What can we defensibly add to IPCC projections? Research updated October 12, 2026; complete-year temperature comparisons through 2025.

Executive summary: the potential IPCC underestimation numbers you can use—and their conditions

No established percentage applies to every IPCC forecast. But our review does support several clearly defined calculations and planning tests. They apply to different quantities. They must not be added together.

In section 16, you will see a summary of defensible rough percentages or ranges that can be added to IPCC temperature forecasts for extended time periods for complete and optimal individual, corporate, and national prudent planning across all likely-to-possible IPCC scenarios. The chart below shows those percentages or ranges. 

For everyday planning, the most useful global test is: keep today’s warming fixed, then compare the selected near-term projected additional warming with a case about 30% higher. Reassess it as observations and forcing estimates improve. The exact calculated central uplift is 31.4%; the rounded 30% is a convenient test value.

For land, use land projections whenever available. For sea ice, evaluate loss and threshold dates separately. We have not established one correction for crops, flooding, deaths, sea level, or ecosystem collapse. Those outcomes need their own evidence and models.

The historical picture is mixed: some sampled temperature curves underestimate later warming, the first sampled IPCC curve overestimates its subsequent trend, and sea ice shows important examples of underestimation. Our new comparisons are exploratory calculations; the formulas, assumptions, and sources are disclosed below.

The planet did not wait for the next executive summary

Imagine buying a smoke alarm and then arguing for twenty minutes about whether its prediction was 12% too cautious. Meanwhile, the curtains are developing opinions about combustion. Climate forecast accuracy matters—but the argument should help us respond to the risk, not become a substitute for responding.

We began with a reasonable question: have successive reports from the Intergovernmental Panel on Climate Change, or IPCC, understated the warming and consequences that followed? We then asked a harder question: if so, can we calculate a defensible adjustment instead of merely saying “it is worse than expected”?

The answer requires more than one scoreboard. A global temperature trend, a land temperature, an Arctic amplification ratio, a shrinking ice area and the date a coastline floods are different measurements. A single impressive percentage cannot do all five jobs. Even a very confident percentage deserves a job description.

Quick navigation

1. Executive summary
2. Why this review matters
3. How this analysis was built
4. The reports and their scenarios
5. Two percentages that must not be confused
6. Report-to-report temperature checks
7. Longer comparisons through 2025
8. The Hausfather study and all 17 projections
9. Forcing, sensitivity and ECS
10. Feedbacks and critical carbon transition levels
11. Northern Hemisphere and Arctic weighting
12. The conditional global adjustment
13. The separate LAND calculation
14. The separate sea ice calculation
15. Consequence timetables and other risks
16. The next 40 years in three-year steps
17. What a stronger accuracy test needs
18. How to use the results
19. Questions and AI review prompts
20. FAQ
21. Glossary
22. Bibliography and reproducibility
23. Next steps and attribution

1. How this analysis was built

We reviewed the major IPCC physical-science assessment sequence, its scenario families, representative archived temperature curves, annual observations, and selected research on land, Arctic warming, sea ice and sea level. We also examined the published Hausfather evaluation and its public analysis code. The publisher’s separate supporting Word appendix could not be retrieved; we therefore do not claim an appendix-based inventory of every process in every original model.

Methodology in plain language: first compare like quantities over the same years. Keep the original scenarios and baselines visible. Calculate differences in degrees Celsius before converting them to percentages. Then ask whether a mismatch came from climate physics, an emissions assumption, natural variability, a measurement definition or an incomplete comparison.

The new temperature checks use NASA GISTEMP v4 annual land–ocean temperatures and selected archived projections. They are not a full reanalysis of all model members or all scenarios. The update ends in 2025 because a final annual 2026 temperature did not yet exist at the research cutoff. [9] [10]

We distinguish three evidence levels throughout: published findings, our reproducible arithmetic, and conditional future sensitivity tests. A test can be useful for planning without being an established correction to climate science.

