Failure to track a stable AMOC state under rapid climate change René M. van Westen, Reyk Börner, Henk A. Dijkstra Abstract The Atlantic Meridional Overturning Circulation (AMOC) is a tipping element of the climate system. The current estimate of the global warming threshold for the onset of an AMOC collapse is +4.0 ∘ C. However, such a threshold may not be meaningful because AMOC stability depends on the rate of radiative forcing and background climate state. Here, we identify an AMOC stabilising mechanism that operates on timescales longer than present-day radiative forcing increase. Slow forcing permits coherent adjustment of surface and interior ocean properties, supported by enhanced evaporation and reduced sea-ice extent, counteracting destabilising feedbacks. This mechanism is explicitly demonstrated in a slow CO 2 increase experiment (+0.5 ppm yr -1 ), in which the AMOC remains stable up to +5.5 ∘ C of global warming. By contrast, under intermediate- and high-emission scenarios, the AMOC collapses at substantially lower warming levels (+2.2 ∘ C and +2.8 ∘ C, respectively). Our findings demonstrate the strong radiative forcing path dependence of AMOC tipping and imply that limiting the rate of radiative forcing is critical for reducing the near-term risk of an AMOC collapse. keywords AMOC, Climate change, Tipping risk Main Earth’s climate system contains various tipping elements in the biosphere, cryosphere and hydrosphere Lenton et al. 2025 . These tipping elements may undergo a transition upon exceeding a critical forcing, often expressed as global warming thresholds Armstrong McKay et al. 2022 . For the Atlantic Meridional Overturning Circulation (AMOC), the threshold for a collapse is estimated to be at global warming levels of +4 ∘ C, with a range from +1.4 ∘ C to +8.0 ∘ C Armstrong McKay et al. 2022 . In particular the lower bound of such estimates is important for society to identify safe operating spaces that minimise the risk of disruptive climate change Wunderling et al. 2023 ; Möller et al. 2024 . However, estimating a critical warming threshold for the AMOC poses several challenges, from limited paleo-climatic evidence and simulated tipping events to model biases causing large uncertainty in the AMOC response to radiative forcing Rahmstorf et al. 2005 ; Kriegler et al. 2009 ; Armstrong McKay et al. 2022 ; van Westen et al. 2025 . A key underlying assumption behind such a threshold is that the present-day AMOC state will lose stability if the corresponding warming level is exceeded, triggering the onset of a collapse Lenton et al. 2008 ; Armstrong McKay et al. 2022 ; Lenton et al. 2025 . Combining this assumption with the future warming levels projected by models forced with the Shared Socioeconomic Pathways (SSPs), it seems possible to determine when an AMOC collapse would start. Yet, many results indicate that the behaviour of the AMOC depends on the rate of radiative forcing change – in observations Dima et al. 2025 , conceptual climate models Stocker and Schmittner 1997 , early global climate models Manabe and Stouffer 1999 , and recently in a low-resolution earth system model Hankel 2025 . In addition, modeling studies have found sustained strong AMOC states in very warm climates Bonan et al. 2022 ; Romanou et al. 2023 ; Willeit and Ganopolski 2024 , for example under quadrupled CO 2 levels. While studies have emphasised the forcing rate effects on AMOC weakening Manabe and Stouffer 1999 ; Hankel 2025 , the question whether the forcing rate alone can determine AMOC tipping has been less investigated. An early study demonstrated in a zonally-averaged three-basin ocean model Stocker and Schmittner 1997 ; Schmittner and Stocker 1999 that a faster rate of radiative forcing can indeed collapse the AMOC than a slower forcing rate. This is one of the first examples showing rate-dependent tipping of the AMOC Ashwin et al. 2012 ; Wieczorek et al. 2023 , meaning that it fails to track its current stable state for sufficiently fast forcing change. An additional problem with a warming threshold is that its destabilising feedback mechanism may operate independently of global temperatures. While certain tipping elements (e.g., Greenland Ice Sheet and West Antarctic Ice Sheet) are losing stability directly due to higher temperatures Waibel et al. 2018 ; Mouginot et al. 2019 ; Boers and Rypdal 2021 ; Turney et al. 2020 , the AMOC is destabilised through the salt-advection feedback, which is particularly sensitive to surface freshwater fluxes in the North Atlantic Ocean Stommel 1961 ; Rahmstorf 1996 ; Marotzke 2000 ; Jackson et al. 2023 ; van Westen et al. 2024 . Increased freshwater input by the Greenland Ice Sheet may further weaken the AMOC Pontes and Menviel 2024 , but this contribution alone requires unrealistically large melt rates to collapse the present-day AMOC van Westen et al. 2025 . Much progress has been made in assessing the stability properties of the AMOC with respect to varying freshwater forcing conditions, but the interacting effects of changing heat and freshwater fluxes under global warming remain insufficiently understood. Motivated by these issues with a global warming threshold for an AMOC, we here investigate the radiative forcing path dependence of AMOC tipping in the Community Earth System Model (CESM, version 1.0.5) and several models participating in the Coupled Model Intercomparison Project (CMIP) phase 6. Guided by a recent quasi-equilibrium freshwater flux simulation with the CESM, in which an AMOC collapse has been found van Westen et al. 2024 , we extend earlier work on rate-dependence in more detail Stocker and Schmittner 1997 ; Schmittner and Stocker 1999 and beyond AMOC weakening Manabe and Stouffer 1999 ; Hankel 2025 . We will demonstrate a strong radiative path dependence of AMOC tipping, questioning the suitability of global warming thresholds for assessing AMOC tipping risk. Stable AMOC under extreme climate change We will present an in-depth analysis of a slow CO 2 increase simulation (+0.5 ppm yr -1 ) with the CESM, together with a comparison of different Representative Concentration Pathway (RCP) scenarios that were already presented previously van Westen et al. 2025 . All climate model simulations are branched from the quasi-equilibrium hosing simulation under constant pre-industrial (PI) forcing conditions van Westen and Dijkstra 2023 ; van Westen et al. 2024 , which we need to describe in more detail first. In the quasi-equilibrium PI hosing simulation, a freshwater flux forcing ( F_{H} ) was applied over the North Atlantic Ocean (20 ∘ N – 50 ∘ N) and this forcing was compensated elsewhere (at the ocean surface) to conserve ocean salinity. The slowly-varying F_{H} ensures that AMOC transitions are caused by intrinsic feedbacks and multiple statistical equilibria (with time-invariant statistics) were obtained at fixed F_{H} van Westen et al. 2024 ; van Westen et al. 2025 , the latter indicated by \overline{F_{H}} . The quasi-equilibrium hosing simulation and the statistical equilibria for \overline{F_{H}}=0.18 Sv and \overline{F_{H}}=0.45 Sv are shown in Figure 1 a. A tipping point (i.e., a saddle-node bifurcation) exists