Model sensitivity to North Atlantic freshwater forcing at 8.2 ka

We compared four simulations of the 8.2 ka event to assess climate model sensitivity and skill in responding to North Atlantic freshwater perturbations. All of the simulations used the same freshwater forcing, 2.5 Sv for one year, applied to either the Hudson Bay (northeastern Canada) or Labrador Sea (between Canada’s Labrador coast and Greenland). This freshwater pulse induced a decadal-mean slowdown of 10–25% in the AtlanticMeridional Overturning Circulation (AMOC) of the models and caused a large-scale pattern of climate anomalies that matched proxy evidence for cooling in the Northern Hemisphere and a southward shift of the Intertropical Convergence Zone. The multi-model ensemble generated temperature anomalies that were just half as large as those from quantitative proxy reconstructions, however. Also, the duration of AMOC and climate anomalies in three of the simulations was only several decades, significantly shorter than the duration of ∼150 yr in the paleoclimate record. Possible reasons for these discrepancies include incorrect representation of the early Holocene climate and ocean state in the North Atlantic and uncertainties in the freshwater forcing estimates.


Introduction
The Atlantic Meridional Overturning Circulation (AMOC) plays a key role in the climate system, particularly through its control on heat transport and storage of carbon in the deep ocean. Changes in the AMOC can have far-reaching effects on the El Niño-Southern Oscillation (Timmermann et al., 2005), Atlantic hurricane development (Zhang and Delworth, 2006), tropical rainfall (Vellinga and Wood, 2002), and marine ecosystems (Schmittner, 2005). Model simulations of the 21st century with prescribed greenhouse gas concentrations increasing according to the Intergovernmental Panel on Climate Change (IPCC) scenario SRESA1B uniformly show a reduction in the strength of the AMOC . This multi-model ensemble yields a mean decrease of 25 % by 2100, but there is a large range in the individual model results that indicates substantial uncertainties in the AMOC response to climate change.
Several previous model intercomparison projects were undertaken to improve understanding of the large spread in modeled AMOC. Schmittner et al. (2005) considered the skill of nine coupled climate models in matching observations of modern hydrography. They found that the models were more successful at reproducing temperature patterns than either salinity patterns or pycnocline depth. Stouffer et al. (2006) examined the response of both Earth system models of intermediate complexity (EMICs) and coupled atmosphere-ocean general circulation models (AOGCMs) to North Atlantic freshwater forcings of 0.1 and 1.0 Sv (Sverdrup = 10 6 m 3 s −1 ) for 100 yr. While there were some robust patterns among the models, important disagreements existed in model sensitivity and in reversibility following AMOC shutdown. Since these were idealized experiments, no comparison to observations was possible. Otto-Bliesner et al. (2007) Wagner et al. (2013) GISS ModelE-R ModelE: M20 (4×5 • ), 20 levels Russell: 4 × 5 • , 13 levels, mass-conserving Schmidt et al. (2006) Russell et al. (2000Russell et al. ( , 1995 LeGrande et al. (2006) LeGrande and Schmidt (2008) LOVECLIM1.2 ECBilt2: T21 (5.625 × 5.625 • ), 3 levels CLIO3: 3 × 3 • , 20 levels, mass-conserving Goosse et al. (2010) paleoclimate proxy evidence indicated serious mismatches for several of the simulations. For the third phase of PMIP, the 8.2 ka event has been targeted for a new model intercomparison. Of past abrupt changes in the AMOC, the 8.2 ka event provides a particularly useful case study because its duration (∼150 yr; Thomas et al., 2007) and forcing are constrained by the proxy record, making an achievable target for climate model simulations (Schmidt and LeGrande, 2005). There are still some uncertainties regarding the hypothesized forcing of the event, including the volume of drainage from proglacial Lake Agassiz-Ojibway (hereafter Lake Agassiz; Barber et al., 1999) into the Hudson Bay (northeastern Canada) and the possibility of multiple meltwater pulses from both the lake and the collapsing Laurentide Ice Sheet (Teller et al., 2002;Gregoire et al., 2012). Model sensitivity to some of these uncertainties has been explored elsewhere (Renssen et al., 2001;Wiersma et al., 2006;LeGrande and Schmidt, 2008;Clarke et al., 2009;Wiersma and Jongma, 2010;Wagner et al., 2013). The target of this intercomparison is to use a median value for the forcing of the 8.2 ka event and compare model sensitivity to North Atlantic surface buoyancy anomalies that have precise dating and a duration short enough to make simulations with state-of-the-art coupled climate models feasible (Schmidt and LeGrande, 2005;Thomas et al., 2007;Kobashi et al., 2007).

