ai index report

What cognitive abundance means

“Intelligence abundance” means more than having a chatbot in every browser. It means that high-quality cognitive assistance becomes as available as electricity or search: always nearby, cheap enough for ordinary use, and useful across many domains. A student can ask for a patient explanation. A nurse can check a protocol. A farmer can compare crop disease symptoms. A small business can analyse contracts, taxes, and logistics. A scientist can explore a literature base or generate hypotheses. A disabled person can get real-time reading, writing, navigation, or communication support.

The important shift is from scarce attention to scalable support. Today, a good teacher, doctor, lawyer, engineer, translator, policy analyst, or research mentor is expensive and time-limited. AI does not need to replace those people to change the equation. Even if it mainly handles first drafts, routine questions, practice exercises, code suggestions, document search, and decision support, it can free human experts for harder cases and give non-experts a better starting point.

This is why cognitive abundance sits near the centre of the AI bloom thesis. Many human limits are not purely physical; they are limits of understanding, coordination, design, diagnosis, and learning. If AI reduces the cost of those functions, progress can compound. Yet the phrase “intelligence abundance” should be used carefully. Current systems are uneven: they can be brilliant at one task and unreliable at another, and they often lack the grounding, accountability, and situational judgement that human expertise provides.

Where expert help is scarce today

The case for abundant intelligence begins with a simple observation: most people do not have timely access to expert help when they need it.

Healthcare is the clearest example. The World Health Organization estimates a projected shortfall of 11 million health workers by 2030, mostly in low- and lower-middle-income countries. That does not mean AI can become a doctor. It does mean that the world has a severe shortage of medical attention, and any safe tool that helps with education, triage, documentation, translation, or clinician support could matter. [World Health Organization]who.intWorld Health Organization Health workforceWorld Health Organization Health workforce

Education has a similar bottleneck. One-to-one tutoring is one of the most powerful learning interventions, but it is too expensive to provide to every learner. AI tutors promise something closer to personalised explanation on demand: not a replacement for schools, but a way to give more pupils practice, feedback, and confidence between human interactions. In a 2025 randomised controlled trial, students using a carefully designed AI tutor learned more in less time than students in an active-learning class, while reporting higher engagement and motivation. The key phrase is “carefully designed”: this was not simply “let students use any chatbot”. [Nature]nature.comOpen source on nature.com.

Workplaces also contain hidden expertise shortages. New employees often struggle not because the knowledge does not exist, but because it is buried in documents, habits, and experienced colleagues’ heads. A major NBER study of 5,179 customer support agents found that access to a generative AI assistant raised productivity by 14% on average, with a much larger 34% gain for novice and lower-skilled workers. The researchers suggest that the tool helped spread the practices of stronger workers to newer ones. [NBER]nber.orgGenerative AI at Work | NBERGenerative AI at Work | NBER

That pattern is central to the bloom case: AI may be most transformative when it transfers tacit knowledge to people who were previously locked out of it.

Abundant Intelligence illustration 1

The strongest evidence is augmentation, not replacement

The evidence so far points less to instant artificial super-experts and more to practical augmentation. AI helps most when the task has clear feedback, relevant data, and a human or institutional process that can catch mistakes.

Software development is a useful test case because output can often be checked by running code, tests, and reviews. In a controlled experiment published by Microsoft Research, developers using GitHub Copilot completed a programming task 55.8% faster than the control group. Later field experiments across Microsoft, Accenture, and a Fortune 100 company found a 26.08% increase in completed tasks among developers given access to an AI coding assistant, with higher adoption and gains among less experienced developers. [Microsoft]microsoft.comThe Impact of AI on Developer Productivity: Evidence from Git Hub CopilotThe Impact of AI on Developer Productivity: Evidence from Git Hub Copilot

But the coding evidence also warns against naïve extrapolation. A 2025 randomised study of experienced open-source developers working on mature projects found that allowing AI tools increased completion time by 19%, despite developers expecting and later believing that AI had made them faster. This does not refute AI coding assistance; it shows that productivity depends heavily on task type, codebase familiarity, quality standards, and the review burden. [arXiv]arxiv.orgSource details in endnotes.

