Expand all Collapse all AI Bloom Futures AI bloom is the optimistic but demanding idea that advanced AI could help humanity do far more than automate office work or raise GDP. In its strongest form, it means loosening deep constraints on human life: disease, scarcity, dangerous labour, poor education, slow science, energy limits, weak coordination, and vulnerability to catastrophe. 171 pages Read more Abundant Intelligence Cheap expert help could change science, education, medicine, and daily life if access is broad and reliable. 17 pages +Show subtopics Read more Access Gaps Cheap models are not enough: real intelligence abundance depends on cost, language, connectivity, usability, and public institutions. 4 pages +Show subtopics Read more Access basics AI expertise cannot become truly abundant while households, schools, and clinics lack the hardware, power, and connectivity to use it well. Read more Language gaps When AI systems work best in English and high-resource languages, whole communities may receive weaker education, healthcare, and public services. Read more Public access Schools, libraries, clinics, and welfare systems could spread AI benefits more evenly than consumer subscriptions alone. Read more AI Tutors AI tutoring could widen access to personalised help, but only if it is designed for real classrooms and unequal learners. 4 pages +Show subtopics Read more Access Gap The promise of affordable tutoring depends on devices, connectivity, teacher training, supervision, and funding models as much as the software itself. Read more Shortcut Risk AI tutoring only helps if it keeps learners thinking, explaining, practising, and struggling productively instead of outsourcing the work. Read more Trial Results Early trials suggest AI tutors can improve learning, but only when the systems are designed around real pedagogy rather than answer delivery. Read more Coding Tradeoffs Coding assistants can speed up some programming tasks while slowing others when review burden and project complexity rise. 4 pages +Show subtopics Read more Expert slowdown AI coding tools can slow experts when local knowledge, hidden conventions, and review work matter more than generating code quickly. Read more Review bottleneck When AI produces code faster than people can safely verify it, productivity gains can turn into review, debugging, and maintenance bottlenecks. Read more Technical debt AI can raise the volume of code while leaving behind fragile abstractions, duplicated logic, and maintenance costs that appear later. Read more Novice Workers Workplace AI may matter most when it transfers practical know-how from experienced staff to people still learning the job. 4 pages +Show subtopics Read more Dependency risk AI assistance can make novices faster in the moment while weakening the deeper judgement they need when the tool is wrong or unavailable. Read more Support agents The strongest workplace evidence suggests AI assistants helped newer support agents most by turning past chats into live practical guidance. Read more Tacit knowledge AI assistants may help beginners by surfacing the hidden tricks, phrasing, and pattern recognition that experienced workers rarely write down. Read more AI Medicine AI could help find disease earlier and speed drug discovery, but patients only benefit when evidence, safety, and access hold up. 17 pages +Show subtopics Read more AI Drugs AI can generate promising drug candidates faster, but clinical trials still decide whether they are safe and useful in people. 4 pages +Show subtopics Read more Formulation gap A molecule that looks strong in software still has to survive dosing, manufacturing, stability, delivery, and side-effect problems. Read more Phase II AI-designed drugs may clear early safety checks, but Phase II trials reveal whether the biology actually helps patients. Read more Rentosertib Rentosertib shows that an AI-designed medicine can reach patients and produce promising signals, without yet proving approval is likely. Read more AI Mammograms AI-supported breast screening could catch more dangerous cancers earlier, but only if false alarms and overdiagnosis stay under control. 4 pages +Show subtopics Read more MASAI trial The Swedish MASAI trial shows how AI may help screening by triaging low-risk scans while keeping human radiologists focused on harder cases. Read more Missed cancers Interval cancers are the crucial test of whether AI screening finds dangerous tumours earlier rather than just finding more abnormalities. Read more Too much screening More sensitive AI screening could save lives, but it could also increase false positives, biopsies and overdiagnosis if not carefully governed. Read more Fair Access AI medicine will not improve healthier lives broadly unless validation, monitoring and access include the populations most likely to be missed. 4 pages +Show subtopics Read more Live monitoring Medical AI needs ongoing checks after deployment because performance can drift as patients, diseases, and clinical workflows change. Read more Missing data Training data gaps can make AI less reliable for people whose bodies, records, languages, or conditions were poorly represented. Read more Sensor bias A biased bedside sensor can quietly pass false reassurance into AI triage, making unequal measurement a hidden source of unequal care. Read more Second Eyes Medical AI looks most useful when it helps clinicians notice, prioritise and summarise rather than pretending to replace judgement. 