Models and experiments
We compare 8.2 ka experiments completed with three models: the Community Climate System Model version 3 (CCSM), the Goddard Institute for Space Studies (GISS) ModelE-R and LOVECLIM version 1.2. CCSM and ModelE-R are atmosphere-ocean general circulation models (AOGCMs) coupled without flux adjustments. LOVECLIM is an Earth system model of intermediate complexity with its most significant simplifications applied to the atmosphere component (Table 1). These simplifications include clouds that are prescribed and vertical profiles of temperature and specific humidity that are limited by three atmospheric levels. LOVECLIM also employs a freshwater flux correction between the atmosphere and ocean subcomponents that removes excess precipitation from the Arctic and Atlantic and adds it to the North Pacific (Goosse et al., 2010).
Of relevance to this study, the ocean models of ModelE-R and LOVECLIM are mass-conserving, in which the addition of freshwater causes a rise in the free surface of the ocean and reduces salinity purely through dilution. The ocean model component of CCSM uses the rigid-lid approximation, which does not permit vertical motion at the top of the ocean and parameterizes the addition of freshwater as a salt extraction while keeping the volume of the ocean constant. Yin et al. (2009) discuss the differences between these two approaches and compare results from two versions of the GFDL CM2.1 model using each formulation. For a large freshwater forcing that is similar in magnitude to that used in 8.2 ka experiments, the rigid-lid version exaggerates the forcing and there are significant regional biases in sea surface salinity (SSS). Despite this, the AMOC behaves similarly in the two versions and many fundamental aspects of the two simulations are qualitatively similar.
Boundary conditions specified for the control simulations are listed in Table 2. Early Holocene orbital forcing increased the seasonality of insolation in the Northern Hemisphere and decreased seasonality in the Southern Hemisphere relative to the present (Berger, 1978). Greenhouse gas concentrations for the early Holocene were nearly identical to those for the recent pre-industrial period (Flückiger et al., 2002;Monnin et al., 2004). Two of the control simulations, CCSM all and LOVECLIM, incorporated the surface albedo and elevation effects of the remnant of the Laurentide Ice Sheet that was present near Hudson Bay at 8.5 ka, as reconstructed by Peltier (2004). These same control simulations also included a small (∼0.05 Sv) background flux of Laurentide meltwater (Licciardi et al., 1999). In CCSM all , this freshwater flux was added to the modeled St. Lawrence River (Canadian/US Great Lakes Basin) at its outflow, and was spread as a virtual salinity flux along the coast near the river's mouth. In LOVECLIM, the freshwater was added  as a volume to the upper layer of the ocean at the Hudson Strait. Since the ocean model in LOVECLIM has a free surface, this effectively means that the surface height was raised. The temperature of the added freshwater in LOVE-CLIM was assigned the same temperature as the water in the ocean cell to which it was added. Both of these control simulations with background meltwater flux were integrated until reaching a quasi-equilibrium, in which SSS of the North Atlantic had stabilized. Global mean ocean salinity decreases slowly throughout these control simulations due to the background meltwater flux, a trend that parallels observed freshening since the Last Glacial Maximum (Adkins et al., 2002). A second CCSM control simulation (CCSM og; OG = orbital and greenhouse gas only) without a Laurentide Ice Sheet and background meltwater flux is included in this study for a more direct comparison to ModelE-R results. For the 8.2 ka event experiments, a meltwater pulse (MWP) of 2.5 Sv for 1 yr was added to each of the control simulations to represent the previously-mentioned drainage of Lake Agassiz. This freshwater volume was the best estimate for the drainage event based on flood hydrograph simulations (Clarke et al., 2004). Following the one-year perturbation, the MWP ceased and the climate was allowed to recover. In the models with a free-surface ocean, the MWP was added as a volume to a limited number of grid cells. In ModelE-R, freshwater was added to the approximately 20 grid boxes in the Hudson Bay and was assigned a temperature of 0 • C. In LOVECLIM, freshwater was added to the upper layers of the ocean at the Hudson Strait and was assigned the same temperature as the water in the ocean cell to which it was added. The virtual salinity flux in CCSM required a larger area for the MWP (50-65 • N, 35-70 • W).