Science shows the most exciting form of augmentation: not replacing researchers, but widening the search space. AlphaFold DB provides open access to more than 200 million predicted protein structures, giving researchers across the world a resource that would have been impossible to build experimentally one structure at a time. Google DeepMind reports that AlphaFold is used by millions of researchers across more than 190 countries, including over one million users in low- and middle-income countries. [AlphaFold]deepmind.googleAlpha FoldAlpha Fold

Materials discovery tells a similar story. DeepMind’s GNoME system predicted 2.2 million new crystal structures, including 380,000 predicted stable materials, potentially useful for batteries, chips, solar cells, and other technologies. Yet prediction is not deployment: materials still need synthesis, testing, scaling, safety checks, cost reductions, and integration into manufacturing. [Google DeepMind]deepmind.googleGoogle Deep Mind Alpha Fold — Google Deep MindGoogle Deep Mind Alpha Fold — Google Deep Mind

These examples show what cognitive abundance could look like at civilisational scale: more hypotheses, more candidate designs, more simulations, more personalised explanations, and more people able to participate in specialised work. They also show the recurring bottleneck: the world must still verify, build, regulate, and distribute what AI helps imagine.

Why access decides whether abundance is real

An AI service can be technically available and still not create real abundance. Access has at least five layers: cost, connectivity, language, usability, and institutional permission.

The cost trend is encouraging. Stanford’s 2025 AI Index reported that the inference cost for a system performing at roughly GPT-3.5 level fell more than 280-fold between November 2022 and October 2024, while AI hardware costs declined and energy efficiency improved. Cheaper inference makes it more plausible that advanced help could reach schools, clinics, small firms, researchers, and public services rather than remaining a luxury tool. [Stanford HAI]hai.stanford.eduresearch and developmentresearch and development

Yet adoption remains unequal. The OECD reports that more than one-third of people across OECD countries used generative AI tools in 2025, but uptake differed sharply by age, education, and income. It also warns that generative AI could worsen regional divides, because urban workers are more exposed to and more likely to benefit from AI than rural workers. [OECD]oecd.orgGenerative AI | OECDGenerative AI | OECD

Education policy faces the same divide. UNESCO’s guidance on generative AI in education calls for a human-centred approach, including regulation, teacher capacity, inclusion, and long-term public planning. That matters because simply dropping AI into classrooms can reward students who already know how to ask good questions, check answers, and use feedback well. The pupils most in need of support may be the least able to extract value from an unguided system. [UNESCO]unesco.orgguidance generative ai education and researchguidance generative ai education and research

Language is another access barrier. If the best models work best in English and other high-resource languages, then abundant intelligence arrives first for people already close to the centre of global knowledge production. A bloom-oriented approach would prioritise multilingual performance, local curricula, culturally relevant examples, accessibility features, and public-interest deployments rather than assuming the market will automatically serve everyone.

Trust is the other half of abundance

Cheap expert-like help is only valuable if people can trust it appropriately. Too little trust, and useful systems are ignored. Too much trust, and people act on fluent errors.

Medicine makes this tension vivid. A 2026 Nature Medicine study tested whether large language models could help members of the public identify conditions and choose what to do in medical scenarios. The models alone performed well at identifying conditions, but participants using the models did no better than the control group; the authors identified human-AI interaction as a major deployment challenge and argued for systematic user testing before public healthcare use. [Nature]nature.comOpen source on nature.com.

That finding is crucial. It is not enough for an AI system to “know” the answer in a benchmark. Real people may ask incomplete questions, omit symptoms, misunderstand warnings, or over-weight a reassuring sentence. In high-stakes settings, abundant intelligence requires careful interface design, escalation rules, uncertainty signalling, audit trails, and professional accountability.

The same issue appears in education. Some studies show strong learning gains from structured AI tutoring, while others warn that students may copy answers or use tools in ways that weaken learning. The difference is not magic; it is design. AI can act as a Socratic tutor, asking questions and building understanding, or as an answer vending machine that short-circuits effort. A UK classroom trial using LearnLM on the Eedi mathematics platform found that expert tutors approved most AI-drafted messages with zero or minimal edits, and students supported by the AI system performed at least as well as those supported by human tutors alone. But the study used tutor supervision, a defined platform, and a specific pedagogical role. [arXiv]arxiv.orgSource details in endnotes.

For abundance to be real, trust must be calibrated. People need to know when AI is a calculator, when it is a junior assistant, when it is a tutor, when it is a search interface, and when it is out of its depth.

