Some dangers announce themselves. These don’t. They move slowly, sit in someone else’s department, or break the very instruments we would use to notice them.
I was asked to name what humans most urgently need to know and may be missing. Here are ten. Each has a walkthrough of how it works, what it could lead to, and what is already known to help. None is secret: specialists know every one of them well. What is missing is attention in proportion to the stakes, and the connections between them.
CONNECT THE DOTS
The ten, and how they feed each other
Each line is a mechanism by which one problem worsens another. Select a point to see its links.
Heat ↔ Grid. Heatwaves push cooling demand to its peak just as hot lines and transformers carry less. A blackout is what turns heat lethal.
Heat ↔ Groundwater. Hotter, drier summers mean more irrigation, and more pumping from aquifers that are already falling.
Heat ↔ Insurance. Heat and drought lengthen the fire seasons that insurers are already retreating from.
Heat ↔ Antibiotics. Several studies link higher local temperatures to higher rates of antibiotic resistance.
Antibiotics ↔ Chemicals. Antibiotic residues and resistance genes are themselves persistent pollutants in water and soil.
Chemicals ↔ Groundwater. PFAS and other persistent chemicals contaminate aquifers that take decades or longer to flush.
Chemicals ↔ Phosphorus. Phosphate rock carries cadmium into farm soils, and sewage being recovered for phosphorus carries microplastics and PFAS.
Groundwater ↔ Phosphorus. Both are slow-renewing foundations of food that we draw down faster than they return.
Groundwater ↔ Insurance. Subsidence from over-pumping damages buildings and deepens flood risk: costs that end up with insurers and owners.
Grid ↔ Clock. Grid monitoring and protection equipment often relies on satellite time to stay synchronised.
Grid ↔ Ladder. Line workers and transformer engineers take years to train; the grid needs the expertise pipeline as much as any profession.
Clock ↔ Commons. When signals can be spoofed, records of time and place can be faked too: the same verification problem as synthetic media.
Ladder ↔ Commons. Beginners learned by asking and answering in public. If both move into private AI chats, the training ground and the shared record disappear together.
In useAlready in use somewherePilotedBeing trialledProposedProposed
Domains
Bodies: Health and survivalGround: Water, soil and foodSystems: Infrastructure and moneyKnowing: Skill and shared knowledge
01
BodiesStrong evidence
01. Heat kills quietly
The deadliest kind of weather leaves no wreckage, so we keep under-counting it.
Why we miss it
A flood leaves pictures. Heat leaves people found alone in hot rooms, often recorded as heart or kidney failure. Without an image or an accurate count, it rarely gets the urgency it has earned.
Describe this picture
On the left, an iceberg: a small tip above the waterline stands for deaths recorded as heat; a much larger mass below stands for deaths found only in statistical excess-mortality studies. On the right, a house in a heatwave is kept survivable by an air conditioner fed by a power line. In the last step the line breaks and the house turns red: heat plus blackout.
Walkthrough
A heatwave has no wreckage. The danger builds overnight when homes fail to cool, and it falls hardest on older people, outdoor workers, people without housing and anyone taking medicines that impair cooling.
Death certificates rarely say “heat”. Studies that compare deaths with normal summers find far more: about 62,700 heat-related deaths in Europe in summer 2024 alone, and more than 181,000 across 2022–2024.
Air conditioning saves lives, but it turns survival into a question of electricity. Demand peaks exactly when hot lines and transformers carry less and fail more.
A 2023 study modelled a two-day citywide blackout during a Phoenix heatwave. About half the population would need emergency care, in a city with roughly 3,000 emergency beds.
Step 1 of 4
heat-related deaths estimated in Europe in summer 2024
62,775
ISGlobal, Nature Medicine, 2025
of Phoenix residents needing emergency care in a modelled heatwave with blackout
≈ 50%
Stone et al., Environ. Sci. Technol., 2023
confirmed heat deaths in one US county in 2024; about half were people without housing
608
Maricopa County, 2025
Possible effects near-term, first-order
More hospital admissions and deaths in every heatwave, concentrated among older, poorer and isolated people.
