If the Atlantic current stoppedENRO

Loading the model run…

Model year25Model years, not calendar years
Current at 26° N

The Atlantic overturning circulation, or AMOC

If the Atlantic current collapsed: what happens to Europe’s winters, the sea ice and the rains

Researchers slowly added fresh water to the North Atlantic in a global climate model until the current that carries warm water north and cold water down and south collapsed. Scroll through the 2,200 simulated years, find your own town, and see why nobody can yet give a date for the real one.

Start the runGo to your town

Model year 25

The North Atlantic, before#

This is the North Atlantic inside CESM, a climate model of the whole Earth, as run by René van Westen, Michael Kliphuis and Henk Dijkstra at Utrecht University. Every colour on the map is the model’s own air temperature, compared with how the run began.

At 26° north the current carries 16 million cubic metres of water a second towards the Arctic, near the surface, and the same amount back south deep down. Oceanographers write that as 16 sverdrups. The real Atlantic, measured by a line of moorings since 2004, carries 17.1.

Model year 1500

Fresh water, a little more each year#

To find where the current breaks, the team added fresh water to the North Atlantic, a little more every year: 0.0003 sverdrups more each year, a slow and steady push. Fresh water is light, and light water does not sink.

By model year 1500 the current is down to 12 Sv and the sea south of Greenland has cooled. These are model years, not calendar years: the run starts from the climate of the 1800s, with no greenhouse warming, and the dose of fresh water grows far beyond anything Greenland’s melting ice delivers today.

Model year 1758

Model year 1758: the tipping point#

Here the current passes its tipping point, as the authors estimate it (between model years 1741 and 1775). Over the next hundred years the fresh water added rises by only 0.03 Sv, from about 0.53 to 0.56 Sv, yet the overturning falls from about 9 Sv to 2 Sv.

The current is now weakening itself: less sinking brings less salty water north, which means even less sinking.

Model year 1850

One hundred years#

Within a century, north-western Europe cools by more than 1 °C a decade, several times faster than the world is warming now. Averaged over Europe, February cools by 1.4 °C a decade and August by 0.3 °C.

In Bergen, on Norway’s west coast, February cools by 3.5 °C a decade.

Model year 2000

The ice comes south#

At the start, March sea ice in the north-east Atlantic stopped near 65° north. Now it reaches 49° north: down past the British Isles, into the English Channel.

White ice reflects sunlight that dark water would absorb, so the ice makes the cold worse.

Model year 2200

Where the heat went#

By model year 2200 the overturning has stopped. The heat it used to carry north stays in the south: the northern hemisphere ends 2.8 °C colder, the southern 0.9 °C warmer.

The heat the Atlantic carries north at 26° falls from 1.0 to 0.27 petawatts. About two thirds of the northern hemisphere’s cooling (66%) happened over model years 1750 to 1850.

How the current works#

Warm, salty water flows north near the surface of the Atlantic, the Gulf Stream and its continuation. Near Greenland and Iceland it gives its heat to the air in winter, becomes cold and dense, and sinks two or three kilometres. Down there it flows back south, all the way past the equator. The loop is called the Atlantic meridional overturning circulation, AMOC for short.

The section below is the model’s Atlantic cut open from south to north. Orange shows water turning clockwise in this view: north at the top, down near Greenland, south at depth. Drag the years and watch the loop shrink.

The overturning in the model’s Atlantic, south to north and surface to sea floor. Colour: sverdrups of water turning in the loop.
Measured at 26° north, monthly, April 2004 to November 2024 (RAPID).

In the real ocean, the moorings across the Atlantic at 26° north have measured an average of 17.1 sverdrups since April 2004, about 14 times the flow of all the world’s rivers put together. With it goes 1.22 petawatts of heat, roughly 60 times the rate at which humanity uses energy. The record wobbles a lot from month to month (from 6.7 to 24.9 Sv), and its last five years averaged 16.6 Sv against 18.4 in its first five.

Why a current can tip, in a model from 1961#

In 1961 the oceanographer Henry Stommel reduced the Atlantic to two boxes of sea water, one warm and one cold, joined by a pipe. Heat makes the flow run; fresh water slows it. His equations already had the strange property that worries scientists now: two stable states for the same amount of fresh water.

Move the slider to add fresh water. Past a point the strong flow has nowhere to go and drops to the weak state. Then try to bring it back: taking fresh water away again does nothing until all of it is gone, and a little more. This is a textbook model, not the real ocean; the climate model above has the same two states.

Strong flow
Stommel’s two-box model: flow strength against fresh water added. The grey line is the unstable middle branch.

Europe’s winters, month by month#

The cooling is not even through the year. Winter takes the worst of it, because in winter the ocean is what keeps the coasts of north-west Europe mild, and because sea ice now forms where there was open water.

The map shows how fast each month cools during the century of the collapse, model years 1750 to 1850, in degrees per decade. Click any town on the map to see its own year below.

How fast each month cools during the collapse (trend over model years 1750–1850), °C a decade.

