For the last 6,000 years, the planet has been doing two things at once, depending on who you ask. Run a modern climate simulation forward from the mid-Holocene and Earth warms, gently, as greenhouse gases creep up and orbital geometry shifts. Pull a sediment core from the seafloor, drill an ice core in Greenland, cross-date a stand of bristlecone pines, or slice open a Caribbean coral head, and the same 6,000 years read the opposite way: a slow, uneven cooling that only reverses in the industrial era.
Both records are built by careful people using calibrated instruments. Both cannot be right.
The mismatch has kept paleoclimatologists arguing in print for years, in papers, in review articles, and in the margins of IPCC drafts. The stakes are not academic. If the proxies are correct, models may be missing a cooling mechanism. If the models are correct, the proxies are quietly lying about which season, or which latitude, they actually record.
Two archives, two answers
The physical setup is straightforward. Around 6,000 years ago, in the middle of what geologists call the Holocene, Earth’s orbit tilted the Northern Hemisphere slightly more toward the Sun in summer than it does today. Carbon dioxide concentrations were lower than modern levels but rising slowly. Ice sheets over North America and Scandinavia had finished collapsing a few thousand years earlier.
Feed those boundary conditions into a general circulation model — the same class of code used to project 2100 — and the global average temperature drifts upward by a few tenths of a degree across the next six millennia. The forcing is small, but the sign is consistent across model families.
The proxy record disagrees. Compilations built from marine sediment cores, lake muds, cave stalagmites, tree rings and coral skeletons — the natural archives catalogued in Nature’s paleoclimate reconstruction summary — show global mean temperature falling between the mid-Holocene and the pre-industrial 1800s. The cooling is gentle, non-monotonic, and punctuated by volcanic cold snaps, but the trend is unmistakable in the aggregated data.
One dataset says warming. The other says cooling. Same planet, same 6,000 years.
How you read a tree, a shell, a mud layer
To see why the disagreement is so hard to resolve, it helps to know what a proxy actually is. A tree ring records the growing season, not the year. A coral skeleton records the temperature of a few metres of tropical surface water. An alkenone from a marine sediment core records the temperature of the summer bloom of a specific alga, in a specific patch of ocean, at a specific depth. A pollen grain in a lake bed records what plants managed to reproduce that decade.
None of them record “global average temperature.” That number is a construction, stitched together from thousands of local, seasonal, biased windows onto the past.
The bristlecone pines of California’s White Mountains are among the most extreme examples. Some of these trees germinated more than 4,800 years ago and have laid down a ring every summer since. They are magnificent recorders — but they record high-elevation, warm-season conditions in one corner of the Great Basin. Extrapolating from that ring width to a hemispheric average requires a chain of statistical assumptions, each of which can wobble.
The Arctic version of this problem is well documented. A 2,000-year extended Arctic proxy temperature database published in Nature pulled together lake sediments, ice cores and tree rings from across the high north and found a long, slow cooling running right up to the 19th century — before the industrial-era spike. That cooling is exactly what models struggle to reproduce.
The seasonal-bias hypothesis
One leading explanation is that the proxies are not lying — they are just not answering the question anyone thought they were answering.
Many of the workhorse proxies in the Holocene database record summer, not the annual average. Alkenones from mid-latitude oceans capture bloom-season temperatures. Pollen assemblages reflect growing-season warmth. In the mid-Holocene, Northern Hemisphere summers were unusually intense because of the orbital tilt, then cooled steadily as that tilt relaxed toward the modern configuration. Winters, meanwhile, may have been warming the entire time.
Add a summer signal that cools and a winter signal that warms, and the annual average — which is what the models compute — can nudge upward even as most proxies point down. This idea, that Holocene temperature trends over land are being overprinted by seasonal and oceanic signals, is the throughline of a Nature collection on millennial-scale climate variability.
It is an elegant fix. It is also incomplete. Even after correcting for seasonal bias, some reconstructions still show more cooling than models produce.
The forcing-and-feedback hypothesis
The other leading explanation flips the blame. Maybe the proxies are close to right, and the models are missing something.
Candidates for the missing ingredient include clustered volcanic eruptions that cool the surface for decades at a time, subtle changes in solar output, dust from expanding deserts, and slow reorganisations of ocean circulation that models struggle to capture at millennial length. A Nature Geoscience study on orbital-accelerated transient simulations of the last 800,000 years showed that when volcanic and solar forcings are updated, simulations develop multi-decadal cold periods driven by eruption clusters and sea-ice feedbacks — cooling episodes the older, cleaner simulations simply did not have.