2. The reports: six assessment generations, several kinds of summary

The main Working Group I physical-science reports appeared in 1990, 1995, 2001, 2007, 2013 and 2021. Their gaps were 5, 6, 6, 6 and 8 years—not a regular three-to-five-year schedule. A Summary for Policymakers belongs to a report; a later synthesis integrates that assessment’s findings. [1] [2] [3] [4] [5] [6]

These are conditional futures, not six attempts to predict one fixed destiny. Different scenarios assumed different emissions, concentrations, energy systems and policies. Their printed temperature numbers also use different reference periods. A larger number in a newer report does not by itself prove that an older prediction failed.

The 1992 supplement updated FAR and introduced IS92 scenarios. The 2000 SRES supplied emissions pathways; the 2011/2012 extremes report examined disaster risks; and the 2018–2019 special reports examined 1.5°C, land, and the ocean and cryosphere. They matter greatly, but they are not additional independent generations of complete global temperature forecasts. [1] [7]

AR6 impacts and mitigation reports followed in 2022; its synthesis appeared in 2023. The synthesis did not rerun a new generation of models. AR7 remains in preparation as of this review, so there is no completed AR6-to-AR7 accuracy interval. [8]

3. The denominator: where innocent-looking percentages get into trouble

Let O be observed additional warming and P be projected additional warming over the same period. We use two different percentages.

Calculation:
Signed shortfall = 100 × (O − P) / O
Required uplift to the projection = 100 × (O − P) / P

Positive shortfalls mean underprojection; negative shortfalls mean overprojection. If the forecast gives 0.20°C and observations give 0.30°C, it misses 33.3% of observed warming. But the forecast must rise by 50% to reach 0.30°C. “Underestimated by a third” and “add half” can describe the very same case.

Neither percentage is an accuracy probability. When observed warming is zero, the first is undefined. When it is very small, the first can become enormous. The thermometer has not suddenly become ridiculous; the denominator has.

4. From one report to the next: the short temperature checks

For each interval, subtract the earlier year’s value from the later year’s value in both observations and the selected projection. The annual NASA anomalies below use 1951–1980 as their baseline. They must not be read as warming above 1850–1900. A constant baseline shift cancels when taking a difference. [9]

FAR-to-SAR has no change between its two annual NASA endpoints, so its shortfall percentage cannot be calculated. The 1991 Pinatubo eruption and other short-term fluctuations complicate this brief window. SAR-to-TAR is approximately a match: its −0.002°C difference is smaller than the annual input’s displayed 0.01°C precision.

TAR-to-AR4 shows a 25.2% shortfall, equivalent to a 33.6% uplift to the selected projected increment. AR5-to-AR6 shows an 18.3% shortfall, equivalent to a 22.4% uplift. These are unadjusted interval calculations, not full assessments of model physics.

AR4-to-AR5 produces −881.9%, because its observed endpoint increase is only 0.01°C. Printing that result is transparent; using it as a transferable error factor would be misleading. Its absolute error is −0.088°C.

For the five-year check, each endpoint becomes its preceding five-year average, for both model and observations. SAR’s apparent slight overprojection becomes a 48.1% shortfall. That sign reversal warns against averaging short-interval percentages into a universal multiplier.

An AR6-to-current completed assessment comparison is unavailable. Its 2021–2040 target average is also unfinished. The observed 2016–2025 global average of 1.26°C and the AR6 2011–2020 average of 1.09°C describe different decades; subtracting them does not measure a forecast error. [6] [11]

5. Longer windows through 2025: a real mismatch, with a narrower meaning

We next fitted a straight line to every annual value from each report’s publication year through 2025. This uses more information than two annual endpoints. It still tests selected curves and one observational product, without updating the forcing-response comparison.

Four later sampled curves fall below the observed warming trend. Their shortfalls are about 33–44% of observed warming; matching observations would require roughly 49–80% increases to those particular projected trends. FAR goes the other way: its projected trend is higher than observed.

Calculation:
Example, AR5:
Shortfall = 100 × (0.357692308 − 0.220385191) / 0.357692308 = 38.4%
Uplift = 100 × (0.357692308 / 0.220385191 − 1) = 62.3%

These windows overlap, use different lengths and include unusually warm recent years. The AR5 window contains only 13 annual values. Actual methane, aerosol and other forcing histories differ from the archived scenario assumptions. No new uncertainty intervals, significance tests or multi-dataset forcing-adjusted evaluation were completed for this table.