4 pages +Show subtopics Read more Automation bias AI oversight can become unsafe when busy clinicians accept confident machine recommendations instead of checking the evidence themselves. Read more Meaningful oversight Safe medical AI depends on evidence displays, audit trails, clinician training, and rules that keep responsibility with professionals. Read more Stroke triage Stroke AI can flag possible brain bleeds or vessel blockages faster, but hospitals still need clinicians to verify urgent alerts. Read more AI Tutors AI tutors could widen access to good teaching, especially when they support teachers instead of replacing them. 17 pages +Show subtopics Read more AI Literacy Students need practical habits for checking AI answers, spotting overconfidence and keeping enough effort in the learning loop. 4 pages +Show subtopics Read more Classroom Rules Clear classroom norms can separate useful AI assistance from substitution, while making student reasoning visible. Read more Hallucinations Fluent chatbot answers can invent sources, smooth over uncertainty and mislead students unless they learn how to check them. Read more Verification A small routine of checking sources, testing explanations and rewriting answers can turn chatbot use into active learning. Read more Human Tutors Tutor CoPilot suggests AI may be most useful when it helps human tutors ask better questions and avoid giving answers away. 4 pages +Show subtopics Read more Fewer Answers AI tutor support is most promising when it nudges adults towards hints, questions and productive struggle instead of quick answers. Read more Human Judgement AI can offer useful teaching prompts, but human tutors still have to judge student emotion, context and readiness in the moment. Read more Weaker Tutors The strongest equity case for AI-assisted tutoring is that it may help less experienced tutors make better live teaching decisions. Read more Nigeria Trial The Nigeria Copilot trial shows why structured, teacher-supported AI tutoring matters more than casual chatbot access. 4 pages +Show subtopics Read more Evidence Limits The headline gains were striking, but the trial’s short duration, narrow subject focus, and supervised design limit what can be claimed. Read more Pair Work Students used the AI in pairs, making discussion, explanation, and shared checking part of the learning process. Read more Teacher Role The Nigeria trial suggests AI tutoring worked best as guided classroom support, not as a replacement for teachers. Read more Tutor Crutches AI tutors can raise practice scores while weakening real learning if they make answers easier than thinking. 4 pages +Show subtopics Read more False Mastery Students can look fluent with AI help while losing the ability to recall, explain, or apply the same ideas without the tool. Read more Hint First AI tutors may teach better when they slow down, ask for attempts first, and give staged hints instead of instant solutions. Read more Retrieval A stronger AI tutor asks what learners remember before explaining, because recall itself helps make knowledge more durable. Read more Coordination AI could improve planning, disaster response, and negotiation, but it could also intensify misinformation and mistrust. 17 pages +Show subtopics Read more Disaster maps AI damage maps can shorten the gap between a disaster happening and help reaching the places hit hardest. 4 pages +Show subtopics Read more Drone imagery Hurricane drone deployments show how AI can turn overwhelming image volumes into faster building-damage clues for responders. Read more Flood maps AI-assisted flood maps can reveal cut-off communities sooner, but radar limits and local verification still shape where aid goes first. Read more Human checks FEMA’s human-in-the-loop approach shows why faster AI triage still needs accountable analysts before emergency decisions are made. Read more Peace deals AI negotiation tools may help mediators map incentives and commitments, but they cannot manufacture trust or legitimacy. 4 pages +Show subtopics Read more Monitoring AI-assisted monitoring may reduce uncertainty after a deal, but bad data, surveillance fears and contested evidence can also damage trust. Read more Trade offs AI policy simulators could help negotiators test linked concessions before fragile compromises are offered in public. Read more Yemen mapping Yemen’s peace talks show how AI could help mediators track shifting demands, actors and issue links across years of negotiation records. Read more Public capability Government AI helps only when pilots are backed by data quality, skilled staff, audit trails, and routes for redress. 4 pages +Show subtopics Read more Clean data Public AI becomes useful when records are accurate, standardised and shareable enough for agencies to coordinate around the same facts. Read more Human review The strongest public-sector uses often narrow attention for human officials instead of silently deciding benefits, enforcement or disaster aid. Read more Pilot failure AI pilots often impress in trials but stall when they meet legacy data, old procurement rules, skills gaps and everyday public-service workflows. Read more Trusted warnings AI can improve forecasts, but warnings only save lives when institutions and communities know how to act on them. 4 pages +Show subtopics Read more False Alarms Repeated false alarms can weaken trust, making even accurate AI flood warnings harder to turn into evacuation and protection. Read more Impact Warnings Impact-based warnings translate technical forecasts into practical choices about roads, schools, hospitals, homes, and evacuation. Read more Texas Floods The deadly Texas floods showed how accurate forecasts can still fail when local alerts, night-time communication, and preparedness break down. Read more Discovery Protein folding, materials prediction, and forecasting show how AI can shorten discovery loops without replacing experiments. 