under an increasing freshwater flux forcing, which is located at F_{H}\approx 0.5 Sv for this CESM version van Westen et al. 2025 , and the existence of a tipping point under varying freshwater flux forcing is consistent with results in a hierarchy of climate models Stommel 1961 ; Rahmstorf et al. 2005 ; Hawkins et al. 2011 ; Hu et al. 2012 . We focus on the statistical equilibrium for \overline{F_{H}}=0.45 Sv in the relatively strong AMOC regime, which is in the destabilising salt-advection feedback regime and hence more prone to AMOC transitions than the statistical equilibrium for \overline{F_{H}}=0.18 Sv van Westen et al. 2025 . Figure 1: Climate model simulations with the CESM. (a): The AMOC strength (at 26 ∘ N and 1,000 m depth) for the quasi-equilibrium PI hosing simulation, including four statistical equilibria. (b): The AMOC strength for the historical and three extended RCP scenarios and \overline{F_{H}}=0.45 Sv. The inset shows the GMST anomaly compared to the period 1850 – 1899. (c): The AMOC strength and GMST anomaly (compared to the first 50 model years) for the CO 2 ramp simulation. (d): The AMOC versus GMST anomaly for the simulations presented in panels b & c. The dashed lines indicate the GMST thresholds for AMOC tipping for the RCP4.5 (+2.2 ∘ C) and RCP8.5 (+2.8 ∘ C) scenarios van Westen et al. 2025 . In all panels, the thin curves are yearly averages, whereas the thick curves are smoothed versions (25-year moving averages). Different radiative forcing simulations were branched from the end of the statistical equilibrium for \overline{F_{H}}=0.45 Sv and under fixed \overline{F_{H}} . The historical forcing (1850 – 2005) was followed by three RCP (2006 – 2100) simulations. These RCP scenarios were subsequently continued to model year 2500 under their fixed greenhouse gas and aerosol concentrations of the year 2100 (Figure 1 b). This was done to determine statistical equilibria under different climate change scenarios van Westen et al. 2025 . The AMOC recovers under RCP2.6, while the AMOC collapses under RCP4.5 and RCP8.5. For RCP4.5 (RCP8.5), the AMOC reaches its tipping point around model year 2110 (2060), with an associated global mean near-surface temperature (GMST) anomaly of +2.2 ∘ C (+2.8 ∘ C) van Westen et al. 2025 . The RCP4.5 warming level could be used as lower bound for the tipping threshold since the AMOC does not collapse under RCP2.6. Do note that this temperature threshold is also dependent on \overline{F_{H}} , as the AMOC recovers under RCP4.5 and \overline{F_{H}}=0.18 Sv van Westen et al. 2025 ; this simulation will also be analysed below. In addition to the RCP simulations, in which different radiative forcing agents (e.g., CO 2 , methane, aerosols) are varied, we performed a separate simulation, below referred to the CO 2 ramp simulation, where the CO 2 concentration increases from 284.7 ppm at a slow rate of +0.5 ppm yr -1 for 1750 model years, while keeping the other forcing agents at their PI levels (and again with \overline{F_{H}}=0.45 Sv). We choose a linear increase instead of the more common approach using exponential growth rates (e.g. Hankel 2025 ) for two reasons. First, it avoids high CO 2 concentrations at the end of the simulation, as our initial growth rate (0.176% yr -1 ) would have resulted in about 6140 ppm (22 \times CO 2 ) at the end of the simulation. Second, it still leads to (roughly linear) global warming of about +0.3 ∘ C per century, even though the radiative forcing depends logarithmically on CO 2 concentration. Nevertheless, the CO 2 ramp simulation reaches 1,120 ppm (4 \times CO 2 ) in model year 1708. Apart from natural variability, the AMOC remains stable in this simulation while the GMST anomaly reaches +5.5 ∘ C over the last 50 model years (Figure 1 c). It is clear that the GMST thresholds of the RCP4.5 and RCP8.5 simulations do not hold for the CO 2 ramp simulation (Figure 1 d). In fact, the CO 2 ramp simulation reaches higher GMST anomalies than the RCP8.5 simulation. This is the first indication that there is a radiative forcing path dependence on AMOC tipping in the CESM. To understand this behaviour in the CESM, we will analyse the CO 2 ramp simulation in greater detail below. Changing AMOC properties under climate change The AMOC in depth coordinates and meridional heat transport (MHT) for model years 1 – 50 and 1701 – 1750 in the CO 2 ramp simulation are shown in Figures 2 a,b, respectively. The overall overturning structures and MHTs do not change much, though the depth of the northward overturning cell does become smaller from about 2800 m depth (model years 1 – 50) to about 2400 m depth (model years 1701 – 1750). The AMOC in density coordinates shifts to lighter water classes between the two periods (Figures 2 c,d), with the section-averaged depth levels of 20 m, 500 m and 3000 m getting lighter by about 0.90 kg m -3 , 0.65 kg m -3 and 0.25 kg m -3 , respectively. The relatively light water masses near the surface are becoming lighter more rapidly than the relatively heavy water masses at depth, which is related to the greater warming closer to the surface (Figure 2 e). The salinity increases over most parts of the Atlantic Ocean (Figure 2 f), which makes water masses more dense and hence partially offsetting the warming-induced response of the density. Figure 2: Oceanic Responses in the Atlantic Ocean (a&b): The time-mean AMOC in depth coordinates for model years a) 1 – 50 and b) 1701 – 1750. The lower panel shows the meridional heat transport (MHT). (c&d): The time-mean AMOC in density coordinates for model years c) 1 – 50 and d) 1701 – 1750. The three curves represent the (section-averaged) depth level, whereas the 20 m depth contour is smoothed to reduce its meridional variability. (e): The zonally-averaged temperature difference between model years 1701 – 1750 and 1 – 50 (shading). The curves are three isotherms of model years 1 – 50 for reference. (f): Similar to panel e, but now for the zonally-averaged salinity. Water mass transformation (WMT, see Methods) primarily takes place over the isopycnal outcropping region (40 ∘ N – 65 ∘ N) in the North Atlantic Ocean, which is crucial for sustaining an adiabatic AMOC Nikurashin and Vallis 2012 ; Wolfe and Cessi 2014 . Under the assumptions of thermal wind balance Vanderborght et al. 2025 and that WMT mainly takes place near the surface, the adiabatic AMOC can be reconstructed from surface buoyancy fluxes Walin 1982 ; Marshall et al. 1999 ; Grist et al. 2009 ; Desbruyères et al. 2019 . Indeed, the AMOC at 40 ∘ N (in density coordinates) and the surface-forced AMOC between 40 ∘ N – 65 ∘ N (indicated by \Delta\Psi_{\mathrm{surf}} ) closely resemble (red and black curves in Figure 3 a, respectively). The surface-forced AMOC is mainly driven by oceanic heat loss to the atmosphere whereas freshwater fluxes have a limited contribution (yellow and blue curves in Figure 3 a), consistent with observations Marsh 2000 . Surface-induced WMT rates that contribute to the water supply of the North Atlantic Deep Water (NADW) are found for density classes of \sigma_{2}\geq\sigma_{2}^{\mathrm{max}} van Westen et al. 2025 , with \sigma_{2}^{\mathrm{max}} being