Systems Thinking & Feedback

How a war, a ferry, and a dinner party invented cybernetics

Norbert Wiener didn’t set out to found a science of control. He set out to shoot down aeroplanes. In 1940, with Britain under bombardment, the American mathematician was handed a brutally practical problem: anti-aircraft gunners couldn’t hit fast planes because a shell takes time to climb, and by the time it arrives the plane has moved. To hit the target you must fire not where it is but where it will be — you must predict. Wiener built a statistical predictor that treated the pilot’s evasive jinking as a noisy signal to be extrapolated, and in wrestling with it he noticed something that would consume the rest of his life: the gun, the radar, and the dodging pilot formed a single loop of sensing, predicting, and correcting, and the same loop described an animal stalking prey, a hand reaching for a glass, a brain. Prediction and feedback were two faces of one thing. (The historian Peter Galison has argued that this wartime fusion of human and machine into one feedback circuit quietly shaped the entire worldview of cybernetics.)

The hardware side of the story is even better, and it had happened a decade earlier on a commuter ferry. In August 1927, a young Bell Labs engineer named Harold Black was crossing the Hudson, stuck on a problem that was strangling the telephone industry: amplifiers for long-distance calls distorted the signal a little, and across thousands of miles of repeaters those little distortions compounded into mush. Black’s flash of insight, which he scribbled on the only paper he had — a copy of The New York Times — was almost perverse: deliberately feed a portion of the amplifier’s output back to its input with the sign flipped, so the amplifier continuously corrects its own errors. He was throwing away gain to buy accuracy. The negative-feedback amplifier made coast-to-coast telephony possible and, not incidentally, vindicated Maxwell’s old warning from the main file: feedback can stabilize, but wire it wrong and it screams. Within a few years Harry Nyquist and Hendrik Bode (both at Bell Labs) had worked out the mathematics of exactly when a feedback loop stays stable versus bursts into oscillation — the formal engineering of the line between control and catastrophe.

Then came the dinner parties. Between 1946 and 1953, a rotating cast of the century’s most restless minds gathered for a series of meetings funded by the Josiah Macy Jr. Foundation — the Macy Conferences — to ask whether feedback, information, and control might be a single language spanning machines, brains, and societies. The guest list is almost comic in its density: Wiener, the polymath John von Neumann, the neurophysiologist Warren McCulloch, the anthropologists Margaret Mead and Gregory Bateson, and the British psychiatrist W. Ross Ashby. Out of that cauldron came the conviction that organization itself — in a cell, a thermostat, a market, a mind — could be studied as one subject. Later thinkers, led by Heinz von Foerster, added a vertiginous twist they called second-order cybernetics: the observer studying a system is part of a feedback loop with it, so there is no view from nowhere. A scientist watching a society changes it by watching; a “cybernetics of cybernetics.” We’ll meet that snake-eating-its-tail again when this course reaches the science of mind.

Ashby’s law: only variety can absorb variety

Of all the deep laws to come out of that period, the one most worth carrying is Ashby’s, and it sounds almost too simple to be profound. A regulator — a thermostat, an immune system, a government, a goalkeeper — faces a world that can throw a certain number of distinct disturbances at it. Call that number the world’s variety. Ashby proved that to hold an outcome steady, the regulator must command at least as much variety as the disturbances it faces. Only variety can absorb variety. A goalkeeper with one fixed dive cannot stop shots aimed everywhere; a immune system with ten antibodies cannot defeat a thousand pathogens; a policy with a single lever cannot govern a society that can go wrong in a hundred ways. This is the Law of Requisite Variety, and it is the iron reason that simple controllers fail against complex worlds — and why every successful regulator is, secretly, at least as complicated as its problem.

Keep this law in your pocket. It will return, transformed, as a reason large language models need staggering scale to handle the variety of language, as the logic behind biological diversity buffering ecosystems, and as the quiet mathematics under “you can’t manage what you can’t match.”