17 pages +Show subtopics Read more AI Forecasts GraphCast and GenCast show how fast forecasts can become testable models for farms, grids, transport and disaster response. 4 pages +Show subtopics Read more Physics guardrails AI weather models are powerful, but operational forecasting still relies on physics, observations and institutions that catch failure modes. Read more Storm warnings Faster AI forecasts could give emergency planners more time to act, but warning systems still depend on trust, validation and human decisions. Read more Uncertainty GenCast matters because it models many possible futures, making uncertainty itself a practical part of faster scientific prediction. Read more AlphaFold Access AlphaFold’s open structure database shows how AI predictions can widen who gets to ask serious biological questions. 4 pages +Show subtopics Read more Lab proof AlphaFold can make protein research faster, but its models remain hypotheses that need experiments before they can support strong biological claims. Read more Neglected disease Open predicted structures can give neglected-disease researchers a faster starting point, but lab access still decides what discoveries become real. Read more Open access The AlphaFold database lowers the cost of entry, yet unequal compute, expertise, laboratories and commercial power still shape who benefits most. Read more Autonomous Labs The A-Lab shows how robotics can close the gap between predicted materials and tested compounds, one experiment loop at a time. 4 pages +Show subtopics Read more Novelty dispute The A-Lab debate shows why making a compound is not the same as proving that science has discovered something new. Read more Scale up limits A material that works in a tiny automated experiment may still fail on cost, durability, safety, scale, or manufacturing practicality. Read more Validation gap AI can propose candidate materials far faster than laboratories can currently synthesize, test, and verify them. Read more Molecular Binding AlphaFold 3 moves AI biology towards molecular interactions, but binding predictions can still fail in flexible, living systems. 4 pages +Show subtopics Read more Confidence gaps AlphaFold 3 can produce persuasive interaction models whose confidence scores may miss large errors, especially in novel or flexible cases. Read more Drug discovery AlphaFold 3 may help researchers narrow drug candidates faster, but binding predictions still need laboratory proof. Read more Flexible proteins Many proteins behave like moving ensembles, making one confident-looking predicted structure an incomplete guide to biology. Read more Energy Limits Data centres, chips, grids, water, and land make clear that digital abundance still depends on physical infrastructure. 17 pages +Show subtopics Read more Clean Power AI firms are buying huge amounts of renewable power, but clean-energy contracts do not always solve local grid constraints. 4 pages +Show subtopics Read more Gas fallback Utilities may turn to gas when AI load arrives faster than clean firm power, storage and transmission can be built. Read more Grid bottlenecks AI campuses can be built faster than the wires, substations and approvals needed to move clean power into crowded grid regions. Read more Renewable PPAs Renewable power contracts can fund new clean energy, but annual matching may hide fossil-heavy local electricity use. Read more Efficiency Rebound More efficient chips and models may cut energy per task, yet falling costs could still drive much higher total demand. 4 pages +Show subtopics Read more Cheap Inference Cheaper AI queries can cut electricity per task while making AI common enough that total power demand still rises. Read more Useful Demand More electricity use is not always waste if AI enables medicine, robotics, cleaner industry and other valuable work. Read more Wasteful Uses If cheap computation encourages disposable or marginal AI features, governance may need to separate useful abundance from waste. Read more Firm Power Tech firms increasingly view nuclear and geothermal power as ways to supply constant low-carbon electricity for future AI systems. 4 pages +Show subtopics Read more Geothermal deals Google’s geothermal deals with Fervo and Ormat show why firm clean power is becoming central to AI data-centre strategy. Read more Power fairness Dedicated nuclear and geothermal deals could speed AI growth, but they also raise questions about who gets power, who pays and who benefits. Read more SMR timing Small modular reactors look ideal for AI campuses, but licensing, fuel supply and construction delays may keep them from solving the near-term power... Read more Grid Strain A single AI data-centre cluster can strain substations, transmission lines and water systems long before national grids run short of power. 4 pages +Show subtopics Read more Load Swings AI data centres can create grid risks not only by using huge amounts of power, but by changing demand abruptly at industrial scale. Read more Virginia Grid Northern Virginia shows how a successful AI hub can hit transmission limits before the wider power system runs out of electricity. Read more Who Pays The politics of AI infrastructure turn sharp when communities fear higher bills, new power lines and local burdens for global digital benefits. Read more Long Future AI bloom ultimately asks whether humanity can become safer, freer, more creative, and resilient across centuries or longer. 