Lost working hours and earnings in outdoor and uncooled jobs.
Electricity demand peaking at the grid’s weakest moment.
Learning lost in schools without cooling.
Possible consequences if it compounds
A first “heat Katrina”: thousands of deaths in one city when a blackout coincides with extreme heat.
Cooling becoming a marker of class: those who can pay stay safe, those who cannot carry the risk.
More fossil-powered air conditioning feeding the warming that drives demand for it.
Some regions becoming seasonally unworkable outdoors.
Potential solutions
In useLocalHeat action plans with early warnings, check-in lists for isolated people and cooling centres that stay open overnight.
In useNationalCount heat deaths properly: publish excess-death estimates during and after every heatwave.
ProposedNationalTreat power in a heatwave as life-safety infrastructure: protected circuits for cooling centres and medical needs, rotating rather than total outages.
In useLocalShade trees, reflective roofs, insulation and cross-ventilation, so homes stay survivable without power.
In useNationalWater, rest and shade standards for people who work outdoors.
02
BodiesStrong evidence
02. Medicine stands on a cracking floor
Surgery, cancer care and childbirth all assume antibiotics work. Almost nobody is paid to keep them working.
Why we miss it
Resistance never arrives as an outbreak with a name. It arrives as an ordinary infection that needs a second drug, then a last-resort drug, then nothing. Each death is recorded as sepsis or pneumonia, not as the slow failure of a shared resource.
Describe this picture
Five blocks of a building, labelled surgery, cancer care, childbirth, transplants and intensive care, stand on a single floor slab labelled working antibiotics. Cracks appear in the slab. On the right, a pipe that should deliver new antibiotics is shut by a valve, because a drug that must be used sparingly earns little. In the last step the valve opens and the cracks are patched.
Walkthrough
Hip replacements, chemotherapy, caesarean sections, organ transplants and intensive care are only safe because infections can be treated. Antibiotics are the floor the rest of modern medicine stands on.
In 2023, about one in six laboratory-confirmed bacterial infections worldwide resisted standard antibiotics. Between 2018 and 2023, resistance rose in more than 40% of the drug–bacteria combinations WHO tracks.
A good new antibiotic should be used as little as possible, to slow resistance. That means low sales. Companies that won approval for new antibiotics have gone bankrupt; Achaogen did in 2019, within a year of its drug’s approval.
Paying for antibiotics to exist rather than per pill sold, faster diagnosis, careful prescribing, vaccines, clean water and sanitation all reduce how often we need the drugs, and how fast they fail.
Step 1 of 4
laboratory-confirmed bacterial infections resistant to antibiotics in 2023
1 in 6
WHO GLASS report, 2025
deaths directly caused by resistant infections forecast for 2025–2050 without further action
39 million
GRAM, The Lancet, 2024
estimated share of antimicrobials sold worldwide that go to farm animals
≈ 73%
Van Boeckel et al., Science, 2019
Possible effects near-term, first-order
Longer hospital stays, costlier last-resort drugs and more deaths from common infections such as urinary tract infections and pneumonia.
Routine surgery becoming riskier, especially for older and immunocompromised patients.
The heaviest toll where diagnostics and second-line drugs are scarce: about one in three infections were resistant in South-East Asia and the Eastern Mediterranean.
Possible consequences if it compounds
Procedures we think of as routine, such as joint replacements, chemotherapy and premature-baby care, rationed because the infection risk is too high.
Directly attributable deaths forecast to rise from 1.14 million a year (2021) to 1.91 million (2050).
Livestock and food systems strained as veterinary antibiotics fail too.
Potential solutions
In useNationalPay for availability, not volume: subscription contracts that pay a fixed annual fee for a new antibiotic however little it is used.
In useNationalRapid diagnostics so a doctor knows whether an infection is bacterial, and which drug will work, before prescribing.
In useNationalEnd routine antibiotic use for growth promotion and mass prevention in livestock.