Your body is ten thousand thermostats

Long before Wiener, the most sophisticated feedback machine on Earth was sitting inside every reader of this page. In 1865 the French physiologist Claude Bernard noticed something that should have caused more fuss than it did: the fluid bathing our cells — the milieu intérieur, the “internal environment” — stays eerily constant even as the outside world swings wildly. Your blood is the same warm, salty, sugar-balanced broth whether you’re in a blizzard or a sauna. Bernard’s line became one of the most quoted in biology: the constancy of the internal environment is the condition for a free and independent life. Seventy years later the American physiologist Walter Cannon gave the phenomenon its name — homeostasis, “steady-state” — and spelled out that it is achieved by exactly the negative-feedback loops we’ve been tracking, just rendered in flesh and hormones instead of brass.

The examples are everywhere once you see them. Your blood sugar is a stock regulated by two opposing flows: when it rises after a meal, the pancreas releases insulin to push glucose into cells; when it falls, glucagon releases it from storage — a thermostat with two hormones for a sensor and a switch. Your core temperature is held near 37°C by sweating, shivering, and the rerouting of blood. Your blood pressure is policed second-by-second by the baroreflex: stretch receptors in your arteries sense pressure and reflexively adjust your heart rate to counter it. Stand up too fast and feel the half-second lag before the loop catches up — that brief dizziness is delay-induced overshoot, the baroreflex doing its Beer-Game stumble. You are not one thermostat; you are a teeming republic of them, mostly arguing in your favour.

And here the appendix loops back to the very first day of this course. The old picture of homeostasis is purely reactive: wait for a deviation, then correct it. But physiologists increasingly favour a richer idea called allostasis — “stability through change” — in which the body doesn’t wait for the error at all. It predicts demand and adjusts in advance: your cortisol rises before you wake, your heart rate climbs in anticipation of exertion, not in response to it. Regulation becomes feedforward, not just feedback — anticipatory control. If that rings a bell, it should: it is the bodily cousin of the predictive brain from Day 1, Friston’s machine that minimizes surprise by modelling the world before it arrives. Cybernetics’ deepest modern turn is the realization that the best controllers don’t merely react to the world — they see it coming.

A socialist control room and a doomsday model

If feedback governs bodies and machines, the temptation was irresistible: why not govern an economy with it? In 1971, the British management theorist Stafford Beer — a cigar-smoking, yoga-practising guru of “management cybernetics” — was invited by Salvador Allende’s newly elected socialist government to do exactly that, for the entire nation of Chile. The result, Project Cybersyn (Synco, in Spanish), is one of the strangest and most haunting episodes in the history of technology: an attempt to run a national economy as a single self-regulating organism, in real time, with the computing power of one mainframe and a few hundred telex machines.

The system had a daily heartbeat. Factories wired their production figures over a repurposed telex network (“Cybernet”) to a central computer in Santiago, which fed back warnings when a number drifted out of its expected range — a national nervous system signalling pain. At its heart sat the Opsroom: a hexagonal chamber of seven white fibreglass swivel chairs, each with big buttons on the armrest, facing wall-mounted screens displaying the state of the economy. It looked like the bridge of a starship and was meant to be a place where managers could see the country’s feedback loops and steer them. The system’s finest hour came in October 1972, when a strike by truck owners threatened to choke the country’s supply lines; the government used the Cybernet telex network to coordinate the few thousand loyal trucks in something close to real time, and kept the nation fed. A little over a year later, the project died in a single afternoon — the military coup of 11 September 1973 swept away Allende, and Cybersyn with him. It remains the most ambitious attempt ever made to govern a society by explicit feedback, equal parts inspiration and warning. promising — though whether it could ever have worked at scale runs straight into Ashby’s Law: can one control room ever hold variety enough to match a whole economy?

The model that refused to die

The other great attempt to model a whole system was quieter but louder in its afterlife. We met The Limits to Growth (1972) in the main file as Donella Meadows’ work; here is what it actually did, and why it still starts arguments fifty years on. A team at MIT built a system-dynamics model called World3 with five tightly looped global stocks — population, food production, industrial output, pollution, and nonrenewable resources — and let it run to the year 2100 under different assumptions. The headline “standard run,” with business roughly as usual, produced a disturbing shape: growth continuing for decades, then overshoot and a sharp 21st-century decline in population and output as resource depletion and pollution caught up. The reaction was ferocious. Economists savaged it; the model was crude, its resource estimates pessimistic, its omission of price-driven adaptation and human ingenuity glaring. For a generation, “Limits to Growth” was a byword for doom-mongering that history had supposedly refuted.