The control simulation for ModelE-R displayed a number of transient, quasi-stable states with either strong or weak AMOC (LeGrande et al., 2006;LeGrande and Schmidt, 2008). For this study, we use an experiment begun from a period of weak AMOC. The weak case was chosen because it exhibited the longest response to the 2.5 Sv × 1 yr forcing, and because it may more closely emulate the early Holocene than periods with strong AMOC since it lacks deep convection in the Labrador Sea between Canada's Labrador coast and Greenland (see LeGrande and Schmidt, 2008 for further detail). Since the weak case exhibits some high amplitude decadal variability, we examined "decadal" results for this model (i.e., the 10-yr mean of the MWP experiment less the 30-yr mean of the closest control years) in order to more clearly show the transient response to the MWP.
We calculated the Student's t test for the differences between control and MWP experiments for the two models with annual output (i.e., LOVECLIM and CCSM). Since annual output is no longer available from the ModelE-R experiments and resources do not exist to re-run these simulations, we are unable to make statements about the statistical significance of the model's response to the MWP, about the skill of the model on the annual time-scale and about the relative amplitude of decadal and annual variability. The model simulations in the present study are an "ensemble of opportunity", meaning that most were completed before this intercomparison was planned. One limitation of not coordinating experiments is that perfect comparisons are sometimes impossible to make. However, the main conclusions of this paper would not change were annual data available from all simulations and ensembles of opportunity are important for informing future coordinated intercomparisons.  Table 2), in Sv (1 Sv = 10 6 m 3 s −1 ). Plotted values are 200 yr means except for CCSM og , which is a 150 yr mean. Values in parentheses following the model names are long-term means for the maximum of the streamfunction below 500 m water depth.
with available annual output (standard deviations: LOVE-CLIM = 0.7, CCSM og = 1.1, CCSM all = 0.9 Sv). AMOC intensity is lower by several Sv in the simulations with a background meltwater flux. AMOC has a similar structure in all the control simulations. The northward flow of warm, salty water occurs in the upper 1000 m, while the southward return flow of North Atlantic Deep Water occurs between 1000-3000 m. The anticlockwise cell in the deep ocean, associated with Antarctic Bottom Water formation, has a strength of about 4 Sv in all control simulations.
The values of AMOC intensity in the control simulations are generally similar to the strength of the modern-day AMOC (Meehl et al., 2007). Proxy evidence suggests that the strength of the AMOC during the early Holocene was probably not that different from today (Bianchi and McCave, 1999;Hall et al., 2004;Oppo et al., 2003;McManus et al., 2004;Praetorius et al., 2008). There is some proxy evidence for lack of convection and deep water formation in the Labrador Sea during the early Holocene, however (e.g., Hillaire-Marcel et al., 2001;Solignac et al., 2004;Fagel et al., 2004). To reconcile a vigorous AMOC with lack of Labrador Sea convection, some other convection area, perhaps in the Irminger Basin, might have been stronger in the early Holocene to offset the weaker Labrador Sea convection (Hall et al., 2010).
The location and strength of convection areas in the North Atlantic varies significantly among the control simulations (Fig. 2). Convection occurs primarily in the Nordic Seas in one of the models (LOVECLIM), primarily in the Irminger Sea in another (ModelE-R), and in both the Nordic Seas and just east of the Labrador Sea in the third model (CCSM). Notably, the background meltwater flux of 0.05 Sv does not shut down convection just east of the Labrador Sea in the CCSM all control simulation (Fig. 2), as that flux is routed to the south of the Labrador Sea by ocean surface currents.
Following the 2.5 Sv MWP for one year, AMOC intensity decreases in all simulations (Fig. 3). The maximum decadalmean decline in LOVECLIM and CCSM is about 10 %,  while for ModelE-R it is about 25 %. The decline in AMOC intensity in LOVECLIM and CCSM is relatively short-lived, on the order of several decades, and generally within the range of natural variability of AMOC in their control simulations. Similarly, mixed-layer depths shoal significantly following the MWP, but this weakening of convection also lasts several decades or less (not shown). The response in ModelE-R is more pronounced and longer-lived, extending on the order of 100-120 yr. Proxy records do not provide a quantitative estimate of AMOC weakening at 8.2 ka, but do suggest a duration of 100-200 yr (Ellison et al., 2006;Kleiven et al., 2008).