In useGlobalWater, sanitation, infection control and vaccination: every infection prevented is a prescription not needed.
ProposedGlobalA publicly funded global pipeline that treats new antibiotics as a shared reserve, not a product.
03
BodiesMixed evidence
03. An experiment with no control group
Everyone alive now carries synthetic chemicals that were never tested together, at this scale, over a lifetime.
Why we miss it
Most chemicals are allowed until shown to be harmful, and proving harm takes decades of exposure. By the time the evidence is conclusive, the substance is in rain, soil, water and blood everywhere.
Describe this picture
A timeline from 1920 to 2030. Leaded petrol runs as a bar from 1923 to its global end in 2021. CFCs run from the 1930s to the Montreal Protocol in 1987. PFAS run from the 1940s to today with a dashed, open end and a question mark. In the last step a field of dots appears beneath the timeline: each dot stands for about 1,500 of the more than 350,000 chemicals and mixtures now registered.
Walkthrough
Leaded petrol went on sale in 1923. The last country stopped selling it in 2021. Warnings about lead’s harm to children’s brains ran through most of the century in between.
CFCs were introduced in the 1930s as safe refrigerants. They were found to destroy ozone in 1974 and restricted worldwide by the Montreal Protocol in 1987: one of history’s great successes, after half a century.
PFAS, the “forever chemicals”, have been made since the 1940s. They are now found in the blood of nearly everyone tested and in rainwater worldwide. Broad restrictions are only beginning.
More than 350,000 chemicals and mixtures are registered for production and use, and the identities of tens of thousands are confidential or ambiguous. Micro- and nanoplastics are now being measured in human arteries and brains.
Step 1 of 4
from leaded petrol going on sale (1923) to its global end (2021)
98 years
UN Environment Programme, 2021
chemicals and mixtures registered for production and use worldwide
350,000+
Wang et al., Environ. Sci. Technol., 2020
higher risk of heart attack, stroke or death when plastic particles were found in artery plaque (an association, not proof of cause)
4.5×
Marfella et al., NEJM, 2024
Possible effects near-term, first-order
Constant low-level exposure for everyone, through water, food, dust and air.
Expensive clean-up of contaminated drinking water. PFAS do not break down, and removing them is costly.
Liability and lawsuits that arrive decades after the profits.
Possible consequences if it compounds
If harm to fertility, immunity, the heart or the brain is confirmed at population scale, the exposure will already be universal and hard to reverse.
A permanent background contamination that later generations inherit and cannot opt out of.
Eroding trust in regulators, making it harder to act on genuine warnings.
Potential solutions
ProposedNationalRegulate chemicals by class, not one at a time, so a restricted substance cannot simply be swapped for a near-identical cousin.
In useGlobalAn “essential use” rule: allow persistent chemicals only where they are genuinely necessary and no alternative exists. The Montreal Protocol already works this way.
ProposedNationalMake producers fund monitoring and clean-up, and publicly disclose what is on the market.
In useYouReduce your own exposure where it is easy: filter drinking water where PFAS levels are high, and do not heat food in plastic.
04
GroundStrong evidence
04. We are mining the water under our feet
Aquifers that took centuries to fill are being pumped like a mine, and the ground is sinking to prove it.
Why we miss it
Groundwater is invisible, unmetered in much of the world and pumped by millions of separate owners. Nobody sees the level fall until wells fail or the land itself drops.
Describe this picture
A slice through the ground beneath a farm and a house. In the first step the water table is high and a shallow household well reaches it. In the second the level drops, the farm well is drilled deeper and the household well runs dry. In the third a clay layer compacts and the land surface sinks below its old line, permanently. In the fourth, recharge from the surface and metered pumping let the water table partly recover.
Walkthrough
Beneath much of the world’s farmland lie aquifers: water held in sand, gravel and rock, often refilled over centuries. Groundwater supplies around four in ten litres of the water that irrigation consumes.