A separate provisional AR5 check gives a different perspective: NASA’s 2016–2025 mean is about 0.56°C above its 1986–2005 mean, within AR5’s assessed 0.3–0.7°C near-term range. The forecast target is 2016–2035, so the test is not complete. A shortfall in one ensemble central trend is not equivalent to observations falling outside that assessment’s range. [5] [9]

Why the tempting “average correction” does not survive inspection

Calculation:
Equal-weight mean of the five uplift percentages:
(−12.0 + 79.9 + 48.6 + 54.1 + 62.3) / 5 ≈ 46.6%
Mean using only the four later positive cases ≈ 61.2%
Ratio of the five mean observed and projected rates − 1 ≈ 41.0%

Each is legitimate arithmetic. None is an established future IPCC correction. Excluding the overprojecting case changes the answer. So does the choice between averaging ratios and taking a ratio of averages. Overlapping years and related models also prevent these five cases from acting like five independent experiments.

6. What Hausfather and colleagues actually tested

Hausfather, Drake, Abbott and Schmidt found 10 of 17 projections consistent with observations on a temperature-versus-time test, and 14 of 17 after accounting for forcing differences. They found no systematic overall underestimation or overestimation over their evaluation periods. Their main analysis ended in 2017. [12]

Source: the paper’s Table 1. AR4 was additionally examined over 2007–2017 but excluded from the principal 17 because the comparison was short and uncertain. [12]

These are not 17 independent IPCC model generations. Several are scenarios from the same physical model. Passing the global-temperature test cannot establish that every regional pattern, feedback, abrupt transition, or consequence timetable is correct.

The main paper and public notebook show a statistical comparison of archived projections—not a new Earth-system simulation. The original models retained their original physical assumptions. Statistical consistency means the estimated difference includes zero within the study’s uncertainty test; it does not mean perfect agreement. [10]

7. Forcing differences are not an ECS adjustment

Radiative forcing describes the energy-balance push from changing gases, aerosols and other drivers, measured in watts per square metre. Equilibrium climate sensitivity, or ECS, describes eventual warming after doubled CO2, after the system approaches equilibrium. They are different variables. [13]

The archived evaluation regressed warming against forcing and expressed the slope as an “implied transient climate response,” using 3.7 W/m² for doubled CO2. Its observational comparison used five temperature products and an ensemble of forcing estimates. It did not select a new ECS for each model to make it agree with observations. [10]

Calculation:
Implied response = slope of temperature versus forcing × 3.7 W/m²

Illustration:
Projected forcing growth 0.6 W/m² per decade × response 0.5°C per W/m² = 0.30°C per decade
Observed forcing growth 0.4 W/m² per decade × the same response = 0.20°C per decade

That invented example overprojects the calendar warming while getting the response per unit forcing right. A policy-driven reduction in a gas or a different aerosol history can change the first comparison without proving that the second is wrong.

This method also has limits. Its approximately linear temperature–forcing relationship, instantaneous forcing definitions available for early models and simplified treatment of time dependence do not reproduce every effect of rapid forcing changes, ocean uptake or nonlinear feedbacks. The public code is an evaluation tool, not an inventory of all climate mechanisms. [10]

AR6 assesses an ECS best estimate of 3.0°C and a likely range of 2.5–4.0°C. Hansen and colleagues’ 2023 paper argues for 4.8 ± 1.2°C, using a different assessment of paleoclimate evidence. These are scientific sensitivity estimates, not the “forcing differences” correction. [13] [14]

Calculation:
At the same forcing and equilibrium assumptions:
(4.5 / 3.0 − 1) × 100 = 50% greater eventual warming
(4.8 / 3.0 − 1) × 100 = 60% greater eventual warming

Those equilibrium ratios cannot automatically become 50% or 60% extra warming by 2036 or 2066. Ocean heat uptake, response times, future forcing and the scope of included feedbacks determine the warming reached by a particular date. Equilibrium warming is also not identical to warming irreversibly committed under future emissions reductions. [14]

8. Feedbacks, nonlinear responses and critical carbon transition levels

A positive feedback reinforces an initial change: warming melts reflective ice; darker surfaces absorb more sunlight; more heat promotes further melting. A nonlinear response changes disproportionately with its driver. A threshold can trigger a change that is difficult to reverse. Scientists often call such thresholds “tipping points”; here we also use critical carbon transition levels when discussing carbon-related transitions.