the density level of the maximum AMOC strength at 40 ∘ N (dashed red line in Figure 3 a). The reconstructed AMOC strength from surface buoyancy fluxes is then defined as \Psi_{\mathrm{NADW}}(t)=\Delta\Psi_{\mathrm{surf}}(\sigma_{2}^{\mathrm{max}}(t)) , which can also be decomposed into a thermal ( \Psi_{\mathrm{NADW}}^{T} ) and haline ( \Psi_{\mathrm{NADW}}^{S} ) contribution. Note that \Psi_{\mathrm{NADW}} approximates the actual AMOC strength, because \Psi_{\mathrm{NADW}} does not consider subsurface mixing and volume tendencies Groeskamp et al. 2019 . In addition, WMT contributions outside the 40 ∘ N – 65 ∘ N latitude band are relatively small van Westen et al. 2025 . Figure 3: Surface-forced AMOC (a): The surface-forced AMOC between 40 ∘ N and 65 ∘ N ( \Delta\Psi_{\mathrm{surf}} , sinking rates) over the first 50 model years in the CO 2 ramp, including the temperature and salinity contributions. The AMOC at 40 ∘ N in density coordinates with its maximum indicated by the dashed red line ( \sigma_{2}^{\mathrm{max}} ) are also shown. (b): The \Psi_{\mathrm{NADW}} and \sigma_{2}^{\mathrm{max}} (inset). (c): The spatially-averaged (40 ∘ N and 65 ∘ N) surface buoyancy flux, decomposed into the heat and freshwater fluxes. (d): The \Psi_{\mathrm{NADW}} differences compared the first 50 model years, including the temperature and salinity contributions. The differences for the maximum AMOC strength at 40 ∘ N are also shown. The time series in panels b,c,d are smoothed through a 25-year running mean (to reduce the variability). The quantity \Psi_{\mathrm{NADW}} is useful as near zero values indicate that an adiabatic AMOC cannot be sustained van Westen et al. 2025 . In the CO 2 ramp simulation, \Psi_{\mathrm{NADW}} primarily increases while \sigma_{2}^{\mathrm{max}} decreases (Figure 3 b). This combined response means that both the interior and surface sinking region (40 ∘ N – 65 ∘ N) are getting lighter, and that surface buoyancy fluxes change accordingly to maintain an adiabatic AMOC. The former is also demonstrated in Figure S1 a, which shows that the surface sinking regions and \sigma_{2}^{\mathrm{max}} change at about the same rate. This is also true for the RCP2.6 simulation (Figure S1 b), but for the RCP4.5 and RCP8.5 simulations (Figure S1 c,d), the surface sinking region is getting lighter faster than \sigma_{2}^{\mathrm{max}} and this closes the adiabatic AMOC pathways. This demonstrates that rate-dependent effects are important for the AMOC responses, not only for AMOC weakening Hankel 2025 , but also for AMOC tipping risk. This is the second indication for a radiative forcing path dependence of AMOC tipping. Apart from the potential density responses, the surface buoyancy fluxes ( B_{\mathrm{flux}} , Figure 3 c) also change and induce \Psi_{\mathrm{NADW}} responses (Figure 3 d). The initial AMOC weakening during the first 300 model years is attributed to changing heat fluxes ( \Psi_{\mathrm{NADW}}^{T} ) under global warming, whereas freshwater fluxes ( \Psi_{\mathrm{NADW}}^{S} ) have the opposite response. Thereafter, both the heat and freshwater fluxes contribute to increasing \Psi_{\mathrm{NADW}} values. The \Psi_{\mathrm{NADW}} and B_{\mathrm{flux}} are closely related ( R^{2}=0.896 ), but note that a part of B_{\mathrm{flux}} does not supply water to the adiabatic AMOC pathways. Nevertheless, the dominant B_{\mathrm{flux}} contributions (e.g., longwave radiation and sea-ice melt) can be analysed in greater detail and recent work suggests that a sign change in B_{\mathrm{flux}} can be used to indicate the onset of the AMOC collapse van Westen et al. 2025 . The tendency of B_{\mathrm{flux}} to become more negative, indicates that the AMOC stabilises in the warmer climate. For the surface heat fluxes (Figure S2 a), the dominant contributions come from more outgoing longwave radiation (stabilising) due to higher sea surface temperatures, more latent heat release (stabilising) due to enhanced evaporation rates, and more shortwave absorption (destabilising) due to a lower albedo. The North Atlantic sea-ice extent retreats northward under higher temperatures (inset in Figure S2 b) and this reduced sea-ice extent limits sea-ice insulation effects that tend to destabilise the AMOC Lin et al. 2023 ; van Westen et al. 2024 . Another stabilising sea-ice effect is the reduced melting contribution to the surface freshwater flux changes (Figure S2 b). This sea ice forms around Greenland and by advection ends up in the 40 ∘ N – 65 ∘ N latitude band where it melts during boreal spring Vanderborght et al. 2025 . Enhanced evaporation (stabilising) and precipitation (destabilising) rates also have a substantial contribution, but their opposing effects compensate to about zero. The remaining B_{\mathrm{flux}} components have a relatively small contribution. The aforementioned sea-ice responses that influence AMOC strength in the CO 2 ramp simulation and the RCP scenarios van Westen et al. 2025 connect well with the proposed mechanism by Hankel 2025 . Moreover, our analysis indicates additional B_{\mathrm{flux}} (or WMT) contributions that modulate AMOC strength. In the CO 2 ramp simulation, the AMOC strength undergoes a relatively rapid increase by about 3 Sv around model year 800 (Figure 1 c), which is related to the retreating sea-ice extent over the North Atlantic Ocean. Upon initialisation, the sea ice extends relatively far south (Figure S3 a), which is related to a weak AMOC strength of about 13 Sv under the \overline{F_{H}}=0.45 Sv hosing. Sea ice limits ocean-atmosphere exchange and deep convection Lin et al. 2023 , hence the mixed layer depth is relatively shallow ( <100 m) where sea ice is present. What is important here is that there is no deep convection over the Labrador Sea due to the extensive sea-ice extent. When sea ice retreats poleward under higher temperatures, the Labrador Sea becomes less sea-ice covered and this allows for deeper mixed layer depths (Figures S3 b,c,d). Vertical mixing brings relatively warm subsurface water masses to the surface, strongly contributing to the retreating sea-ice extent over the Labrador Sea, and giving rise to non-linear responses under the linear CO 2 concentration increase. Labrador Sea deep convection starts from model year 770 and onwards, which is followed by the 3 Sv AMOC strengthening around model year 800. The role of deep convection on WMT rates is quite complex Brüggemann and Katsman 2019 , and this convection certainly plays an important role in modulating AMOC strength and AMOC tipping Drijfhout et al. 2025 . Hence, we expect that the 3 Sv increase in AMOC strength can only be found when additional deep convection sites are activated as a result of retreating sea ice. In summary, the analysis of the CO 2 ramp and the comparison with the RCP scenarios (Figure S1 , more details in van Westen et al. 2025 ) show clear indications of rate-induced effects on AMOC responses. In the CO 2 ramp simulation, in which the warming rate is (much) smaller than the RCP scenarios, the AMOC remains stable up to +5.5 ∘ C warming. In fact, from the \Psi_{\mathrm{NADW}} and B_{\mathrm{flux}} analysis, it appears that the AMOC is getting