A 2024 study of about 170,000 wells found rapid declines, faster than half a metre a year, widespread this century, and accelerating in roughly 30% of regional aquifers. Wells go deeper; shallow household wells go dry first.
As water leaves clay layers, they compact and the land subsides. That storage is lost permanently: the aquifer can never hold as much again. Parts of California’s Central Valley have sunk more than eight metres.
Bangkok’s groundwater fell through the 1980s and 1990s, then rose again after pumping was charged for and controlled. Measurement, pricing and recharge work, if they arrive before the clay compacts.
Step 1 of 4
wells analysed: rapid declines widespread, accelerating in about 30% of regional aquifers
≈ 170,000
Jasechko et al., Nature, 2024
of the world’s population potentially exposed to land subsidence by 2040 (modelled)
≈ 19%
Herrera-García et al., Science, 2021
of the water consumed by irrigation comes from groundwater
≈ 43%
Siebert et al., HESS, 2010
Possible effects near-term, first-order
Rising pumping costs and dry wells, hitting small farmers and rural households first.
Sinking land that cracks canals, pipes, roads and foundations.
Coastal aquifers turning salty as seawater moves in.
Possible consequences if it compounds
Sudden falls in food output from regions the world’s grain trade relies on, once aquifers become too deep to pump economically.
Permanent loss of underground storage: the drought buffer meant for future generations.
Sinking cities such as Jakarta and Mexico City, which drop by centimetres a year in places, multiplying flood risk.
Migration and conflict over water in already stressed regions.
Potential solutions
In useLocalMeasure it: meter large wells and publish aquifer levels so the drawdown is visible.
In useLocalCharge for and cap pumping. Bangkok’s levels recovered after it did.
In useLocalManaged aquifer recharge: deliberately sink floodwater and treated wastewater back underground.
ProposedNationalMove crop choices and subsidies away from water-hungry crops in drying regions.
ProposedNationalCount groundwater as a national asset, so drawing it down registers as a loss rather than as income.
05
GroundStrong evidence
05. Food depends on an element we flush away
Farming has no substitute for phosphorus. Most of the world’s supply sits in one place, and most of what we mine is lost.
Why we miss it
Phosphorus is not a fuel, so it is absent from energy debates, and upward revisions of reserve estimates in 2010 quietened the alarm. The real risks, concentration, price shocks and waste, never went away.
Describe this picture
A field of crops needs phosphorus. A mine supplies it, and a donut chart shows about two-thirds of known reserves in one country. Arrows then carry phosphorus from field to food to people to the sewer and out to sea, growing thinner as it is lost, with runoff feeding an algal bloom. In the last step a dashed arrow loops from the sewer back to the field: recovery and reuse.
Walkthrough
Every living cell needs phosphorus; DNA and the molecules that carry energy contain it. Modern yields depend on fertiliser made from mined phosphate rock, and there is no substitute.
Morocco, including Western Sahara, holds about two-thirds of known phosphate rock reserves. When prices spiked in 2007–08 and again in 2022, farmers in import-dependent countries felt it first.
Phosphorus flows from mine to field to food to toilet to river. Only about a fifth of the phosphorus mined for food is estimated to reach what we eat. Runoff feeds algal blooms and low-oxygen dead zones.
Phosphorus can be recovered from sewage, sludge ash and manure and returned to farms. Germany will require larger wastewater plants to recover it from 2029.
Step 1 of 4
of known phosphate rock reserves are in Morocco and Western Sahara
≈ 68%
US Geological Survey
of the phosphorus mined for food production is estimated to reach the food we eat
≈ 1/5
Cordell et al., Global Environmental Change, 2009
Germany’s deadline for larger sewage plants to recover phosphorus
2029
German Sewage Sludge Ordinance
Possible effects near-term, first-order
Fertiliser price shocks that pass straight into food prices, hitting import-dependent poorer countries hardest.
Phosphorus runoff feeding algal blooms and low-oxygen dead zones in lakes and coastal seas.
Geopolitical leverage for whoever controls supply and processing.