Ocean circulation is the other suspect. The North Atlantic in particular is a lever big enough to swing hemispheric temperature. A synthesis in Eos on 2,000 years of North Atlantic climate change compiled sediment cores and shelled-organism remains and found progressively cooler surface waters through parts of the last two millennia. The authors compared three modern models built specifically for that region and found agreement with proxies in some places and periods, disagreement in others. The Atlantic Meridional Overturning Circulation — the conveyor of warm water from the tropics toward Europe — is the usual suspect, but its exact role remains under active study.
Warming holes and regional weather
Some of the sharpest recent work argues that neither side needs to be entirely wrong — the planet’s response to Holocene forcing was just spectacularly uneven. A 2025 Nature Communications paper on North American “warming holes” and European heat during abrupt Holocene cooling events found the opposite of what a simple ocean-to-land link would predict: in two abrupt episodes, roughly 5,800 to 4,800 years ago and again after 2,200 years ago, the North Atlantic and eastern North America cooled together while Europe counter-intuitively grew warmer and drier. Averaged globally, these opposing regional signals can partially cancel. Sampled locally, by a sediment core off Iceland or a pollen record in Bavaria, they look like whiplash.
The lesson is uncomfortable for anyone who wants a single tidy curve for the Holocene. There may not be one. The planet’s temperature history at millennial resolution is a mosaic, and different proxies see different tiles.
Echoes of the hockey stick
The conundrum did not appear from nowhere. It grew out of the same methodological soil that produced the famous “hockey stick” graph — Michael Mann’s 1998 reconstruction of temperature over the past six centuries, extended a year later to cover the full millennium, in which centuries of gentle variation gave way to a sharp modern spike.
That graph became the most fought-over image in climate science. As The Guardian documented in 2010, leaked emails from the University of East Anglia revealed that in 1999, IPCC lead authors were quietly arguing over whether to feature Mann’s version or one of two rival curves: a multi-proxy reconstruction by Phil Jones built from tree rings, ice cores and coral, or a tree-ring-only chronology from Keith Briffa. According to reports on the hockey stick controversy, Met Office scientist Chris Folland indicated that a millennial temperature diagram was a leading candidate for the policy makers’ summary. Keith Briffa reportedly expressed concerns about treating Mann’s reconstruction as definitive, noting scientific uncertainties in the temperature record. Briffa’s own commentary on the reconstruction noted that medieval warmth, while generally below 20th-century levels, approached it within the uncertainty range.
The disagreement was resolved, in the end, in favour of Mann’s shape, and later reconstructions have largely confirmed the modern spike. But the deeper methodological argument — how do you weigh a tree ring against a coral, a sediment layer against a model output? — never went away. The Holocene conundrum is that argument, extended six times further back in time.
Why the mismatch matters for the next century
A climate model is validated, in part, by how well it reproduces the past. If the model says warming and the mud says cooling, the calibration for the next 100 years becomes shakier. A Nature Geoscience paper on data-model discrepancy in the last millennium laid out the possible culprits explicitly: proxies may be seasonally biased, models may underestimate low-frequency variability, or the reconstructed forcings themselves — solar output, volcanic aerosol loading — may be wrong. Each of those possibilities has a different implication for climate sensitivity, the number that governs how much warming a given dose of CO₂ eventually delivers.
If proxies systematically underestimate past warmth in warm periods and overestimate cold in cold periods, climate sensitivity might be lower than the current consensus. If models underestimate natural variability, some of the warming attributed to greenhouse gases might be background noise — or, more worryingly, the system might swing further than models predict. Neither conclusion is comfortable, and neither is settled.
The instruments getting sharper
The good news is that the tools are improving faster than the argument. Proxy system models — frameworks that simulate exactly how a climate signal becomes a tree ring or an alkenone — let researchers generate synthetic proxies from a model run and compare them like-for-like against real cores. Pseudoproxy experiments test whether a reconstruction method could even recover the true signal if it were handed perfect data. Multi-method ensembles produce not one temperature curve but a family of them, with quantified uncertainty.
New proxies keep coming online too. Lipid biomarkers from lake sediments, clumped isotopes from carbonates, sub-annual laser scans of coral skeletons — each adds a channel to a dataset that used to be dominated by a handful of long tree-ring chronologies and a small number of famous ice cores.
Whether these tools eventually collapse the conundrum into a single coherent story, or reveal that the Holocene really was a patchwork of regional trends that no global average can honestly describe, is still open. The bristlecones will keep laying down rings while it gets sorted out. Some of them have been recording the argument since before it started, and will keep recording long after this decade’s papers are cited, contested and shelved.