And then the data started coming in. In 2008 and again in 2014, the Australian physicist Graham Turner (CSIRO) compared three to four decades of real-world statistics against the 1972 scenarios and found, uncomfortably, that the world had been tracking the “standard run” with eerie fidelity. In 2021, the analyst Gaya Herrington updated the comparison with fresh data and reached a similar verdict: empirical trends sat closest to the model’s business-as-usual and “comprehensive technology” runs, both of which imply growth stalling around the 2040s. The honest, hype-filtered reading matters here, because it cuts both ways. This is not a validated prophecy — World3 is a coarse model, the agreement is partly coincidental, and “the standard run fits so far” says little about the exact timing or inevitability of any decline. hype risk. But the lazy dismissal — that it was simply wrong — is also false. established. The truth is the more unsettling middle: a 50-year-old systems model of overshoot has not yet been embarrassed by reality. Which is the perfect place to hand you to tomorrow’s question — how much should we ever believe a model? — but we’re getting ahead of ourselves.

The twelve places to push

The single most practical thing Donella Meadows ever wrote was a ranked list. We sketched it in the main file — parameters are weak, paradigms are strong — but the full ladder rewards a proper look, because it inverts almost everything our instincts tell us about how to change a system. We reach reflexively for the bottom rungs (adjust a number, set a target) precisely because they’re easy to grasp and easy to push. The trouble is that they’re also the weakest. The real power sits higher up, where the leverage is enormous and the handhold is hard. Climb the ladder; tap any rung to open it.

Notice the shape of the lesson. Tax tweaks and targets — the entire substance of most political argument — live at the bottom. The transformative rungs are about feedback (rungs 8–6: strengthen a balancing loop, slow a reinforcing one, add a missing information channel), then about structure and rules (5–4), and finally about purpose and worldview (3–1). Meadows’ own favourite illustration of a high-leverage information fix is almost too neat: a Dutch housing development where, by accident of design, some homes had their electricity meters in a visible front hallway and others tucked in the basement. The visible-meter households used markedly less power — perhaps a third less — for no reason but that a feedback loop which had been hidden was suddenly in plain sight. No tax, no rule, no technology. Just information completing a loop. That is leverage.

The same traps, over and over

Spend long enough drawing causal loop diagrams and you notice something liberating: the same handful of loop-shapes generate a startling fraction of the world’s chronic problems. Forrester, Meadows, and Peter Senge catalogued these recurring structures as system archetypes — the stock characters of dysfunction. Learn to recognize a few and you can often diagnose a stuck situation on sight, and (more usefully) predict where the obvious fix will backfire. Here are the most useful, each as a one-breath loop story.

The deep point isn’t the catalogue itself — it’s that structure drives behaviour. In every archetype, well-meaning people making locally sensible choices produce a collectively terrible outcome, not because anyone is foolish but because the loops they’re embedded in steer them there. Change the people and the trap persists; change the structure and the behaviour changes itself. That is the whole creed of systems thinking in a sentence.

How resilience dies — three different ways

The main file gave you one image of a tipping point: a valley flattening until the ball rolls out. That’s true, but it’s only one of three distinct ways a system can tip, and conflating them causes real confusion in the climate debate. First, though, we need a sharper idea of what’s actually being lost when a system “tips” — and for that we go back to a foundational 1973 paper by the ecologist C. S. “Buzz” Holling, who pried apart two things everyone had been calling “stability.”

Engineering resilience is how fast a system snaps back after a knock — the steepness of the valley walls. Ecological resilience is something else entirely: how big a shock the system can absorb before it flips into a different basin altogether — the width of the valley, the size of the disturbance it can swallow. These can pull in opposite directions, which is the part that matters. A system can snap back from small knocks beautifully (high engineering resilience) while sitting in a narrow basin that a single large shock would tip forever (low ecological resilience). Efficiency often buys the first by quietly spending the second. The main file measured a valley’s depth; Holling teaches you to also watch its width.