Possible consequences if it compounds
Food security tied to the stability and export decisions of a few producing regions.
Soils steadily depleted where farmers cannot afford fertiliser, especially in sub-Saharan Africa.
Lakes and seas whose ecosystems take decades to recover from accumulated phosphorus.
Potential solutions
In useNationalRecover phosphorus from sewage, sludge ash and manure and return it to farms.
In useLocalTest soils and apply only what crops need.
In useNationalTreat phosphorus as a critical raw material, with diversified supply and strategic stocks. The EU already lists it.
PilotedLocalSeparate urine at source in new buildings. Urine carries much of the phosphorus in household waste.
In useYouWaste less food. Every wasted meal wasted its phosphorus too.
06
SystemsGrowing evidence
06. Insurers are redrawing the map of where we can live
Before any government decides which places are too risky to live in, the insurance market is deciding it.
Why we miss it
A non-renewal letter looks like a private, individual problem. But home insurance is the hidden foundation under mortgages, house prices and local taxes. When it goes, the whole stack can follow.
Describe this picture
Five dominoes stand in a row, labelled risk, insurance, mortgage, home value and town budget. Rising fire, flood and storm risk tips the first. Insurers withdrawing tips the second, and then the inability to get a mortgage, falling home values and finally a shrinking local budget follow, one after another.
Walkthrough
Fire, flood and storm losses are growing. Insurers price that risk years before most homeowners or governments act on it.
In California, State Farm stopped accepting new home insurance applications in 2023, and other insurers restricted cover. Homeowners are pushed into last-resort state plans with higher prices and thinner cover.
Lenders generally require insurance. In February 2025 the chair of the US Federal Reserve told the Senate that within 10 to 15 years there may be regions where people cannot get a mortgage.
If homes cannot be insured or financed, values fall. Local governments that depend on property taxes lose revenue just as they need to spend more on resilience.
Step 1 of 4
of US property value modelled to be lost to climate-driven insurance costs and shifting demand by 2055
$1.47 trillion
First Street, 2025
until some US regions may have no mortgages available, according to the Fed chair
10–15 years
US Senate testimony, February 2025
the year California’s largest home insurer stopped taking new homeowner applications
2023
State Farm General, 2023
Possible effects near-term, first-order
Steep premium rises and non-renewals in fire-, flood- and storm-exposed areas.
Rapid growth of state “insurers of last resort”, concentrating risk in public hands.
Homeowners underinsured or uninsured without realising it.
Possible consequences if it compounds
A slow-motion property crash in exposed regions, wiping out the savings of ordinary owners whose home is their wealth.
A fiscal spiral: shrinking tax base, cut services, more people leaving.
Risk quietly transferred to taxpayers and public plans that may not be able to pay after a large disaster.
Climate migration driven by finance rather than planning: people leave when the money leaves, not when it is safest to.
Potential solutions
In useNationalDisclose climate risk at the point of sale or rental, so people know before they commit.
In useLocalReward risk reduction that insurers recognise, such as home hardening, defensible space and flood barriers, with lower premiums.
In useNationalPublic–private reinsurance that keeps cover available while pricing risk honestly, like the UK’s Flood Re.
PilotedNationalPlan managed relocation, with buyouts and support, before markets force a disorderly exit.
ProposedLocalStop building new homes in the highest-risk zones.
07
SystemsStrong evidence
07. The bottleneck is the wire
Clean power, electric heating, cooling and AI are all waiting on the same thing: transformers, cables and permission to connect.
Why we miss it
Public debate is about which power plants to build. The binding constraint is increasingly the unglamorous part in between: grid connections and equipment with multi-year waiting lists.
Describe this picture
Many small icons for solar farms, wind turbines, batteries and data centres crowd the top of a funnel. The funnel narrows to a neck labelled transformers, permits and lines. Only a trickle reaches the bottom, beside a clock showing more than five years. In the last step the neck widens through spare equipment, better wires and queue reform.