Ocean salinity and temperature
Significant freshening of the North Atlantic occurs following the MWP in all simulations (Fig. 4). The largest anomalies are generally along the coast of Labrador and are up to 1 psu when averaged over the first fifty years following the MWP. Areas of positive SSS anomalies at the mouth of the St. Lawrence River in CCSM all are caused by cessation of the 0.05 Sv background meltwater flux once Lake Agassiz has drained. Globally, negative anomalies greater than 0.2 psu are confined to the North Atlantic and Arctic oceans (not shown).
Patterns of SSS anomalies suggest that freshwater travels eastward from the Labrador Sea into the North Atlantic in all simulations. For most of the simulations, SSS decreases in both the Greenland-Iceland-Norwegian Seas and in the subtropical gyre. This pathway is different from that inferred by Keigwin et al. (2005), who used δ 18 O of planktic foraminifera to suggest salinity was decreased near Cape Hatteras around 8.2 ka. Also, it has been argued that freshwater released from Hudson Strait would not reach the Nordic Seas, instead being trapped along the North American coast (e.g., Wunsch, 2010) or circulating in the subtropical gyre   5. Anomalies of annual-mean sea surface temperature in the first fifty years following the MWP relative to the control simulation, in degrees Celsius. Stippling shows statistical significance at the 95 % level according to a Student's t test. Statistical tests were not performed for ModelE-R since only decadal averages were available. (Condron and Winsor, 2011). However, several proxy records from the Irminger and Labrador Seas combine δ 18 O and Mg/Ca of planktic foraminera to infer decreases in δ 18 O of seawater at 8.2 ka of up to 1 ‰ (Came et al., 2007;Thornalley et al., 2009;Ellison et al., 2006;Winsor et al., 2012;Hoffman et al., 2012), which would be equivalent to a freshening of ∼0.7 psu assuming the Laurentide Ice Sheet meltwater was about −25 ‰ (Hillaire- Marcel et al., 2007). Also, the location of detrital carbonate layers deposited around 8.2 ka indicate greater freshwater transport in the outer branch of the Labrador Current, which typically mixes with the North Atlantic Current and travels to the Nordic Seas (Lewis et al., 2012). The model simulations presented here, as well as others published by Born and Levermann (2010) and Spence et al. (2008), tend to support some amount of freshwater transport into the Nordic Seas.
Likewise, sea surface cooling is concentrated in the North Atlantic in all simulations (Fig. 5). Mean anomalies across the North Atlantic for the first fifty years following the MWP are on the order of 1 • C, though they exceed 2 • C locally in the CCSM and ModelE-R experiments. Maximum anomalies in the LOVECLIM simulation are on the order of ∼0.5 • C and are located in the far North Atlantic. ModelE-R shows cooling on the order of several tenths of a degree Celsius across most of the Southern Hemisphere. The other simulations show little significant change south of 30 • N with the exception of CCSM all , which has some significant warming in the south Atlantic.

Barotropic streamfunction
A common model diagnostic for the ocean circulation, including the strengths of the subtropical and subpolar gyres in the North Atlantic, is the vertically-integrated mass transport (barotropic) streamfunction. Values for this quantity are available for three of the simulations (Fig. 6). In these three simulations, transports in both the subtropical gyre and the subpolar gyre weaken for a few decades following freshwater forcing. This result is consistent with the concept that reduction of deep convection in the core of the subpolar gyre, as occurs briefly in these simulations in the Labrador and/or Irminger Seas, weakens this circulation (e.g., Häkkinen and Rhines, 2004). The barotropic streamfunction is not available as standard output for the fourth simulation (LOVECLIM) and is not easily calculated offline. Neither deep convection (Sect. 3.1) nor upper-ocean velocities (top 100 m, not shown) in this model show a large or long-term change forced by the MWP, though, suggesting that the response of the gyres is at least qualitatively similar to the other models.
On the other hand, Born and Levermann (2010) found a prolonged strengthening of the subpolar gyre circulation in a simulation of the 8.2 ka event with the CLIMBER-3α model. In this model, a reduction of deepwater formation in the Nordic Seas intensified the subpolar gyre and triggered internal feedbacks to increase and maintain deep convection

Sea ice
All of the simulations have areas of significantly expanded sea ice following freshwater forcing, particularly in the Labrador Sea and in the Norwegian and/or Barents Sea (Fig. 7). Generally, these changes for the first fifty years following the MWP are on the order of 5-10 %, although they can be as large as 20-25 % in some areas. Sea ice changes in the Southern Ocean have a heterogeneous spatial pattern and generally are not statistically significant.