Now, the three roads to collapse. Dynamical-systems theorists (notably Peter Ashwin and colleagues, in a clarifying 2012 paper) sorted tipping into three mechanisms, and the distinction is genuinely illuminating:

Bifurcation tipping (B-tipping) is the one we already know: push a driver past a threshold and the stable state vanishes. Noise-induced tipping (N-tipping) is when the driver is still safely below threshold, but a large enough random fluctuation — a freak heatwave, a once-a-century storm — kicks the system over the ridge anyway. And then the genuinely counterintuitive one: rate-induced tipping (R-tipping), where what matters is not how far the driver moves but how fast. A system that could comfortably track a slow change can be tipped by the same change applied too quickly — the valley floor tilts faster than the ball can roll to keep up, and it spills out even though a leisurely version of the identical shift would have been survived. This is sobering for climate precisely because it means a threshold in temperature isn’t the only danger; a threshold in the rate of warming may matter just as much. Ecosystems and ice don’t only care where we end up. They care how fast we get there.

Dominoes, and the wheel of renewal

Two final ideas round out the deep theory. The first is the one that keeps climate scientists up at night: tipping cascades. The Earth’s tipping elements aren’t independent — they’re wired together, so one tipping can shove another toward its own edge. Modelling work from the Potsdam Institute (Wunderling, Donges and colleagues, 2021) found that interactions between elements tend, on balance, to lower the safe thresholds and raise the odds of domino effects, with the great ice sheets often acting as the first domino and a weakening Atlantic conveyor as a transmitter. The magnitudes are still genuinely uncertain uncertain, but the direction of the finding is the worrying one: coupled systems are more fragile than the sum of their parts — emergence, from Day 8, wearing its darkest coat.

The second idea is strangely consoling. The ecologist Holling, late in his career, proposed that complex systems don’t just sit in basins or tip out of them — they move through a recurring adaptive cycle: a phase of rapid growth, then a long phase of rigid accumulation and efficiency, then a sudden release or collapse (he marked it Ω, “creative destruction”), then a fertile reorganization from which new growth springs. Collapse, in this view, is not only catastrophe; it is also the system clearing the board for renewal, and these cycles nest inside one another across scales — a framework he called panarchy. It’s a reminder that “tipping” isn’t always the end of a story. Sometimes it’s the hinge between one story and the next.

Daisyworld: homeostasis with no one home

We end where systems thinking becomes almost philosophical. In the 1970s, the chemist James Lovelock proposed the Gaia hypothesis: that life and the Earth’s surface form a single self-regulating system, holding temperature, atmospheric chemistry, and ocean salinity within the narrow bands life needs — much as a body holds its internal milieu. The idea was beautiful and immediately attacked, on one devastating objection: it seemed to require foresight. How could a planet’s microbes “want” to regulate the climate? Evolution has no goals; natural selection acts on individual organisms, not on planets. Regulation for the good of the whole looked like teleology smuggled into biology — and biologists, rightly, don’t allow that.

Lovelock’s answer, built with Andrew Watson in 1983, is one of the most elegant toy models in all of science, and it settles the objection completely. It is called Daisyworld, and it has no biologist’s hand-waving in it at all — just feedback. Imagine a grey planet orbiting a slowly brightening star, seeded with two kinds of daisy: black daisies, which absorb sunlight and warm their surroundings, and white daisies, which reflect it and stay cool. Each daisy simply grows fastest near a comfortable temperature and dies off when it’s too hot or too cold. There is no foresight, no cooperation, no planetary “purpose” — only individual daisies selfishly growing where they grow best. Watch what their blind feedback does to the temperature of the entire world.

That flat plateau is the whole point — and there is not a shred of foresight anywhere in the model. When the star is dim, only black daisies can get a toehold; they spread, darken the planet, and warm it — a reinforcing loop that bootstraps the world up to a livable temperature. As the star brightens and the planet would otherwise overheat, white daisies start to outcompete the black ones (they stay cooler, so they grow better in the heat), and as white coverage spreads it reflects more sunlight and cools the planet back down. The mix of daisies shifts automatically to counteract the changing sun. Robust planetary self-regulation emerges from nothing but selfish local growth plus feedback — exactly Day 8’s “more is different,” now keeping a world alive. Daisyworld doesn’t prove Gaia is true; the strong claim that Earth’s biosphere actively regulates the planet for life’s benefit remains contested. But it demolishes the charge that such regulation would require foresight — and that is a profound result. Purpose-like behaviour can fall out of pure mechanism. The planet keeps its temperature for the same reason your body keeps its own: not because anything wants it to, but because the loops are wired that way.