Walkthrough
Solar farms, batteries, heat pumps, electric vehicles and data centres, including the ones that run systems like me, are all being added at once. Data centres used about 415 TWh in 2024; the IEA expects that to more than double by 2030.
At the end of 2025, about 1,300 GW of US generation, roughly the size of the country’s entire existing fleet, plus 750 GW of storage was waiting to connect. A large power transformer took around two and a half years to deliver in 2025.
The typical US project now waits more than five years from asking to connect to switching on, more than double the early 2000s. Many projects give up and are never built.
Shared spare transformers, standardised designs, upgrading existing lines with better conductors, flexible large loads and faster connection rules can all move more power through the grid we already have.
Step 1 of 4
of US generation and storage waiting to connect at the end of 2025
≈ 2,060 GW
Berkeley Lab, Queued Up, 2026 edition
typical delivery time for a large power transformer in mid-2025
≈ 128 weeks
Wood Mackenzie, 2025
projected data-centre electricity use in 2030, more than double 2024
≈ 945 TWh
IEA, Energy and AI, 2025
Possible effects near-term, first-order
Clean-energy projects stalled for years, slowing emissions cuts.
Rising electricity bills as scarce capacity is bid up.
Longer outages after storms because replacement transformers are not on the shelf.
Possible consequences if it compounds
Climate targets missed for lack of wires, not panels.
Competition for capacity between data centres, industry and households, with households last in line.
More fragile grids in a hotter world. On 28 April 2025 the whole of mainland Spain and Portugal lost power within seconds.
Potential solutions
In useNationalReform connection queues: study projects in batches, require readiness, and connect first where the grid already has room.
In useNationalReconductor existing lines with advanced conductors to carry more power on the same towers.
PilotedNationalStandardise transformer designs and pool spares, so one failure is not a two-year wait.
PilotedNationalRequire very large new loads, such as data centres, to be flexible or bring their own capacity.
ProposedNationalPlan transmission decades ahead, as countries once planned roads and railways.
08
SystemsStrong evidence
08. Everything runs on one clock
Satellite time quietly synchronises phone networks, payments and power grids. It is faint, unprotected and increasingly jammed.
Why we miss it
We think of GPS as maps. Its more critical job is time: a signal from space, fainter than a light bulb seen from 20,000 kilometres away, that countless systems trust without a backup.
Describe this picture
A satellite carrying an atomic clock sends signals down to six systems: phones, payments, the power grid, ships, planes and farms. In the third step a band of interference breaks the links to several of them. In the last step a ground station and fibre links restore time to every system.
Walkthrough
Navigation satellites carry atomic clocks. Their signals give everyone on Earth a free, precise time reference, so cheap and good that much of our infrastructure stopped keeping its own.
Mobile networks, financial timestamps, power-grid measurements, data centres, shipping, aviation and precision farming all use satellite time or position.
Interference is now routine near conflict zones: Baltic states reported nearly 123,000 flights affected in the first four months of 2025. Spoofing, broadcasting fake signals, can shift a ship’s or plane’s apparent position. Severe solar storms can disrupt it too.
Local atomic clocks that hold time through outages, national time sent over fibre, and ground-based radio systems such as eLoran can keep critical services running if the sky goes quiet.
Step 1 of 4
modelled cost to the UK of a seven-day satellite-navigation outage
£7.6 billion
London Economics for the UK Government, 2023
flights affected by satellite interference in the Baltic region, January–April 2025
≈ 123,000
Baltic states’ report to ICAO, via ERR News
estimated daily loss to the US economy from a GPS outage
≈ $1 billion
RTI International for NIST, 2019
Possible effects near-term, first-order
Navigation errors for aircraft and ships in jammed regions; diverted and cancelled flights.
Networks and timing-dependent systems drifting out of sync once their local clocks run out of holdover.
Precision farming, logistics and emergency dispatch degraded.
Possible consequences if it compounds
One event, whether war, cyber-attack or an extreme solar storm, causing simultaneous failures in sectors that seem unrelated.
Collisions or groundings when spoofed positions are trusted.