The Hausfather comparison did not add a new suite of feedbacks to the old models. We must inspect the original models to learn what was represented. The following is a process-level guide, not a claim that every one of the 17 included every item.

A concentration-driven model can correctly respond to specified CO2 while failing to test how thawing soils or damaged ecosystems might change future CO2. That is a real distinction between validating a temperature response and validating the future carbon cycle.

The appropriate way to add a missing process is to quantify its genuinely additional emissions or energy effect, its timing and its interactions. A list of omitted mechanisms is not a list of known temperature increments. Some effects are already included elsewhere in the assessed response.

Calculation:
Physical accounting rule:
Additional warming from a proposed missing process = temperature in a coupled run that includes the process − temperature in the matched baseline run

Keep other drivers comparable, and check whether the assessed baseline already includes an estimate of that process.

9. Northern Hemisphere warming: a weighting question, not a missing hemisphere

Modern coupled climate models calculate regional temperatures within a global system. The Northern Hemisphere, Arctic, land, ocean and sea ice interact in those calculations. Scientists then area-weight the regional fields to obtain a global average. They do not simply impose a universal 2–4-times Northern Hemisphere multiplier. [19]

The Northern Hemisphere is half of Earth. The Arctic is a much smaller northern region. “The Arctic warms nearly four times as fast” does not mean the entire Northern Hemisphere does. If NH and SH are hemispheric average anomalies with consistent definitions, the global anomaly is (NH + SH) / 2.

Rantanen and colleagues measured an Arctic/global warming-trend ratio of about 3.8 for 1979–2021, north of 66.5°N. The corresponding ensemble means were about 2.5 in CMIP5 and 2.7 in CMIP6. Internal variability contributes to the observed ratio, and individual simulations vary substantially. [20]

Calculation:
Regional amplification uplift = 100 × (3.8 / 2.7 − 1) = 40.7%
This changes an amplification ratio, not the planet’s total temperature by 40.7%.

A separate, explicitly hypothetical global calculation

The spherical area north of 66.5°N is about 4.15% of Earth. Suppose we change the Arctic/global trend ratio from 2.7 to 3.8 while keeping the average non-Arctic trend fixed. Because each amplification ratio is measured relative to its own global mean, the denominator must change too.

Calculation:
Arctic area fraction f = (1 − sin 66.5°) / 2 = 0.041470
If A = Arctic trend / global trend, then non-Arctic trend = global trend × (1 − fA) / (1 − f)
Global multiplier = (1 − f × 2.7) / (1 − f × 3.8)
= 1.05415, or a 5.4% increase in the global trend
Using 2.5 instead of 2.7 gives about 6.4%.

This is our spatial sensitivity calculation, not a validated future correction. It assumes the historical amplification ratios can be reused and excludes changes in the rest of the world. Carbon release and circulation effects could change that rest-of-world assumption; they require further modelling.

Warmer land and a rapidly warming Arctic are already part of the area-weighted global average. Their higher temperatures can be underrepresented in a particular model, but their existence does not establish a geographical weighting error. Demonstrating that 40–50% of Earth exceeds a particular warming level would require a gridded, area-weighted calculation for the same period.

10. A defensible conditional global adjustment

Here is the most practical adjustment from this project. It does not claim we have discovered a permanent IPCC error. It asks what happens if specified recent warming rates persist over the next decade instead of following one selected archived central curve.

Calculation:
G_test(t) = G_observed(t0) + kG × [G_IPCC(t) − G_IPCC(t0)] + δ_additional(t)

G = global warming anomaly; t0 = the starting year.
kG scales only the future increment.
δ_additional is used only for a separately quantified extra effect, with δ_additional(t0) = 0.