more stable in warmer climates. Freshwater budget under climate change The Atlantic freshwater budget is crucial for understanding AMOC stability through different processes, which include overturning, gyre, atmosphere and sea-ice feedbacks Vanderborght et al. 2025 . The dominant overturning feedback is the so-called salt-advection feedback Stommel 1961 ; Rahmstorf 1996 ; Marotzke 2000 and is quantified by the AMOC-induced freshwater transport at 34 ∘ S, indicated by F_{\mathrm{ovS}} (see Methods). A more negative F_{\mathrm{ovS}} is linked to a less stable and weaker AMOC Vanderborght et al. 2025 , a relation found in conceptual models van Westen et al. 2025 and climate models Van Westen and Dijkstra 2024 ; van Westen et al. 2025 . The Atlantic freshwater budget and the meridional freshwater convergences are shown in Figures 4 a,b. We stress that the Atlantic freshwater budget and the F_{\mathrm{ovS}} as stability indicators are only applicable under (quasi-)equilibrium conditions Rahmstorf 1996 ; Vanderborght et al. 2025 , which certainly do not hold for the RCP scenarios van Westen et al. 2025 . The warming rate in the CO 2 ramp simulation is on average 0.03 ∘ C per decade, a factor of 10 slower than the RCP8.5 scenario over the 21st century and of that being currently observed Hansen et al. 2023 . Figure 4: The Atlantic freshwater budget (a): The Atlantic Ocean (34 ∘ S to 65 ∘ N) freshwater budget for the CO 2 ramp simulation with the freshwater content ( \overline{W} ), freshwater convergence ( F_{\mathrm{con}} ), surface freshwater fluxes ( F_{\mathrm{surf}} ) and changes in the freshwater content ( \frac{\mathrm{d}\overline{W}}{\mathrm{d}t} ). The quantity \overline{W} is split into an upper 1,000 m contribution ( \overline{W}_{1000\uparrow} ) and below 1,000 m contribution ( \overline{W}_{1000\downarrow} ) and are displayed as their differences compared to their time mean over the first 50 model years. (b): The meridional freshwater convergences (omitting the contribution by the Strait of Gibraltar) for the different freshwater transport components. (c): The salinity along 34 ∘ S over the first 50 model years, where the dashed lines indicate the different water masses which are based on the meridional velocity profile (see procedure outlined in Van Westen and Dijkstra 2024 ). (d): The zonally-averaged salinity at 34 ∘ S over time, which are displayed as differences compared to the first 50 model years. The dashed lines are the different water masses. The arrows are indicative of the meridional velocity (right = northward, left = southward) associated with the AMOC. The freshwater budget decomposition indicates that the Atlantic Ocean is salinifying through surface freshwater fluxes (i.e., net evaporation) and this is partially opposed by the horizontal convergences (see Methods), where both the (azonal) gyre ( \Delta F_{\mathrm{az}}=F_{\mathrm{azS}}-F_{\mathrm{azN}} , before model year 800) and overturning ( \Delta F_{\mathrm{ov}}=F_{\mathrm{ovS}}-F_{\mathrm{ovN}} , after model year 800) components are important. The increase in \Delta F_{\mathrm{az}} is primarily caused by a declining gyre freshwater transport at 65 ∘ N over the first 800 model years and remains constant thereafter (Figure S4 b). This response appears to be related to the retreating North Atlantic sea-ice extent. The freshwater transport by the overturning component at 34 ∘ S starts to contribute after model year 800 (Figure S4 a), which is also reflected by the salinity responses at 34 ∘ S (Figures 4 c,d). The AMOC starts to import freshwater anomalies over the upper 1000 m and to export salinity anomalies between 1000 – 3500 m depths (see also Figure S4 c). The decreasing F_{\mathrm{ovS}} variance (Figure S4 d), together with larger F_{\mathrm{ovS}} values, are indicative of increasing AMOC stability van Westen et al. 2024 ; van Westen et al. 2025 . This freshwater budget analysis supports the argument that there is no specific GMST threshold as the AMOC is getting more stable in warmer climates, meaning that AMOC tipping under anthropogenic climate change is primarily controlled by the warming rate. The salinity responses at 34 ∘ S influence the F_{\mathrm{ovS}} and it is therefore important to understand the drivers. One driver is the higher oceanic temperatures across the Atlantic Ocean (Figure 2 e), which result in higher evaporation rates at the surface and contribute to near-surface salinity anomalies (Figure 2 f). These near-surface salinity anomalies are spreading through the Atlantic Ocean by horizontal advection, vertical mixing, and WMT at the higher latitudes. A part of the salinifying water masses that undergo WMT eventually end up in the southward flowing NADW, explaining the saltier NADW that contributes to increasing F_{\mathrm{ovS}} values (Figure S4 c). These NADW salinity anomalies become more pronounced after model year 800 in the CO 2 ramp simulation (Figure 4 d), which is again indicative of a more stable AMOC. To further explore this stabilising response (i.e., salinifying NADW), we consider the Hist/RCP4.5 and \overline{F_{H}}=0.18 Sv and Hist/RCP2.6 and \overline{F_{H}}=0.45 Sv cases of our CESM simulations, in which the AMOC remains sufficiently strong, and several CMIP6 models under the extended SSP scenarios Romanou et al. 2023 ; Drijfhout et al. 2025 . In the CESM, the stabilising NADW response is found for the Hist/RCP4.5 and \overline{F_{H}}=0.18 Sv, but not for the Hist/RCP2.6 and \overline{F_{H}}=0.45 Sv (Figure S5 ). This difference between the two RCP scenarios could be related to the different \overline{F_{H}} values but it is more likely related to the GMST anomaly. Higher temperature anomalies induce more evaporation which eventually result in a saltier NADW; this is indeed the case for the CO 2 ramp simulation (Figure 4 d). The NADW salinity anomalies become more pronounced after model year 700 and under GMST anomalies higher than +2.5 ∘ C in the CO 2 ramp. The Hist/RCP4.5 and \overline{F_{H}}=0.18 Sv exceeds this warming level around model year 2200 and thereafter NADW salinities increase, whereas the Hist/RCP2.6 and \overline{F_{H}}=0.45 Sv stays below +2 ∘ C warming without a clear salinifying NADW response. Next, we investigate the stabilising NADW response in different CMIP6 models, starting with the GISS-E2-1-G under the extended SSP1-2.6 and SSP2-4.5 (Figure S6 ). We find the stabilising NADW response under the intermediate emission scenario of SSP2-4.5, but this is less pronounced under SSP1-2.6. Most CMIP6 models under the extended SSP1-2.6 scenario show an increase in NADW salinity between 1000 – 2000 m depths (Figure S7 ), but these responses are quite small and the NADW becomes even fresher for the IPSL-CM6A-LR. As was argued above, we expect a saltier NADW under higher emissions scenarios, but this cannot be verified as all CMIP6 models show a collapsing AMOC under the high-emission scenario of SSP5-8.5 Drijfhout et al. 2025 . Although the signal-to-noise ratio is relatively small, most CMIP6 models do align with our CESM results, indicating a robust stabilising response in warmer climates. The near-surface freshening at 34 ∘ S (Figure 4 d) is