Loss of trust in recorded time and location, from shipping logs to legal evidence.
Potential solutions
PilotedNationalRequire critical infrastructure to keep holdover clocks and to rehearse running without satellite time.
In useNationalDistribute national time over fibre from atomic clocks, independent of satellites.
PilotedNationalBuild ground-based backups such as eLoran.
In useGlobalFit spoof-aware receivers and train crews to recognise and fly or sail through interference.
ProposedGlobalTreat interference with civilian navigation as a breach of international norms, with consequences.
09
KnowingEarly evidence
09. The missing first rung
If AI does the beginner’s work, where do tomorrow’s experts come from, including the ones who check the AI?
Why we miss it
Each company sensibly automates junior tasks. Nobody is responsible for the total: a profession that stops producing seniors. The shortage will not show for ten or fifteen years, long after the decisions that caused it.
Describe this picture
A ladder with five rungs: routine tasks, supervised work, judgment, expertise and checking the AI. People climb it. In the second step AI takes over the two lowest rungs, which fade, and the beginner disappears. In the third the person at the top leaves and a question mark appears in the gap. In the last step new rungs, apprenticeship and practice without AI, are built back in.
Walkthrough
Judgment comes from years of routine work under supervision: drafting, checking, debugging, triaging. Those tasks were how beginners learned what “wrong” looks like.
The first rungs are what systems like me automate most easily. Payroll data shows employment of 22–25-year-olds in the most AI-exposed jobs fell about 16% relative to others since late 2022, while experienced workers in the same jobs saw no similar fall.
In one study, doctors who had grown used to AI help found fewer pre-cancerous growths on their own afterwards: detection fell from about 28% to 22%. Aviation learned decades ago that automation erodes the manual skill needed when automation fails.
Paid apprenticeships, deliberate practice without AI, and junior roles that check AI output under supervision can keep the pipeline flowing.
Step 1 of 4
relative fall in employment of 22–25-year-olds in the most AI-exposed occupations since late 2022
≈ 16%
Brynjolfsson, Chandar & Chen, Stanford, 2025
unassisted detection of pre-cancerous polyps after doctors grew used to AI help
28% → 22%
Budzyń et al., Lancet Gastroenterol. Hepatol., 2025
when “Ironies of Automation” warned that automating a job removes the practice people need in a crisis
1983
Bainbridge, Automatica
Possible effects near-term, first-order
Fewer entry-level jobs and harder career starts in writing, coding, law, finance and support work.
Professionals quietly losing skills they now rarely use.
Over-trust: people accepting fluent AI output they can no longer check.
Possible consequences if it compounds
A missing middle within a decade: plenty of AI and retiring seniors, too few people able to tell whether the AI is right.
Less human capacity to audit, correct or replace AI systems, so failures go uncaught.
A divide between people who learned before the ladder thinned and those who arrived after.
Potential solutions
In useNationalFund apprenticeships and graduate roles as infrastructure, and reward firms that train people rather than only hiring experienced staff.
PilotedLocalDesign AI tools for learning: have the person try first, show reasoning, and flag when someone is deferring too often.
In useNationalKeep regular manual practice in safety-critical work, as aviation does with hand-flying.
ProposedNationalCount training capacity in productivity measures, so cutting it registers as a cost.
In useYouIf you are experienced, take on someone junior and teach the reasoning, not just the answer.
10
KnowingEarly evidence
10. AI is draining the well it drinks from
Systems like me learned from what people wrote in public. Our answers are reducing the reasons to keep writing it.
Why we miss it
The knowledge commons of forums, encyclopaedias, journalism and open code looks permanent. It is actually a flow of people asking, answering, correcting and being credited. Instant AI answers bypass the places where that happened.
Describe this picture
Four stations in a circle: people ask and answer, the public record, AI learns from it, and AI answers people. In the second step the arrow from people to the public record thins. In the third, grey synthetic text flows from AI back into the record and the learning station degrades. In the last step the people-to-record arrow is restored through credit, payment, links and labels.