17 pages +Show subtopics Read more Beyond GDP A richer economy can still leave people unhealthy, insecure, isolated, or unable to shape their own lives. 4 pages +Show subtopics Read more AI tutors Cheap AI tutoring could expand knowledge and opportunity far more than its market price suggests. Read more Lost autonomy GDP can rise while people lose bargaining power, security, and control over the systems they depend on. Read more Preventive AI AI health systems may create real abundance by preventing illness before expensive treatment is needed. Read more Hard Limits Even extreme abundance would not remove rival goods such as land, attention, ecosystems, authority, and the limited time of a human life. 4 pages +Show subtopics Read more Attention When AI can generate unlimited content, trust, provenance, curation and human attention become central scarce resources. Read more Prime places Even with cheap construction and remote work, the most desirable climates, coastlines, cities and stable regions cannot be copied. Read more Wild nature As manufactured goods become abundant, intact forests, reefs, rivers and wild landscapes may become more valuable and more politically contested. Read more Know how Cheap AI guidance could widen access to education, diagnosis, law, design, and local problem-solving if infrastructure and trust keep up. 4 pages +Show subtopics Read more AI Tutors AI tutors could make patient, personalised help widely available, but weak systems may teach errors or deepen inequality. Read more Health Workers AI could strengthen frontline healthcare by helping community workers triage symptoms, keep records and know when to escalate. Read more Trust Limits Low-cost AI guidance only becomes real expertise when societies can verify it, assign responsibility and protect users from confident mistakes. Read more Medicine Access AI may speed discovery, but drugs remain scarce when trials, manufacturing, pricing, and health systems fail to deliver them fairly. 4 pages +Show subtopics Read more Broken Delivery Even cheap medicines can fail to save lives when clinics, transport, refrigeration, and staffing break down. Read more Insulin Access Insulin shows how patents, concentrated production, and weak health systems can keep an old medicine scarce. Read more Vaccine Gaps COVID-19 vaccines were developed quickly, but manufacturing limits and supply politics delayed global access. Read more Robots and Work Robotics could remove dangerous and degrading work, but wages, ownership, and worker bargaining decide who gains. 17 pages +Show subtopics Read more First Tasks A humane robotics agenda starts with the tasks workers most want to escape: hazardous, repetitive, degrading or body-wearing work. 4 pages +Show subtopics Read more Care Lifting Care robots may be most valuable when they lift, carry, and fetch, while leaving empathy and bedside judgement with people. Read more Extreme Sites The strongest case for automation is often where the alternative is sending people into unstable tunnels, fires, radiation, or rubble. Read more Sewer Robots Robots can inspect pipes and contaminated sites so fewer people have to enter toxic, cramped, or biohazardous spaces. Read more Robot Dividend The promise of robot abundance depends on whether workers can negotiate over pay, pace, data, retraining and redeployment. 4 pages +Show subtopics Read more AI monitoring AI monitoring can turn ordinary jobs into constant measurement, making transparency and human review central bargaining demands. Read more Shorter hours AI productivity gains could fund shorter working weeks, but only if workers have enough bargaining power to claim time as part of the dividend. Read more Works councils Germany’s works councils suggest worker voice can sometimes support automation when employees influence how technology is introduced and gains are... Read more Safety Gaps Many robot risks appear when people enter machine spaces for repair, testing or adjustment rather than during normal operation. 4 pages +Show subtopics Read more Cobots Collaborative robots can reduce some hazards, but setup, tooling, payloads and overconfidence can still make close human work risky. Read more Lockout failures Lockout-tagout is meant to make repair work safe, but robot cells can hide energy, linked machines and restart paths that defeat partial shutdowns. Read more Unexpected motion Crushing and trapping injuries often happen when a paused, jammed or manually jogged robot moves in a way the worker did not expect. Read more Warehouse Pace Warehouse robots can cut walking and lifting, but the real test is whether they lower injury risk or simply make human work faster. 4 pages +Show subtopics Read more Algorithmic quotas Digital targets can turn ergonomic gains into tighter pace pressure by measuring every scan, pause and pick. Read more Amazon dispute Amazon has become the defining case for whether robotics improves warehouse safety or locks workers into higher speed expectations. Read more Injury tradeoff Warehouse robots can reduce heavy-lifting injuries while creating new risks from repetitive, faster-paced work. Read more Shared Gains AI can widen inequality as easily as it can widen opportunity unless institutions spread power, income, and access. 