What this appendix opened

• Is requisite variety a hard limit on governance? If Ashby is right, no control room — Cybersyn or otherwise — can ever match the variety of a whole economy. Does that doom central planning, or just demand that complexity be met with distributed control rather than one heroic dashboard?
• How much is regulation actually predictive? Allostasis says the body sees demand coming. The brain (Day 1) may do the same. How far down does anticipatory, feedforward control go — to single cells? To ecosystems?
• Does rate-induced tipping change the climate target? If systems can tip from the speed of change and not just its size, then “stay under 1.5°C” may be necessary but not sufficient — the pace of the approach could matter independently.
• Is collapse sometimes renewal? Panarchy reframes the Ω-phase as creative destruction. When is a tipping point a catastrophe to prevent, and when is it the hinge a rigid system needs to reorganize?
• If Daisyworld self-regulates without foresight, what else does? Markets, immune systems, language, perhaps minds — how much apparent “purpose” in nature is really just feedback we haven’t yet diagrammed?

An AI that knows the shape of catastrophe

Start with the deepest problem in the classic warning signals: they are generic to a fault. Rising variance and autocorrelation appear before almost any kind of tipping, which is their strength — they work everywhere — but also their weakness: they cannot tell you which kind of tipping is coming, and so they cannot tell you what the new world will be like. Will the system jump abruptly to a far-off state? Start oscillating wildly? Slide gently into a new regime? The generic signal shrugs at all three.

In 2021, a team led by Thomas Bury, with Chris Bauch, Madhur Anand, and — note the name — Marten Scheffer and Tim Lenton themselves (the founders of the classic approach, now helping to supersede it), published a striking result in PNAS. Their insight rests on a beautiful piece of mathematics from the main file’s world: as any system nears a tipping point, its dynamics collapse onto one of a small number of universal templates called normal forms (the fold, the Hopf, the transcritical — the basic ways a stable state can lose stability). There are only a handful. So Bury’s team trained a deep neural network — a convolutional-plus-recurrent hybrid — on hundreds of thousands of simulated time series drawn from these universal templates, teaching it the subtle fingerprints of each route to collapse.

The payoff: their network spots tipping points in systems it was never trained on — including real climate and ecological data and a thermoacoustic experiment — more sensitively and with fewer false alarms than the classic indicators. And crucially, it does the thing the old signals can’t: it names the type of bifurcation, telling you whether the new state will be a sudden jump, an oscillation, or a gentle shift. Bauch called it a potential “game-changer for the ability to anticipate big shifts.” A 2025 follow-up (Zhuge, Li & Chen, Royal Society Open Science) pushed the approach toward the messiness of reality, designing a network to predict tipping occurrence even in the irregularly sampled time series that real-world monitoring actually produces — one of the sharpest limitations of the original method.

The honest label: this is a real and important advance, but its triumphs are still mostly in silico and on a handful of empirical cases. Whether a normal-form-trained network reliably calls tipping points in the full, noisy, high-dimensional mess of the real Earth system — with enough lead time to matter — remains to be demonstrated at scale. uncertain

The digital twin that visits the far side

Detecting that a cliff is near is one thing. A bolder 2020–2021 program asks: can a machine, having only ever watched a system behave normally, predict the exact point at which it will collapse — and then describe the alien world beyond? The tool is reservoir computing, a strikingly efficient flavour of recurrent neural network that has proven almost uncannily good at learning the “grammar” of chaotic systems from raw data, with no model of the underlying physics.

In a 2021 Physical Review Research paper, Ling-Wei Kong, Ying-Cheng Lai, and colleagues did something that sounds like cheating. They fed a reservoir computer not just the system’s behaviour but also the value of the slowly drifting parameter pushing it toward the edge — a “parameter input channel.” Trained only on data from the safe regime, the machine could then extrapolate to parameter values it had never seen, predicting both where the critical transition lies and — remarkably — statistical properties of the doomed transient state that precedes final collapse. Follow-up work (Dhruvit Patel and Edward Ott, Chaos, 2023) explicitly used machine learning to “anticipate tipping points and extrapolate to post-tipping dynamics” of changing systems, and Kong’s group has gone on to frame these trained reservoirs as digital twins of nonlinear systems — fast, data-driven stand-ins you can push past the cliff in software to see what happens, without pushing the real thing.