Surface air temperature
The North Atlantic region and the Arctic become significantly colder in most simulations during the first fifty years following the MWP, with mean annual temperatures in the multi-model ensemble decreasing less than ∼0.5 • C over Europe and ∼1.0 • C over Greenland (Fig. 8) . Temperature changes are minimal in the tropics and the Southern Hemisphere. This spatial pattern agrees well with proxy records, which clearly indicate colder conditions across the Northern Hemisphere during the 8.2 ka event but suggest that any Southern Hemisphere temperature changes were likely regional (Fig. 8).
The magnitude of circum-North Atlantic temperature changes inferred from proxies is somewhat larger than those in the models. Temperature reconstructions from pollen and δ 18 O in Europe consistently show anomalies of about −1.1 to −1.2 • C in mean annual temperature during the 8.2 ka event, although standard errors of these reconstructions are nearly as large as the anomalies themselves (Veski et al., 2004;von Grafenstein et al., 1998;Sarmaja-Korjonen and Seppä, 2007;Feurdean et al., 2008). Nitrogen isotopes from Greenland indicate temperatures decreased about 2.2 • C averaged over the duration of the event, with an even larger decrease of 3.3 • C during the most extreme 60 yr period (Kobashi et al., 2007).
Anomalies over the North Atlantic in the LOVECLIM and CCSM experiments are short-lived; generally, temperature values are outside the range of natural variability (defined as the mean ±2 standard deviations of the control) for less than two decades (Fig. 9). Anomalies are longer-lived in the ModelE-R simulation, lasting on the order of 100 yr. These longer-lived anomalies are a better match to high-resolution proxy records from Europe and Greenland, which consistently show an event duration of 100 to 150 yr . Stippling shows grid cells where at least three of the simulations agree on the sign of the temperature anomaly. (Bottom) Qualitative and quantitative mean-annual temperature anomalies relative to the early Holocene background climate, in degrees Celsius, inferred from proxy records for the 8.2 ka event, as summarized by Morrill et al. (2013).

Precipitation
Despite the noise inherent in precipitation, a number of features are common among the model simulations for the fifty years following the MWP. In all cases, the most important changes are a reduction in precipitation over the North Atlantic and Northern Hemisphere tropics, and an increase in precipitation over the Southern Hemisphere tropics (Fig. 10). The tropical pattern, consistent with a southward shift of the mean position of the Intertropical Convergence Zone, is clearest over the Atlantic Ocean (Fig. 11). Tropical proxy records from both speleothem δ 18 O measurements and indicators of lake water balance support this spatial pattern (Fig. 10).
Several quantitative estimates of drying exist from proxies in high northern latitudes; these include an ∼8 % reduction in accumulation in central Greenland ice cores and an ∼17 % reduction in rainfall inferred from pollen north of the Mediterranean, although again the standard errors of these reconstructions are nearly as large as the anomalies themselves (Feurdean et al., 2008;Pross et al., 2009;Hammer et , 1997;Rasmussen et al., 2007). The model simulations generally match the magnitude of drying in central Greenland, but typically do not match either the direction or magnitude of change in southeastern Europe. Additionally, evidence for wetter conditions at 8.2 ka from pollen and lake geochemical records in northern Europe is not matched by the freshwater experiments (Fig. 10).

Discussion and conclusions
To summarize, the models generally do a good job in reproducing large-scale patterns of temperature and precipitation changes at 8.2 ka inferred from proxy records. These patterns include cooling across most of the Northern Hemisphere and a southward shift of the Intertropical Convergence Zone. The models have less success in matching the magnitude and duration of climate anomalies. Temperature changes in the multi-model ensemble are about half the size of those of quantitative proxy records from Europe and Greenland. For all but one of the simulations, the duration of the 8.2 ka climate anomalies is on the order of several decades rather than the ∼150 yr observed in proxy records. Also, there are discrepancies between model and data for some regional-scale anomaly patterns, including precipitation changes in Europe. These patterns are less well-constrained by proxy evidence, however.