Your town after the collapse#

Type a town or click the map. The grey line is the monthly mean temperature around your town in 1880–1909, from Berkeley Earth’s reconstruction of the measurements, the decades closest to the model’s starting climate. The blue line adds the model’s change between the start of the run and its end, once the current has stopped: an illustration from one model, not a dated forecast. The dashed line is the same months in 1995–2024, for comparison.

London: once the current has stopped, the year is 8.1 °C colder, and February 16.6 °C colder.

No big city’s climate around 1900 matches its new year month by month; the closest, Erzurum, is off by 3.0 °C a month. Yearly rain and snow in the model: −20% from the start of the run to its end.

1880–1909, estimated from measurements After the collapse (model change added) 1995–2024, estimated from measurements

Monthly mean temperature, °C.

The rain belt moves south#

The tropical rain belt follows the warmest part of the planet. When the north cools and the south warms, the belt slides south, and with it the wet season of whole countries.

In the model’s Amazon the seasons swap. January to March, part of the wet season, become the driest months; August and September, part of the dry season, turn wet. The Sahel, south of the Sahara, loses about 38% of its rain and South Asia about 9%.

Change in yearly rain and snow after the collapse (model years 2151–2200 against 1–50), % of the starting amount. Deserts are left grey.
The Amazon’s rain, month by month: before (orange) and after (blue) the collapse, mm.

Higher seas on Atlantic coasts, and the place where the water stops sinking#

The Gulf Stream holds the sea surface along the American coast lower than further out, the way a spinning current tilts the water it carries. When the current weakens, the tilt relaxes and water piles up against the coasts of the North Atlantic: 80 cm higher at New York, 79 cm at Boston, 77 cm off the Netherlands, on top of any rise from melting ice and warming water. The water comes from elsewhere: on average the North Atlantic rises 69 cm and the Pacific drops 20 cm.

The two small maps show the deepest winter mixing in the model, where surface water sinks. Before the collapse it reaches 1,650 metres in the Labrador and Nordic Seas. After, nowhere in the North Atlantic mixes deeper than 290 metres.

Change in sea level caused by the shift in currents (model years 2151–2200 against 1–50), cm.
Deepest winter mixing, before and after, m.

With global warming in the mix#

The run on this page starts from a climate without greenhouse warming, so its numbers show the collapse on its own. In June 2025 René van Westen and Michiel Baatsen published runs of the same model under the RCP4.5 scenario, which reaches about 2 °C of warming by 2100 and is then held at 2100 levels for centuries, with and without a collapsed current, each run until its climate settled, well beyond 2200. North-western Europe still cooled: the warming did not cancel the collapse.

Their worked example is De Bilt in the Netherlands. In that 2 °C world with the current collapsed, the coldest night that comes about once in ten years falls to about −20 °C; in the old pre-industrial climate it was about −6 °C with the current running and about −38 °C with it collapsed. The hottest day of a typical decade reaches 31.6 °C, against 34.1 °C in the same warmer world without the collapse. In late winter the sea ice reaches about 53° north and partly covers the North Sea. In their runs with about 4 °C of warming by 2100 (RCP8.5) the strong cooling and the ice did not appear: north-western Europe warmed, only less than elsewhere.

The patch of ocean south of Greenland#

Since 1900 the surface of the world’s ocean has warmed by about 0.6 °C a century on average. South of Greenland, one of the first regions to cool in the model, it warmed by about 0.2 °C, and about a third of that region cooled. Scientists call it the cold blob or the warming hole.

A weaker overturning would bring less warm water there, which is why the patch is used as a fingerprint of the current. The fingerprint is not proof: winds, cloud and air pollution also change the temperature of that sea, and direct measurements of the current go back only to 2004.

Trend in sea-surface temperature, 1900–2025, °C per century (HadISST). Cells covered by sea ice in some years are left out.
The fingerprint: the sea south of Greenland minus twice the global ocean average, yearly, °C, rebuilt from HadISST; 2021–2025 in colour.

When could it happen? The 2065 forecast, re-run#

In 2023 Peter and Susanne Ditlevsen fitted the fingerprint above with a model of a system approaching a tipping point. Their estimate, around 2057, travelled around the world. In August 2025 they corrected an error in their code, and the estimate became 2065, with a 95% confidence interval from 2037 to 2109. Both hold only if their statistical model and the fingerprint are right.

The method can be rerun. On their own series, the corrected method gives 2065 here too. Rebuilt from HadISST with a patch of sea that only approximates theirs, it gives 2056 with data to 2020: choices that small move the answer by years. Adding 2021–2025 to that rebuilt series moves the estimate to 2060. Five more years of data pushed the date 4 years later, and the years that fit nearly as well still run from 2041 to 2096.

The fingerprint rebuilt from HadISST, and the path the corrected method fits to it with data to 2025.
How well each tipping year fits the rebuilt series, with data to 2020 (grey) and to 2025 (blue). Lower is better; the shaded band is where the fit is nearly as good as the best.