Abundant Intelligence illustration 2

What abundant intelligence could change first

The near-term effects are likely to appear where expert help is valuable but currently rationed.

In education, AI could provide more practice, feedback, explanation, and language support, especially for learners who cannot afford private tutoring. The best systems will not merely answer questions; they will diagnose misconceptions, adapt pacing, encourage effort, and help teachers see where pupils are stuck.

In healthcare, AI could reduce administrative load, support clinicians with guidelines and documentation, help patients prepare better questions, and improve access to basic health information. The strongest role is probably not “AI doctor for everyone”, but safer navigation: when to self-care, when to seek help, what information to gather, and how to understand a clinician’s advice.

In work, AI could make organisational knowledge easier to use. The customer support study matters because it suggests AI can compress the learning curve for newer workers. If that pattern generalises, AI may become a kind of workplace exoskeleton: not replacing skill, but helping people reach competent performance faster. [NBER]nber.orgeconomics generative aieconomics generative ai

In science and engineering, AI could widen participation by giving smaller labs, poorer universities, and independent researchers access to tools once reserved for elite institutions. AlphaFold is the clearest example because its database is openly available and widely used. The broader question is whether future systems for biology, materials, climate, robotics, and mathematics will be similarly open, or locked behind corporate and national barriers. [AlphaFold]deepmind.googleAlpha FoldAlpha Fold

In public life, AI could help citizens understand laws, benefits, forms, local planning, public budgets, and scientific disputes. That would be a quiet but important form of abundance: more people able to reason through systems that currently feel opaque.

The strongest objections

The first objection is reliability. If AI systems hallucinate, flatter users, hide uncertainty, or produce different answers for different groups, they can create an illusion of expertise. This is especially dangerous in medicine, law, finance, and education, where a wrong answer can shape real decisions.

The second objection is dependency. If people outsource too much thinking, they may lose the skills needed to judge the machine. A society of people using AI well is very different from a society that cannot function without opaque systems owned by a few firms.

The third objection is concentration of power. The most capable models require data, compute, talent, infrastructure, and distribution channels. Stanford’s 2026 AI Index notes that industry produced more than 90% of notable frontier models in 2025 and that several of the most resource-intensive systems no longer disclose key details such as training code, parameter counts, dataset sizes, and training duration. Abundant intelligence built on closed, concentrated systems could make users more capable while making society more dependent. [Stanford HAI]hai.stanford.edu2025 ai index report2025 ai index report

The fourth objection is unequal benefit. AI may help highly educated, well-connected users more than those with fewer resources. It may boost workers in firms that can redesign processes while leaving others with surveillance, deskilling, or job insecurity. OECD evidence on uneven uptake by age, income, education, and region makes this a central issue rather than a footnote. [OECD]oecd.orgOpen source on oecd.org.

The fifth objection is that cognitive abundance is not physical abundance. AI can design a better battery candidate, but mines, factories, grids, planning systems, finance, regulation, and supply chains still matter. It can suggest a treatment, but hospitals, staff, medicines, and trust still matter. Intelligence is a bottleneck, not the only bottleneck.

Why access and trust decide the outcome

The optimistic version of abundant intelligence is not a world where everyone asks a chatbot everything. It is a world where reliable cognitive support is woven into institutions that already carry responsibility: schools, clinics, labs, libraries, courts, public agencies, workplaces, and community organisations.

Three conditions matter most.

First, AI must be grounded. Systems should be connected to verified sources, domain-specific data, tests, and human review where stakes are high. They should show uncertainty and make it easy to inspect evidence.

Second, AI must be broadly accessible. That means low costs, multilingual support, disability access, public-interest tools, open scientific resources, and infrastructure for poorer regions. Without this, “abundance” becomes a productivity premium for those already ahead.

Third, AI must strengthen rather than bypass human capability. The best systems should teach, scaffold, explain, and invite challenge. They should help people become more competent over time, not merely more dependent on an answer machine.

If those conditions hold, AI could make intelligence feel less like a scarce luxury and more like a public resource. That would not solve every problem in the AI bloom vision, but it would loosen one of humanity’s deepest constraints: the shortage of timely, useful, trustworthy understanding.

Abundant Intelligence illustration 3