also contributing to increasing values of F_{\mathrm{ovS}} (Figure S4 c), although one may instead expect salinity anomalies by increased evaporation rates under higher surface temperatures. To understand this response, the upper 100 m salinity and surface freshwater flux differences are displayed in Figure 5 a and 5 b, respectively. The Atlantic near-surface waters are salinifying, whereas other ocean basins show a near-surface freshening. Indeed, the surface freshwater flux differences are mostly negative (i.e., salinifying) over the Atlantic basin, but the same holds for the Indian Ocean. This means that the horizontal convergence of relatively fresh water masses are dominant and offset the negative surface freshwater flux differences over the Indian Ocean. For example, the near-surface Pacific Ocean is freshening and, through the Indian Throughflow, contributes to the freshening of the Indian Ocean. These fresher Indian Ocean water masses enter the Atlantic Ocean through Aghulas Leakage Großelindemann et al. 2025 and then contribute to increasing F_{\mathrm{ovS}} values. There are large inter-model differences when comparing the near-surface water masses between the CMIP6 models (Figures S6 and S7 ), suggesting contributions from the atmosphere Van Westen and Dijkstra 2024 and Aghulas Leakage Großelindemann et al. 2025 , but these are not further explored here. Figure 5: Salinity and freshwater flux responses (a): The depth-averaged (upper 100 m) salinity differences between model years 1701 – 1750 and 1 – 50. (b – f): The surface freshwater flux differences and its decomposition between model years 1701 – 1750 and 1 – 50 (positive = relative surface freshening). In all panels, the circled markers indicate non-significant ( p\geq 0.05 , two-sided Welch t-test) differences, markers were not displayed when no sea ice was present (panel f). The total surface freshwater flux can be decomposed into its evaporation (Figure 5 c) and precipitation (Figure 5 e) contributions, with their combined contributions (Figure 5 d) very close to the total (Figure 5 b). Evaporation rates increase over almost all ocean surfaces and this results in higher precipitation rates, indicative of an intensified hydrological cycle in warmer climates. Note that evaporation and precipitation do not capture all freshwater flux differences at the higher latitudes, where sea ice also modulates the local salinity budget through melt and brine rejection. For example south of Greenland, the negative freshwater flux differences are strongly connected to sea-ice melt (Figure 5 f) and not so much by precipitation and evaporation. As was argued in the previous section, the sea-ice cover retreats poleward (Figure S3 ) and hence this melting contribution reduces to near zero values, explaining the negative freshwater flux differences there. The Atlantic basin is a net evaporative basin (Figure 4 a) and this characteristic becomes more pronounced in warmer climates. Freshwater anomalies of Atlantic Ocean origin are transferred to other ocean basins via atmospheric bridges, while residual salinity anomalies feed into the NADW. The retreating North Atlantic sea-ice extent salinifies the North Atlantic surface, also contributing to the salinifying NADW and limiting sea-ice insulation effects van Westen et al. 2024 . At 34 ∘ S, the NADW becomes saltier and contributes to increasing F_{\mathrm{ovS}} values, the latter indicating that the AMOC is getting more stable when the AMOC remains close to its equilibrium state Vanderborght et al. 2025 . This NADW response is also found in most CMIP6 models under the extended SSP1-2.6 and SSP2-4.5 scenarios and, together with the results of an intermediate-complexity climate model Willeit and Ganopolski 2024 and CMIP5 models Bonan et al. 2022 , demonstrates a robust stabilising response for the AMOC in warmer climates. This also means that the destabilising effects of anthropogenic climate change at high enough forcing rate and can become larger than the stabilising responses, such that the AMOC is brought out of equilibrium and subsequently collapses through a rate-induced AMOC tipping event. The critical radiative forcing rate for which the AMOC collapses can in principle be obtained for the CESM, but this requires substantial computational resources and is out the scope of this paper. Dynamical Perspective To provide a dynamical perspective on the radiative forcing path dependence of AMOC tipping, we use an idealised 5-box ocean model (Figure S8 a) that captures the essential AMOC dynamics of the CESM van Westen et al. 2025 . We focus on the response to destabilising heat and stabilising freshwater fluxes in the absence of sea ice; sea-ice insulation effects were previously investigated van Westen et al. 2024 . We will show that the different effective timescales of these processes can lead to rate-dependent tipping under climate change. The two main forcing parameters are the atmospheric temperature anomaly over the subpolar Atlantic box ( \Delta T_{\mathrm{n}}^{a} ) and the asymmetric freshwater flux forcing ( E_{A} ). The effects of global warming are incorporated by increasing \Delta T_{\mathrm{n}}^{a} to lower the meridional atmospheric temperature gradient between the lower and higher latitudes, mimicking polar amplification van Westen et al. 2025 . Additionally, we consider an effective freshwater flux forcing: E_{A}=\overline{E_{A}}-\gamma\Delta T_{\mathrm{n}} , with \overline{E_{A}} a fixed freshwater flux forcing (similar to the CESM), \Delta T_{\mathrm{n}} the oceanic temperature anomaly of the subpolar Atlantic box ( \Delta T_{\mathrm{n}}=T_{\mathrm{n}}(t)-T_{\mathrm{n}}(\overline{E_{A}},\Delta T_{\mathrm{n}}^{a}=0 )), and \gamma the temperature–freshwater coupling strength. This temperature–freshwater coupling is motivated by our CESM results, where in the slow CO 2 ramp simulation, the GMST anomaly and the Atlantic surface freshwater flux scale linearly ( R^{2}=0.98 , 25-year windows), with the surface freshwater flux declining by 0.069 Sv per degree warming (giving \gamma\approx 0.069 Sv per ∘ C). First, we consider no temperature–freshwater coupling ( \gamma=0 ). The steady states and bifurcations of the box model can be determined using continuation techniques and we choose E_{A} as the bifurcation parameter (Figures S8 b). Increasing the freshwater flux forcing reduces the distance of the ‘AMOC on’ state to the saddle-node bifurcation, but we consider a fixed freshwater flux forcing ( \overline{E_{A}} ). The distance can also be reduced by imposing \Delta T_{\mathrm{n}}^{a}=5^{\circ} C, for which the saddle-node bifurcation shifts in freshwater forcing space from E_{A}=0.486 Sv to E_{A}=0.342 Sv (black curve in Figure 6 a). In other words, for \overline{E_{A}}>0.342 Sv the AMOC always collapses under the imposed warming anomaly as the ‘AMOC on’ state disappears, which is a case of bifurcation-induced tipping (red curve in Figure 6 a). The ‘AMOC off’ state remains the only stable solution and the system transitions to the collapsed state. Note that our CESM results suggest that even for \overline{F_{H}}=0.45 Sv, no such bifurcation occurs the under CO 2 