The harder test is a world where the answer is known. The Utrecht group tried the method on their own model in 2024 and found that most 150-year stretches of data gave a wrong date. Applied here to the model’s own fingerprint, the corrected method made 128 forecasts: 32 overlapping 151-year stretches, each with four lengths of the calm period. The forecasts spread from model year 1697 to 2193; their median was 1787, against the paper’s estimated tipping point at 1758, and only 36% landed within 25 years of it. The closer the data came to the tipping point, the better the forecast: with data ending a century or more before it, they ranged from 1697 to 2193; with data ending 53 to 68 years before it, all 16 landed between 1746 and 1778.

The method tested inside the model: each dot is one forecast, made from 151 years of model data ending at the year on the left. The line marks the tipping point the authors estimate for this run.

Other studies give other answers. A set of climate models analysed by the Utrecht group in 2025 shows the AMOC could begin to collapse by 2063 under moderate emissions, with a full collapse taking more than a century. A study of 34 models in Nature, also in 2025, found a collapse (a fall below 6 Sv) unlikely this century, because winds around Antarctica keep pulling deep water up. The IPCC’s 2023 report says an abrupt collapse is not expected before 2100, with medium confidence. They measure different things and disagree about how likely a collapse is and when; none gives a reliable date.

Abrupt cooling of the North Atlantic has happened before#

About 12,900 years ago, as the last ice age was ending, the North Atlantic suddenly turned cold again for over a thousand years, a period called the Younger Dryas. The leading explanation is a weakened overturning, after fresh meltwater from the North American ice sheet reached the ocean.

It happened with ice sheets far larger than today’s, so it is a precedent for abrupt change, not a copy of what is described above. Ice drilled from the top of Greenland records it. In this reconstruction the summit averaged −48 °C during the cold spell. When it ended, the temperature climbed about 8 °C within a century.

Temperature at the top of the Greenland ice sheet over the last 20,000 years, from the GISP2 ice core (Alley 2000), °C.

What a collapse would not do#

The Gulf Stream would not stop. Most of its flow is driven by the winds and the Earth’s rotation; it would carry less heat north, but the current along the American coast would keep running.

It would not freeze the world in days, as in the film The Day After Tomorrow. In the model the fall from strong to weak takes about a hundred years, still very fast for a climate, and far faster than people and farms adapt to.

And it would not cancel global warming. In the model the southern hemisphere warms, because the heat that used to go north stays there, and the greenhouse gases would still be in the air.

How this page was made#

The map, the section, the monthly changes, the rain, the sea level and the mixing are the published output of van Westen, Kliphuis and Dijkstra (2024), a run of the Community Earth System Model with slowly increasing fresh water over 2,200 model years (output and scripts on Zenodo, CC BY 4.0). The map averages each decade; the section averages ten years around each frame. The run starts from a pre-industrial climate without greenhouse warming, so model years are not calendar years.

Town values combine Berkeley Earth’s measured monthly temperatures (1° grid; 1880–1909 and 1995–2024) with the model’s change from its start (model years 1–50) to its end (2151–2200), interpolated from its grid of about 2°. The model stores that change only as a yearly average, so each month gets the seasonal shape of the model’s monthly trends during the collapse (1750–1850), shifted so that the twelve months average to the direct yearly change. The model’s error is the difference between its starting climate in that grid square and the 1880–1909 measurements. The closest-city match compares the twelve monthly temperatures with those of every city over 300,000 people in 1880–1909; a city is offered as a match only when it differs by under 1.5 °C a month on average, otherwise the page names the closest one and its larger difference.

The forecast uses the Ditlevsen and Ditlevsen estimator, ported from the Utrecht group’s code, with the 2025 correction applied. The rebuilt series regresses the published fingerprint (HadISST as downloaded for the 2023 paper, 1870–2020) on HadISST (as downloaded in October 2026) averaged over 45–65° N, 50–20° W and over the global ocean (correlation 0.975), and applies that to every month from 1870 to 2025. It is not spliced onto the published series: in a separate test calibrated over 1990–2015, replacing only the published 2016–2020 moved the estimate from 2065 to 2070. Best years and the “nearly as good” band come from the profile likelihood (a drop of 1.92 in log-likelihood), not the authors’ bootstrap. The in-model test applies the same estimator to the model’s subpolar sea-surface temperature minus its global average, for 151-year windows (the authors’ grid) starting every 5 years from model year 1400 to 1555, with the calm part ending 50, 60, 70 or 80 years in.

Sea-surface trends are least-squares fits to yearly HadISST means; ERSST v5 shows the same pattern (in ERSST, about 23% of the area between 40° and 66° north and 60° and 10° west cooled, in HadISST 13%). The measured current is RAPID-MOCHA-WBTS version 2024-2.

Sources#

One sverdrup (Sv) is a million cubic metres of water a second: about 150 times the Danube where it reaches the sea, and close to all the world’s rivers put together (about 1.2 Sv).