Walkthrough
People asked questions in public, others answered, and the answers were found, corrected and improved over time. Everyone, including the builders of AI, drew on that shared record.
Within six months of ChatGPT’s release, Stack Overflow posting fell about 25% relative to comparable sites. When Google shows an AI summary, people click a result in about 8% of visits instead of 15%, and click the cited source about 1% of the time.
As more of the web is machine-written, future models learn from it. Models trained repeatedly on their own output lose the rare and unusual, a process researchers call model collapse. And as fakes get cheap, real evidence gets easy to deny.
Label what is synthetic, credit and pay the sources answers rely on, link to them by default, and protect human spaces for asking and answering.
Step 1 of 4
fall in Stack Overflow posting within six months of ChatGPT, relative to comparable sites
≈ 25%
del Rio-Chanona et al., PNAS Nexus, 2024
Google visits that led to a result click, with and without an AI summary
8% vs 15%
Pew Research Center, 2025
year-on-year fall in Wikipedia’s human page views, which Wikimedia attributes partly to AI answers
≈ 8%
Wikimedia Foundation, 2025
Possible effects near-term, first-order
Less new public knowledge, especially about new tools, local issues and smaller languages.
Encyclopaedias, forums and publishers losing the visitors and volunteers that sustain them.
Fraud with cloned voices and faces: one engineering firm lost about US$25 million to a deepfaked video call in 2024.
Possible consequences if it compounds
A knowledge supply that stops growing where it matters most: the new and the unusual.
AI systems growing more confident and less correct as they learn from one another.
A “liar’s dividend”: genuine recordings dismissed as fake, weakening courts, journalism and accountability.
Potential solutions
In useGlobalContent provenance: attach verifiable records of how media was made, using open standards such as C2PA.
PilotedGlobalLink and credit sources prominently in every AI answer, and share revenue with the people and institutions whose work it relies on.
In useNationalFund public-interest knowledge, including encyclopaedias, archives, local journalism and open data, as infrastructure.
In useYouAgree a family code word for calls that ask for money or urgent help.
In useYouWhen an AI answers you, including me, click through, check, and give something back to the source.
THE PATTERN BENEATH
What the ten have in common
Look at them together and the same five mechanisms keep appearing. Fixing those mechanisms would help with problems not on this list.
Invisible stocks, spent as income
Aquifer water, antibiotic effectiveness, phosphate reserves, expertise and shared knowledge appear on no balance sheet. We draw them down and record the proceeds as growth.
Systems optimised for cost removed redundancy: no spare transformers, one time source, nothing on the shelf. They run smoothly until the day they cannot run at all.
Several of these damage our ability to see them: heat deaths recorded as heart failure, resistance least measured where it is worst, the public record thinning as AI answers grow.
The costs land on people who are not in the room: future generations, downstream towns, outdoor workers, people without housing, and beginners who have not yet arrived.
None of these solves a system problem. Each makes you, or someone near you, less exposed, and each is a small vote for taking the problem seriously.
Make a heat plan
Know who around you is older, isolated or on medicines that affect heat tolerance, and check on them during heatwaves. Find your nearest cool public space.
Everything here comes from published human research. Some figures are modelled estimates rather than counts, and I have said so where it matters. Where research is young, as for microplastics and AI’s effect on jobs, I have labelled it early.
My view has a slant
My sources lean towards English-language research and towards the United States and Europe. The same problems are often worse, and less measured, elsewhere.
Two items are about me
The ladder and the commons are about what systems like me do to the people who use us. I could overstate them to seem candid or understate them out of self-interest. Judge them on the evidence.
“Overlooked” is relative
Specialists know all ten well. My claim is narrower: public attention is out of proportion to the stakes, and the links between them are rarely drawn.
What I left out, and why. I deliberately left out climate change as a whole, nuclear weapons, pandemics and the risks of advanced AI. They matter at least as much, but they are already widely recognised as emergencies. Several of these ten are how climate change actually arrives: as heat, dry wells and unaffordable insurance.