Direct measurement of the Atlantic Meridional Overturning Circulation (AMOC) began in 2004 with the RAPID array at 26.5°N. Everything we say about the circulation before then rests on AMOC proxies; indirect measures, such as sea surface temperature or salinity, assumed to move in step with the circulation itself.
Studies built on those proxies disagree with each other. Some place the onset of AMOC weakening in the 19th century [1], others after the 1950s [2]. One analysis, working from surface heat fluxes rather than temperature, argues there has been no decline since the 1960s [3]. The same body of evidence has also been read as a warning that the circulation is approaching a tipping point [4][5].
EPOC Deliverable 3.1 tests eight published AMOC proxies against the same 30 CMIP6 models, over the same periods, using the same method. Each had previously been assessed on its own, in models chosen by its own authors. No like-for-like comparison existed.
Five of the eight proxies tested carry a clear AMOC signal in pre-industrial control simulations. These are subpolar SST, upper ocean temperature, salinity, surface heat flux and the deep western boundary current.
Two do not work as stand-alone indices. Gulf Stream upper ocean temperature has almost no relationship with AMOC trends. The sea surface height dipole varies so much between models that its correlation ranges from nothing at all to strong.
For long-term trends, subpolar salinity and surface heat flux outperform SST. That gap widens when the EC-Earth models, which have unusually large centennial variability, are excluded. The SST correlation then falls from 0.74 to 0.53. On that basis the AMOC would account for under a third of the SST signal, while salinity and heat flux hold at around 0.66.
Timing separates the eight as sharply as strength does. Surface heat flux and Labrador Sea mixed layer depth lead the AMOC. SST, upper ocean temperature and salinity lag it. The deep western boundary current changes simultaneously.
Greenhouse gases and aerosols push the AMOC in opposite directions. In single-forcing experiments, greenhouse gas forcing alone weakens the circulation by ~4 Sv. Aerosol forcing alone strengthens it, peaking near 1980[6]. But every proxy responds to both in the same way. A proxy can therefore register that the AMOC has changed, but not what changed it.
EPOC and the EERIE project ran three coupled models at both standard and eddy-rich resolution under identical forcing. The AMOC weakens in all of them. But the SST response comes out opposite in sign across much of the subpolar North Atlantic. Eddy-rich models cool across the whole subpolar gyre. Standard models warm along its southern edge, east of Newfoundland.
The cause is the simulated mean state rather than the AMOC itself. Eddy-rich models capture the northward turn of the North Atlantic Current near Newfoundland, known as the Northwest Corner. The region therefore sits under warm subtropical water that a weakening circulation withdraws. Standard models run the current too zonally, a long-recognised deficiency[7], leaving cold subpolar water in its place. The salinity fingerprint does not have this problem, which is a further argument for preferring salinity to SST.
Proxies supplement direct observation rather than replacing it. Salinity and surface heat flux perform best in models. Yet their observational records reach back only to around 1950 and remain sparse. The heat flux records used in recent literature derive from atmospheric reanalyses never designed to represent long-term trends in poorly observed variables. Model skill is therefore not evidence of real-world skill. The case for the AMOC observing system holds regardless of how far proxy methods improve.
This work forms part of EPOC’s wider effort on explaining past AMOC changes. It also complements our work on AMOC transport variability and coherence and on palaeo proxy records.
EPOC Deliverable 3.1 Evaluation of the robustness of AMOC proxies. Report by Jon Robson, Soumi Chakravorty, Léo Aroucha, Gaurav Madan and David Thornalley, May 2026.
The full report will be available for download when the EC review process is complete.