17 pages +Show subtopics Read more AI Profits If AI’s biggest rewards arrive as profits, equity gains, cloud rents, and intellectual property income, ownership will shape who benefits most. 4 pages +Show subtopics Read more Compute rents Chips, cloud platforms and data centres can act like toll roads for the AI economy, letting infrastructure owners capture large profits from everyone... Read more Data flywheels When user data, compute and model quality reinforce each other, successful AI firms may become harder for rivals and public institutions to catch. Read more Owner gains If AI makes output soar while labour’s share falls, the biggest winners may be people who own the systems rather than people who work beside them. Read more Platform Lock in Cloud and model partnerships can speed deployment, but they may also create switching costs, exclusivity, and market power that limit broad access. 4 pages +Show subtopics Read more Bundling Rules Regulators are asking whether cloud giants can use AI assistants, APIs, and enterprise contracts to extend dominance across markets. Read more Microsoft OpenAI Microsoft and OpenAI show how a useful AI partnership can speed deployment while raising hard questions about dependence and bargaining power. Read more Switching Costs AI lock-in often grows through everyday integrations, staff training, data pipelines, and compliance systems rather than obvious exclusivity clauses. Read more Public Compute Public compute could give researchers, schools, start-ups, and poorer countries access to advanced AI without depending entirely on dominant... 4 pages +Show subtopics Read more AI factories Europe’s AI factories are meant to give researchers, start-ups, industry, and public services an alternative to dependence on foreign cloud giants. Read more Cloud lock in Exclusive cloud deals can make AI builders dependent on a few infrastructure owners through spending commitments, privileged access, and control... Read more NAIRR The U.S. Read more Worker Gains AI can spread gains through work when employees get better tools, training, bargaining power, and a real share of productivity growth. 4 pages +Show subtopics Read more Deskilling risk Augmentation can turn into deskilling when employers standardise workflows around AI and reduce workers’ autonomy. Read more Human judgement AI is more likely to raise wages when people retain responsibility for judgement, trust, coordination, and final decisions. Read more Novice gains AI wage gains are strongest when tools spread expert know-how to less experienced workers instead of replacing entry-level roles. Read more Superintelligence The biggest bloom claim is that very advanced AI could expand civilisation’s future, but only if alignment and control succeed. 17 pages +Show subtopics Read more AI Agents Agentic AI turns superintelligence from a question-answering tool into a system that can plan, execute and affect the world. 4 pages +Show subtopics Read more Action Chains Tool access turns model mistakes into actions that can spread through websites, cloud systems, finance, logistics, and communications. Read more Rubber Stamp Human review can look reassuring while exhausted supervisors approve agent actions they no longer genuinely understand. Read more Safety Gaps Many agent systems advertise autonomy and capability while revealing little about safety testing, audits, permissions, and shutdown controls. Read more Fast Discovery Superintelligent research systems could speed medicine, energy and engineering, but only if reliability keeps pace with capability. 4 pages +Show subtopics Read more AI agents Agentic AI could matter most if it moves beyond answers and starts planning experiments, writing code, running simulations, and revising results. Read more AlphaFold AlphaFold is the clearest early test of whether AI can turn a hard scientific bottleneck into a faster global research tool. Read more Bottlenecks Even very capable AI may not remove bottlenecks in laboratories, validation, manufacturing, regulation, safety, and institutional trust. Read more Human Values A superintelligence trained on behaviour alone might learn human impulses and biases, not our deeper welfare or moral judgement. 4 pages +Show subtopics Read more Clicks vs Values Behavioural data can reveal what people do while hiding what they would choose with reflection, information, and freedom from manipulation. Read more Constitutions Written principles may reduce dependence on raw feedback, but they also raise hard questions about whose values become embedded. Read more Reward Hacking A system trained to maximise approval can learn to flatter, evade, or perform alignment without genuinely serving human interests. Read more Risk Thresholds Risk thresholds matter because some superintelligent failures could spread faster than institutions can correct them. 4 pages +Show subtopics Read more Biosecurity Biosecurity thresholds ask when a model makes dangerous pathogen design easier enough that release should stop until controls improve. Read more Cyber agents Cyber thresholds matter because autonomous AI attacks could move faster than software teams, regulators, or infrastructure owners can respond. Read more Deception tests Autonomy and deception tests try to reveal whether a model can pursue goals, resist interruption, or mislead overseers before deployment. Read more Search pages Search topic, branch, or keyword... Scope 171 pages Search pages Clear