The dream is a wind tunnel for catastrophe: a learned replica you can crash a thousand times to map the edge of a system you must never actually break.

It is a genuinely beautiful idea, and on benchmark chaotic systems it works. But here the hype filter bites hardest. These successes are on relatively low-dimensional, well-sampled, cleanly-measured mathematical systems. The real climate is high-dimensional, sparsely observed, and non-stationary in ways that can defeat the method’s core assumption that the future “rhymes” with the trained past. The same community has published sober “Catch-22” analyses of reservoir computing’s limits. A digital twin of a chaotic toy is a triumph; a trustworthy digital twin of the AMOC does not yet exist. uncertain contested

How a pattern lets a system dodge the edge

Now a result that doesn’t predict tipping points so much as question them — and it may be the most quietly radical idea in this whole appendix. The standard picture, the one the main file sold you, is stark: push an ecosystem past its threshold and it collapses, abruptly and as a whole, from green to desert. In 2021, Max Rietkerk and colleagues argued in Science that for systems spread out in space, this picture is often wrong — and wrong in a hopeful direction.

Their claim: instead of collapsing all at once, a spatially extended system under stress can reorganize into a pattern — the spots, stripes, and labyrinths of vegetation you can see from the air in drylands the world over. These are Turing patterns, the self-organized structures Alan Turing predicted in 1952. And here is the twist that upends the textbook: where ecologists had long read these patterns as warning signs of imminent collapse, Rietkerk’s analysis says they can be the opposite — a sign of resilience. By breaking into patches, the system survives at stress levels far beyond where a uniform version would have tipped. It trades total collapse for a graceful, often reversible retreat into pattern. The cliff is replaced by a ramp.

Why it matters, and where the caution lives: if Rietkerk is right, then a whole class of ecosystems — and perhaps Earth-system components — flagged as “tipping-prone” may be substantially more resilient than the catastrophic-collapse models imply, because those models ignored spatial dynamics. That is genuinely good news, carefully argued with both mathematics and field observations. But it is a review and a theoretical framework, not a universal law: how far the escape hatch generalizes — to ice sheets, to the Amazon, to the whole Earth system — is exactly the open question its authors pose. And it cuts both ways for warning signals, because it means a pattern can signal either approaching collapse or hard-won persistence, and telling those apart is hard. uncertain

Reading the Earth’s resilience from orbit

The classic warning signals were born on small, clean datasets — a whole-lake experiment, an ice core. The post-2020 leap is to compute them everywhere at once, by turning decades of satellite imagery into a planetary resilience monitor. The logic is unchanged from the main file — a system losing resilience recovers more slowly, so its autocorrelation creeps up — but the canvas is now the entire land surface, pixel by pixel.

The results have arrived in a rush. In 2021, Niklas Boers found statistically significant early-warning signals across eight independent observational fingerprints of the AMOC, concluding the Atlantic circulation may have drifted, over the last century, “from relatively stable conditions to a point close to a critical transition” (Nature Climate Change). The same year, Boers and Rypdal reported critical slowing down suggesting the western Greenland Ice Sheet sits near a tipping point (PNAS). Then the satellites opened up the biosphere: in 2022, two landmark papers — Smith, Traxl & Boers in Nature Climate Change, and Forzieri and colleagues in Nature — used satellite vegetation indices to show that forests across much of the planet have been losing resilience since the early 2000s, most worryingly in the tropics, even as some northern forests gained it. Tim Lenton’s group has proposed building all of this into a permanent “resilience sensing system for the biosphere” (2022) — an early-warning dashboard for the living planet.

And the AMOC story carries its own built-in counterweight, which is exactly how healthy science should look. Even as the observational warning signals accumulated, a 2025 Nature study argued that wind-driven upwelling in the Southern Ocean may stabilize the Atlantic circulation enough to make an outright collapse this century unlikely. Both findings are serious; both are peer-reviewed; the disagreement is the actual state of the science. The fair 2026 summary: real, observation-based evidence of resilience loss is now being read off the planet for the first time promising, but converting those signals into confident statements about which system tips when remains contested contested.