ramp forcing to at least +5.5^{\circ} C global warming. Figure 6: AMOC transitions in the 5-box model (a): The position of the saddle-node bifurcation of the stable AMOC on state for increasing \Delta T_{\mathrm{n}}^{a} . Three different forcing paths, which include: bifurcation-induced tipping (B-tipping, \gamma=0 Sv per ∘ C, any positive T_{\mathrm{n}}^{a} trend), fast warming case ( \gamma=0.04 Sv per ∘ C, \Delta T_{\mathrm{n}}^{a} = +0.05 ∘ C yr -1 ), and slow warming case ( \gamma=0.04 Sv per ∘ C, \Delta T_{\mathrm{n}}^{a} = +0.01 ∘ C yr -1 ). (b): The critical temperature trend of T_{\mathrm{n}}^{a} without collapsing the AMOC for varying \overline{E_{A}} and temperature-freshwater coupling strength \gamma , the lowest explored trend is 1\times 10^{-4}~{}^{\circ} C yr -1 . (c): AMOC trajectory (black curve) for \overline{E_{A}}=0.45 Sv and climate change (+0.01 ∘ C yr -1 ), which is visualised as a classical double well potential ( V(x)=\frac{1}{4}x^{4}-\frac{1}{2}x^{2}+\mu x , with \mu the tuning parameter and x the state variable), with the left minimum the AMOC on state and right minimum the AMOC off state. (d): Similar to panel c, but now with a faster temperature increase (+0.05 ∘ C yr -1 ). (e & f): Similar to panels c & d, but now for \gamma=0.04 Sv per ∘ C. (g & h): Similar to panels c & d, but now for \gamma=0.08 Sv per ∘ C. In all panels, the simulations are initialised from the stable AMOC on state for the given \overline{E_{A}} and without climate change ( \Delta T_{\mathrm{n}}^{a}=0^{\circ} C), and then \Delta T_{\mathrm{n}}^{a} increases linearly up to +5 ∘ C and is kept constant thereafter. We repeat the warming experiment for \overline{E_{A}}=0.45 Sv but now with \gamma>0 , comparing relatively slow (+0.01 ∘ C yr -1 ) and relatively fast (+0.05 ∘ C yr -1 ) warming rates; the temperature anomaly increases up to a maximum of \Delta T_{\mathrm{n}}^{a}=5^{\circ} C. For \gamma=0.04 Sv per ∘ C, the system continues to track the ‘AMOC on’ state under the slow warming case (green curve in Figure 6 a). The equilibrated \Delta T_{\mathrm{n}} is 3.17 ∘ C, resulting in an effective hosing strength of 0.32 Sv, which is indeed below the bifurcation threshold at 0.342 Sv for the given \Delta T_{\mathrm{n}}^{a} . However, in the fast warming case, the AMOC collapses as the system fails to track the ‘AMOC on’ state after crossing the moving saddle-node bifurcation (blue curve in Figure 6 a). To visualize the cases described above, we sketch the stability landscape of the model as a classical double-well potential, with the left and right minima representing the ‘AMOC on’ and ‘AMOC off’ states, respectively (Figures 6 c–h). In this landscape, the system state is remapped based on the distance to the saddle-node bifurcation, where a larger distance means that the AMOC is more stable and is reflected by a lower left minimum. We use the meridional density difference as state variable as it appears better suited to link transient and equilibrium dynamics than the AMOC strength (see Figures S8 c,d). The stability landscape changes for increasing \Delta T_{\mathrm{n}}^{a} and for varying effective freshwater flux forcing (for \gamma>0 ). For \gamma=0 , the AMOC tips faster in the fast warming case (Figure 6 d) compared to the slow warming case (Figure 6 c), but eventually both trajectories reach the same collapsed state which is the only remaining steady state. For \gamma>0 , whether the AMOC collapses or continues to track a stable strong AMOC state depends on the ratios of atmospheric warming (forcing timescale) and ocean warming (adjustment timescale), as well as the interaction strength between thermal and freshwater effects (Figures 6 e–h). Due to the state-dependent forcing, the rate-dependent tipping scenario shown here is more complex than the classical setting of rate-induced tipping in which a single parameter is varied Ashwin et al. 2012 ; Wieczorek et al. 2023 . The maximal temperature trend that does not cause an AMOC collapse (i.e., the critical warming rate), is shown in Figure 6 b as a function of \overline{E_{A}} and \gamma . For sufficiently large temperature-freshwater coupling strengths, the system can always track the stable ‘AMOC on’ state as long as it exists and the imposed forcing rate is slow enough. These results are qualitatively robust for even larger temperature anomalies of up to +10 ∘ C (Figure S8 e), demonstrating that even for extreme warming, forcing pathways exist for which the AMOC does not collapse. Discussion In this study, we argue that AMOC tipping is dependent on the radiative forcing path (i.e., climate change scenario) and not governed by a specific global warming threshold. To demonstrate this, we presented an analysis of a slow CO 2 ramp (+0.5 ppm yr -1 ) simulation using the Community Earth System Model. In this simulation, the GMST anomaly reaches +5.5 ∘ C and the AMOC remains stable throughout, while under the RCP4.5 and RCP8.5 scenario simulations the AMOC starts to collapse at +2.2 ∘ C and +2.8 ∘ C warming van Westen et al. 2025 , respectively. These RCP warming levels are much lower than the maximum warming in the CO 2 ramp simulation, showing that AMOC tipping is affected by the radiative forcing path. This result corroborates early work using idealised climate models Stocker and Schmittner 1997 , while now being demonstrated with a much more comprehensive climate model. Present-day and projected warming simulations using CMIP6 models show substantial AMOC weakening Caesar et al. 2018 ; Weijer et al. 2020 ; Bonan et al. 2025 ; Michel et al. 2025 ; Dijkstra and van Westen 2026 and possibly an AMOC collapse Romanou et al. 2023 ; Drijfhout et al. 2025 . For the CESM under relatively rapid radiative forcing conditions ( \geq RCP4.5 and \overline{F_{H}}=0.45 Sv), the near-surface water is getting lighter much faster than the interior of the Atlantic Ocean, resulting in a shutdown of the adiabatic pathways followed by an AMOC collapse van Westen et al. 2025 . When the imposed radiative forcing is sufficiently slow, both the interior and surface waters adjust their densities accordingly such that they support a relatively strong adiabatic AMOC. With an idealised AMOC model subjected to heat and freshwater forcing, we illustrated how the interplay between the changing forcing and the changing climate state can alter the system’s stability landscape in a way that leads to rate-dependent AMOC tipping. In the CO 2 ramp simulation, the relatively small temperature-induced AMOC weakening is outweighed by two dominant effects. First, higher evaporation rates make the evaporative characteristic of the Atlantic basin more pronounced, which effectively results in a negative freshwater flux forcing. Second, the North Atlantic sea-ice extent is reduced in a warmer climate, which limits the sea-ice insulation effects that influence the AMOC Lin et al. 2023 ; van Westen et al. 2024 . Both effects contribute to a salinifying Atlantic Ocean and, after water mass transformation, the AMOC starts to export these salinity anomalies out of the Atlantic basin at greater depth (1000 – 3500 m). This response at depth is indicative of a more stable AMOC (i.e., larger F_{\mathrm{ovS}} values) and a similar response is found in most CMIP6 models under the extended SSP scenarios. These stabilising mechanisms explain the stable AMOC states which have been found in warmer climates Bonan et al. 2022 ; Romanou et al. 2023 ; Willeit and Ganopolski 2024 ; Hankel 2025 . There are now first indications that the AMOC can collapse under anthropogenic climate change in the latest generation of climate models Romanou et al. 2023 ; Drijfhout et al. 2025 . This implies that the stabilising mechanism in warmer climates, as described above, operates on timescales slower than the current rate of radiative forcing. Hence, it may be still useful to explore warming thresholds for AMOC tipping on relatively short timescales (years to decades), but this becomes less useful if warming rates slow down (as projected beyond the 21st century) and should be complemented with threshold estimates for other relevant forcing parameters. Based on our results, we propose to convert the AMOC tipping warming threshold to a warming rate threshold. Yet, this also poses challenges as the critical warming rate is dependent on the background climate state and the strength of the stabilising mechanism. Therefore, it is more useful to analyse physics-based indicators, such as \Psi_{\mathrm{NADW}} and B_{\mathrm{flux}} , that capture the effects of changing forcing conditions (both stabilising and destabilising). The distance to the tipping point could be estimated using B_{\mathrm{flux}} , with a B_{\mathrm{flux}} sign change marking the onset of an AMOC collapse van Westen et al. 2025 . According to the B_{\mathrm{flux}} indicator, as determined in CMIP6 models, the AMOC starts to collapse around +2.5 ∘ C warming which could be reached by around 2060 van Westen et al. 2025 . This is substantially lower than the previous threshold of +4 ∘ C warming Armstrong McKay et al. 2022 that lacks a time horizon. Observations show a present-day warming at the +1.5 ∘ C level (2023 – 2025, C3S/ECMWF) and, if the AMOC were to collapse at +2.5 ∘ C warming in 2060, this would imply a critical warming rate of +0.29 ∘ C per decade for AMOC tipping. This rate aligns with projected warming rates of +0.27 ∘ C per decade to +0.36 ∘ C per decade Hansen et al. 2023 , which at least suggests that rate-dependent effects are highly relevant for the fate of the AMOC. It is therefore critically important to reduce the rate of radiative forcing as quickly as possible to limit the risk of AMOC tipping. Methods Climate Model Simulations The CESM is a fully-coupled climate model and the simulations here have a 1 ∘ horizontal resolution for the ocean/sea-ice components and a 2 ∘ horizontal resolution for the atmosphere/land components. The ocean component is the Parallel Ocean Program version 2 (POP2, Smith et al. 2010 ), the atmospheric component is the Community Atmosphere Model version 4 (CAM4, Neale et al. 2013 ), and the sea-ice component is the Community Ice Code version 4 (CICE4, Hunke and Lipscomb 2008 ). The CESM is either forced under varying freshwater flux forcing and fixed pre-industrial radiative forcing conditions or fixed freshwater flux forcing and varying radiative forcing conditions. For more details on the precise CESM set-up, we refer to previous work van Westen and Dijkstra 2023 ; van Westen et al. 2024 ; van Westen et al. 2024 ; van Westen et al. 2025 and the main text. In addition to the CESM, model output from CMIP phase 6 (indicated here as CMIP6 models) were analysed. We retained the historical forcing (1850 – 2014) followed by the extended SSP1-2.6 scenario (2015 – 2300). Only for the GISS-E2-1-G model, the scenarios were extended to 2500 and are also available for the extended SSP2-4.5 scenario. Note that the forcing scenarios are slightly different between the CESM simulations (the RCPs) and CMIP6 (the SSPs). Water Mass Transformation (WMT) We use the same procedure and naming convention as described in van Westen et al. 2025 , which are briefly repeated here for completeness. We consider the volume ( V_{\sigma} ) conservation of a fixed horizontal domain and that is bounded by the ocean surface and isopycnal Groeskamp et al. 2019 : \frac{\partial V_{\sigma}}{\partial t}=M_{\sigma}-G_{\sigma}, (1) with M_{\sigma} is the advective transport convergence through the open boundaries of the domain and G_{\sigma} is the diapycnal transformation rate taken place in the domain. The WMT rates from the surface are dominant in G_{\sigma} and are indicated by \Psi_{\mathrm{surf}} : \Psi_{\mathrm{surf}}(y,\sigma_{2})=\frac{1}{\Delta\sigma_{2}}\int_{x_{W}}^{x_{E}}\int_{y}^{y_{N}}-\frac{\rho_{0}}{g}B_{\mathrm{flux}}(x,y)~\Pi(\sigma_{2})~\mathrm{d}y^{\prime}\mathrm{d}x^{\prime} (2) where: \Pi(\sigma_{2})=\begin{cases}1~\hskip 14.22636pt~\text{if}~\sigma_{2}-\frac{\Delta\sigma_{2}}{2}\leq\sigma_{2}<\sigma_{2}+\frac{\Delta\sigma_{2}}{2}\\ 0~\hskip 14.22636pt~\text{elsewhere}\end{cases} (3) and we use a potential density bin size of \Delta\sigma_{2}=0.05 kg m -3 . The quantity B_{\mathrm{flux}} is the surface buoyancy flux: B_{\mathrm{flux}}(x,y)=\frac{g\alpha}{\rho_{0}C_{p}}Q_{\mathrm{heat}}+\frac{g\beta S_{\mathrm{surf}}}{\rho_{0}}Q_{\mathrm{fresh}}=B_{\mathrm{flux}}^{T}+B_{\mathrm{flux}}^{S} (4) In the relations above, \alpha is the thermal expansion coefficient, C_{p} the specific heat capacity, Q_{\mathrm{heat}} the net heat flux into the ocean, \beta the haline contraction coefficient, S_{\mathrm{surf}} the sea surface salinity, Q_{\mathrm{fresh}} the net freshwater flux into the ocean, g (= 9.8 m s -2 ) is the gravitational acceleration and \rho_{0} (= 1027 kg m -3 ) a reference density. For each quantity, we used its local value and for \alpha , \beta and C_{p} we used the Thermodynamic Equation of SeaWater 2010 (TEOS-10) toolkit McDougall and Barker 2011 . All the analyses are conducted on monthly-averaged fields (due to strong seasonal cycle) and the related time series are subsequently averaged to yearly values. The Atlantic Freshwater Budget The freshwater budget over the Atlantic Ocean (34 ∘ S – 65 ∘ N) is defined as: \displaystyle\frac{\mathrm{d}\overline{W}}{\mathrm{d}t} \displaystyle=F_{\mathrm{con}}+F_{\mathrm{surf}}+F_{\mathrm{mix}} (5a) \displaystyle\overline{W} \displaystyle=-\frac{1}{S_{0}}\int_{-H}^{0}\int_{34^{\circ}\mathrm{S}}^{65^{\circ}\mathrm{N}}\int_{x_{W}}^{x_{E}}(S-S_{0})\mathrm{d}x\mathrm{d}y\mathrm{d}z (5b) where \overline{W} is the freshwater content, F_{\mathrm{con}} the freshwater convergence and is determined as the total freshwater transport through the three boundaries (i.e., 34 ∘ S, 65 ∘ N and the Strait of Gibraltar), F_{\mathrm{surf}} the surface freshwater flux and F_{\mathrm{mix}} is a residual term which closes the budget and captures for example diffusion Jüling et al. 2021 . The F_{\mathrm{con}} is primarily governed by the total meridional freshwater transports at 34 ∘ S and 65 ∘ N, where the latter is defined by: F_{\mathrm{tot}}(y)=-\frac{1}{S_{0}}\int_{-H}^{0}\int_{x_{W}}^{x_{E}}v(S-S_{0})\mathrm{d}x\mathrm{d}z (6) and can be further decomposed in an overturning ( F_{\mathrm{ov}} ), azonal gyre ( F_{\mathrm{az}} ), barotropic ( F_{\mathrm{bt}}\approx 0 ), and eddy ( F_{\mathrm{eddy}} ) contribution: \displaystyle F_{\mathrm{ov}}(y) \displaystyle= \displaystyle-\frac{1}{S_{0}}\int_{-H}^{0}\left[\int_{x_{W}}^{x_{E}}v^{*}\mathrm{d}x\right]\left[\langle S\rangle-S_{0}\right]\mathrm{d}z, (7) \displaystyle F_{\mathrm{az}}(y) \displaystyle= \displaystyle-\frac{1}{S_{0}}\int_{-H}^{0}\int_{x_{W}}^{x_{E}}v^{\prime}S^{\prime}\mathrm{d}x\mathrm{d}z, (8) \displaystyle F_{\mathrm{bt}}(y) \displaystyle= \displaystyle-\frac{1}{S_{0}}\int_{-H}^{0}\int_{x_{W}}^{x_{E}}\hat{v}\left(\hat{S}-S_{0}\right)\mathrm{d}x\mathrm{d}z, (9) \displaystyle F_{\mathrm{eddy}}(y) \displaystyle= \displaystyle-\frac{1}{S_{0}}\int_{-H}^{0}\int_{x_{W}}^{x_{E}}\tilde{v}\left(\tilde{S}-S_{0}\right)\mathrm{d}x\mathrm{d}z. (10) Here, v^{*} is defined as v^{*}=v-\hat{v} , where v is the meridional velocity and \hat{v} ( \hat{S} ) is the section spatially-averaged meridional velocity (salinity). The quantity \langle S\rangle indicates the zonally-averaged salinity and primed quantities ( v^{\prime} and S^{\prime} ) are deviations from their respective zonal averages. The \tilde{v} and \tilde{S} are the eddy terms for velocity and salinity, respectively, with more details provided in Jüling et al. 2021 . Declarations • Funding – R.M.v.W. and H.A.D. are funded by the European Research Council through the ERC-AdG project TAOC (project 101055096). R.B. is funded by the ClimTip project, which has received funding from the European Union’s Horizon Europe research and innovation programme under grant agreement No. 101137601. • Conflict of interest – The authors declare no competing interest • Ethics approval – Not applicable • Availability of data and materials – The (processed) model output will be made available upon publication. • Code availability – The analysis scripts will be made available upon publication. • Authors’ contributions – R.M.v.W., R.B. and H.A.D. conceived the idea for this study. R.M.v.W. conducted the analysis and prepared all figures. All authors were actively involved in the interpretation of the analysis results and the writing process. Supplementary Figures Figure S1: Atlantic Ocean potential densities. The potential density in the North Atlantic Ocean for the surface ( \sigma_{2}^{\mathrm{surf}} , 40 ∘ N – 65 ∘ N, yellow curve) and at the maximum AMOC strength at 40 ∘ N ( \sigma_{2}^{\mathrm{max}} , red curve). The yellow shading represents the variations in surface potential density, where for each month the lightest and heaviest potential densities (over 40 ∘ N – 65 ∘ N) are retained and are subsequently converted to yearly averages. The inset shows the surface area where the monthly surface potential density are heavier than \sigma_{2}^{\mathrm{max}} , which is subsequently converted to yearly averages. All time series are smoothed through a 25-year running mean to reduce the variability. The results are for the a) CO 2 ramp, b) historical and RCP2.6, c) historical and RCP4.5, and d) historical and RCP8.5 simulations, all simulations have \overline{F_{H}}=0.45 Sv. Figure S2: Surface buoyancy flux decomposition. The surface buoyancy flux differences (compared to the first 50 model years) between 40 ∘ N – 65 ∘ N for the CO 2 ramp simulation, decomposed into the different heat (panel a) and freshwater (panel b) fluxes. The inset in panel b shows the yearly-averaged sea-ice area (grid cells with sea-ice fractions of at least 15%) between 40 ∘ N – 65 ∘ N. The time series are smoothed through a 25-year running mean to reduce the variability, note the different vertical ranges for the panels. Figure S3: Mixed layer depth and sea-ice extent. (a): The March mixed layer depth and sea-ice extent (i.e., 15% sea-ice fraction contour) for model years 1 – 50 of the CO 2 ramp simulation. (b & c): Similar to panel a, but now for b) model years 701 – 750 and c) 1701 – 1750. (d): The spatially-averaged March mixed layer depth and sea-ice fraction over the Labrador basin (black outlined region in panel a). The thin curves are yearly averages, whereas the thick curves are smoothed versions (25-year moving averages). Figure S4: Freshwater transport decomposition. (a & b): The freshwater transports at 34 ∘ S and 65 ∘ N for the CO 2 ramp simulation. (c): The F_{\mathrm{ovS}} decomposition, following the procedure outlined in Van Westen and Dijkstra 2024 . (d): The F_{\mathrm{ovS}} variance, where the variance is determined over 50-year sliding windows. A linear trend is removed over the 50-year window before determining the variance. In panels a,b,c, the thin curves are yearly averages, whereas the thick curves are smoothed versions (25-year moving averages). Figure S5: Climate model simulations with the CESM (a): The AMOC strength (at 26 ∘ N and 1,000 m depth) for the historical and RCP4.5 with \overline{F_{H}}=0.18 Sv, and historical and RCP2.6 with \overline{F_{H}}=0.45 Sv. (b): Similar to panel a, but now for the GMST anomaly compared to 1850 – 1899. (c): Similar to Figure 4 d, but now for the historical and RCP4.5 with \overline{F_{H}}=0.18 Sv. (d): Similar to Figure 4 d, but now for the historical and RCP2.6 with \overline{F_{H}}=0.45 Sv. In panels a,b, the thin curves are yearly averages, whereas the thick curves are smoothed versions (25-year moving averages). Figure S6: Climate model simulations with the GISS-E2-1-G (r2i1p1f2) Similar to Figure S5 , but now for the GISS-E2-1-G (r2i1p1f2) under the historical and SSP1-2.6 and Historical and SSP2-4.5. Figure S7: Climate model simulations under the historical and SSP1-2.6 Similar to Figure S5 , but now for 7 different CMIP6 models under the historical and SSP1-2.6. The specific realisations are indicated in the captions of panels c through h. Figure S8: The idealised 5-box AMOC model (a): The idealised 5-box AMOC model, with an AMOC on state (clockwise circulation, dashed red arrows) and AMOC off state (anticlockwise circulation, dotted red arrows), in which temperatures, salinities and pycnocline depth are dynamically resolved. The model has the asymmetric freshwater flux forcing, E_{A} , as control parameter. In addition, the model is forced under a linearly increasing atmospheric temperature anomaly for T_{\mathrm{n}}^{a} (and also T_{\mathrm{s}}^{a} as they are coupled) and the temperature anomaly is then hold constant. More details, parameter settings and sensitivity experiments are provided in van Westen et al. 2024 ; van Westen et al. 2025 , the version analysed here does not consider sea-ice insulation effects. (b): The different steady states under varying E_{A} ( \Delta T_{\mathrm{n}}^{a}=0^{\circ} C), including the steady states under climate change ( \Delta T_{\mathrm{n}}^{a}=5^{\circ} C). (c & d): The evolution of the steady states and the system for the AMOC strength and the meridional density difference between boxes n and ts. Here, the model is initialised from the steady state of \overline{E_{A}}=0.33 Sv (and with \gamma=0 ) and is forced under an atmospheric temperature trend of +0.125 ∘ C per year (up to +5 ∘ C). The steady states are determined for every \Delta T_{\mathrm{n}}^{a}=0.1^{\circ} C. (e): Similar to Figure 6 b, but now up to \Delta T_{\mathrm{n}}^{a}=10^{\circ} C.