The background climate state of the early Holocene, and the location of convection areas in the North Atlantic more specifically, might explain some of the differences we see between models and proxy data. The ModelE-R simulation has the best match to proxies for event duration, and it has been previously demonstrated for this model that the lack of Labrador Sea convection is essential for this response (LeGrande and Schmidt, 2008;LeGrande et al., 2006). Previous work with the ECBilt-CLIO model also supports this interpretation; when Labrador Sea convection is weakened by the background meltwater flux, the ocean's ability to transport freshwater anomalies away from the North Atlantic is diminished and the response to freshwater forcing is prolonged (Wiersma et al., 2006). On the other hand, lack of convection in the Labrador Sea does not lead to a long-lived climate response in this model's successor, LOVECLIM. While the exact reasons for this have yet to be determined, the background meltwater flux used in the LOVECLIM experiment is less than that in the ECBilt-CLIO experiments (0.05 vs. 0.17 Sv; Wiersma et al., 2006;Li et al., 2009), and it seems that LOVECLIM is also less sensitive to freshwater perturbations than its predecessor. Also adding uncertainty to the importance of convection strength in the Labrador Sea, proxies indicate that AMOC strength was not too different from today during the early Holocene. In this case, some other convection area, perhaps in the Irminger Basin (between Greenland (Bottom) Anomalies of Atlantic annual-mean precipitation for the first fifty years following the MWP relative to the control simulation, in cm yr −1 . and Iceland), might have been stronger in the early Holocene to offset the weaker Labrador Sea convection (Hall et al., 2010). If this was true, the strengthened convection areas elsewhere might be able to compensate for decreased freshwater divergence in the Labrador Sea.
Another factor in the model-data mismatch could be the size or the complexity of the MWP. The model simulations were forced with 2.5 Sv for one year, which was the best estimate of the flood hydrograph simulations of Clarke et al. (2004). As these authors point out, though, the total volume of Lake Agassiz could have generated twice this forcing and more complex multipulse patterns are possible (Teller et al., 2002). Their flood model generates a stable drainage channel that prohibits complete drainage, but this result might be unlikely for an outburst flood from Lake Agassiz. In addition, uncertainties in the reconstructed position of the ice-margin on the northern side of Lake Agassiz translate into a range of possible lake volumes spanning 45-200 % of the best estimate (Tornqvist and Hijma, 2012).
Reconstructions of sea level rise at 8.2 ka support the idea of a larger freshwater drainage. Using peat deposits from the Mississippi River delta (US), Li et al. (2012) reconstructed a total eustatic sea level rise of 0.8 to 2.2 m at 8.2 ka. Another reconstruction from the Rhine-Meuse delta (the Netherlands and Belgium) implies a sea level rise of 3.0 ± 1.2 m (Hijma and Cohen, 2010). These are significantly larger than the forcing of 2.5 Sv for one year (∼0.2 m sea level equivalent) or even than the best estimate of the entire volume of Lake Agassiz (∼0.4 m sea level equivalent). Recent model simulations suggest that the collapse of the Laurentide ice-sheet saddle around 8.2 ka provided this larger volume of freshwater (Gregoire et al., 2012), and that this forcing results in a cooling event that matches many proxy records (Wiersma and Jongma, 2010;Wagner et al., 2013).
The difference in boundary conditions between the control simulations does not obviously account for divergent model responses. As shown in the comparison of the two CCSM simulations, CCSM og and CCSM all , the addition of a remnant Laurentide Ice Sheet and a background meltwater flux does not alter the model response to freshwater forcing, either in magnitude or duration. It is worth noting, however, that these boundary conditions were important in previous experiments with ECBilt-CLIO for prolonging the AMOC response to Lake Agassiz drainage (Wiersma et al., 2006). Thus, the effects of these boundary conditions might be very model-dependent. Differences between early Holocene and preindustrial orbital forcing and greenhouse gas concentrations are relatively minor, and are not expected to have an important influence. This should be verified, though, with additional model experiments.
Another explanation for the model-data discrepancies is that the models are not sensitive enough to freshwater perturbations. If true, this finding would have important implications for future climate projections, particularly as models suggest that continued melting of the Greenland Ice Sheet at its current rate will have a significant impact on the AMOC (Hu et al., 2009). There are few model intercomparisons to determine whether the sensitivity of these three models to freshwater perturbations is representative of coupled climate models as a whole. For hosing experiments of 0.1 Sv for 100 yr under modern boundary conditions, earlier versions of the CCSM3 (CCSM2) and LOVECLIM (ECBilt-CLIO) have AMOC and surface air temperature responses close to the multi-model ensemble mean (Stouffer et al., 2006). For hosing experiments in a Last Glacial Maximum climate, however, AMOC decreases somewhat less in the CCSM3 and LOVECLIM compared to the multi-model ensemble mean (Kageyama et al., 2012). Improved constraints on the size of freshwater forcing and its location with respect to early Holocene convection areas are necessary to rule out the possibility of inadequate model sensitivity.