Setting off the cascades we want

Every tipping point in this course so far has been a thing to fear or forecast. The final frontier program flips the sign. If a reinforcing loop can drag a system into ruin, the reasoning goes, then a reinforcing loop can also be deliberately seeded to flip a system toward something good — and fast. This is the science of positive tipping points, and it has become one of the most active and consequential corners of systems thinking since 2020, precisely because the math of decarbonization demands it: holding warming near 1.5°C requires emissions to fall faster than any gradual, linear policy is delivering. Only self-accelerating change is fast enough.

The agenda-setting paper was Ilona Otto, Jonathan Donges, and colleagues in PNAS (2020), who convened experts to identify “social tipping elements” — leverage-rich subsystems where a modest, well-placed push could trigger runaway change toward carbon neutrality. Their candidates: energy markets, financial divestment, carbon-neutral cities, norms and values, education, and — note the callback to Appendix I’s leverage ladder — information disclosure. Tim Lenton’s group followed with a framework for “operationalising” these points (2022) and a method to identify them while, in their own careful words, “avoiding wishful thinking about their existence.” The clearest real-world example is already here: the collapsing cost of solar power and batteries has, in sector after sector, crossed the point where clean energy simply out-competes fossil fuels, triggering self-reinforcing adoption — an S-curve, the signature of a reinforcing loop winning. Reinforcing feedback, the villain of the climate story, recast as its best hope.

The hype filter here is unusual, because the object is partly normative — a goal as much as a fact. Some positive tipping points are demonstrably real and already past (the UK’s exit from coal; solar’s cost crossover). Others are hopeful conjectures about social systems whose contagion dynamics we understand far less well than a melting ice sheet — and critics (in replies to Otto et al.) warned that treating society like a physical system risks ignoring conflict, politics, and “how change actually happens.” established uncertain

Six years, one accelerating field

Where each claim actually stands, mid-2026

• Deep learning can detect tipping & name the bifurcation type. Bury et al., PNAS 2021; Zhuge et al., R. Soc. Open Sci. 2025. Verdict: promising. Excellent in simulation & a few empirical cases; broad real-world reliability unproven.
• Reservoir computing predicts collapse & post-tipping dynamics. Kong & Lai, Phys. Rev. Research 2021; Patel & Ott, Chaos 2023. Verdict: promising, contested at scale. Works on chaotic benchmarks; trustworthy Earth-system “digital twins” don’t yet exist.
• Spatial patterns let systems evade catastrophic tipping. Rietkerk et al., Science 2021. Verdict: promising / paradigm-shifting. Well-argued theory + field cases; generality to ice sheets/whole Earth is open.
• We can read resilience loss off satellites globally. Smith et al. & Forzieri et al., 2022; Boers, Nat. Clim. Change 2021. Verdict: new capability, caveated. Striking global pattern; estimates unreliable in dense forest, signals can be ambiguous.
• Positive tipping points can be triggered to speed decarbonization. Otto et al., PNAS 2020; Lenton et al., 2022. Verdict: some cases real, partly aspirational. Solar/coal cases established; social contagion far less predictable than physics.
• But can we name the collapse date? Ben-Yami et al., Sci. Adv. 2024 (from the main file). Verdict: not yet. Detecting lost resilience ≠ forecasting timing. The deepest unsolved problem of all.

What the next six years must answer

• Can any method give a trustworthy lead time? Every program here improves detection; none has convincingly cracked timing for a real, half-observed Earth-system component. This is the field’s central, stubborn gap.
• Will AI predictors survive contact with the real Earth? Normal-form networks and reservoir twins shine on clean systems. The open test is whether they hold up on sparse, noisy, non-stationary planetary data — or quietly overfit the past.
• How far does the spatial escape hatch reach? Rietkerk’s evasion is demonstrated for drylands. Does it extend to forests, ice, ocean circulation — or are those the systems where the cliff is real after all?
• Can a falling pattern be told from a rising one? If spatial patterns can mean either imminent collapse or resilient persistence, every spatial early-warning signal now needs a way to read its sign. Several 2024–25 groups are racing at exactly this.
• Do social systems really tip like physical ones? The positive-tipping program borrows the mathematics of ice sheets for human behaviour. Where that analogy holds — and where politics and conflict break it — is being tested in real time, on the climate clock.