Use δ_additional = 0 for the examples here. We have not calculated an extra temperature allowance for the missing-feedback inventory. Starting conditions and uncertainty should also be updated separately.

The archived assessed AR6 SSP2-4.5 mean curve adds 0.235926891°C between 2026 and 2036. At constant warming rates of 0.27, 0.31 and 0.35°C per decade, the corresponding increments are 0.270, 0.310 and 0.350°C. [21]

Calculation:
Central test: kG = 0.310 / 0.235926891 = 1.313966
Uplift = 100 × (1.313966 − 1) = 31.4%
With the illustrative 1.4°C starting level:
Unscaled curve in 2036 = 1.4 + 0.235926891 ≈ 1.64°C
Central persistence test = 1.4 + 0.310 = 1.71°C

The 0.27°C rate comes from the latest human-induced warming assessment. The 0.35°C case approximates faster recent rates in Foster and Rahmstorf’s piecewise analysis; their quadratic approach yields lower rates. The 0.31°C middle value is our chosen midpoint, not a published best estimate. [11] [22]

The archived Figure SPM.8 curves use global surface-air temperature and 20-year means; the observational warming anchor and attributed rate use a blended global surface-temperature metric. Our re-anchoring is an approximation for illustration. A formal recalibration must harmonise those temperature definitions and averaging periods. [21]

These are different rate estimates and methods, not a single probability interval. The acceleration study removes estimated ENSO, volcanic and solar effects and finds evidence of acceleration; it also notes observations remain within AR6 model prediction ranges. A mismatch with a central curve is not necessarily a miss of the assessed range. [22]

Conditions for using about +30%: identify the archived SSP2-4.5 curve; apply the multiplier only to its 2026–2036 increment; disclose that 0.31°C per decade is assumed to persist; and retain comparison cases. Other scenarios, dates, updated forcings or regional metrics require recalculation. Do not carry this multiplier unchanged through the next 40 years.

11. LAND: the separate calculation for where most people live

Sources: AR6 and the latest annual update. These are historical average anomalies, not forecasts for the same future year. [6] [11]

Calculation:
Land/global ratio = 1.81 / 1.26 = 1.436508
Land excess = 100 × (1.81 − 1.26) / 1.26 = 43.7%
Difference = 1.81 − 1.26 = 0.55°C

Illustrative area-weighted global check:
0.29 × 1.81 + 0.71 × 1.03 = 1.256°C ≈ 1.26°C

The arithmetic shows why the hotter land is not missing from the global average. Roughly 71% of Earth is ocean, with a smaller surface warming anomaly. The global figure is appropriate for the planet’s surface average; it is insufficient by itself to describe the heat exposure of people, crops or infrastructure on land.

The hypothetical “miss” from using 1.26°C as if it were a land prediction would be (1.81 − 1.26) / 1.81 = 30.4% of the land anomaly. But it was never a land prediction. That is a category mistake, not an observed IPCC land forecast error.

A cautious future land translation

Calculation:
L_test(t) = L_observed(t0) + 1.44 × [G_test(t) − G_test(t0)]
Example: 1.44 × 0.310°C = 0.4464°C, or about 0.45°C additional average land warming.

This deliberately assumes the observed historical land/global level ratio is a reasonable coefficient for the future increment. That assumption is provisional. A ratio of historical levels need not equal a ratio of future changes. Use assessed regional or land projections when available, and anchor them to a corresponding land starting level.

Do not use the 2016–2025 land average as though it were a measured 2026 land starting value. Do not add 44% to an existing land-specific forecast. And do not convert a 44% land temperature difference into 44% more crop loss, fire damage or deaths. The exposure–damage relationship must be evaluated separately.

12. Sea ice: a separate loss-sensitivity calculation

Diebold and Rudebusch compared September Arctic sea ice area with cumulative CO2 emissions over 1979–2019. Their observed sensitivity magnitude was 2.891 m² of area loss per tonne of CO2. For six paired model families, the median slope biases were +0.873 for CMIP5 and +0.982 for CMIP6. [23]

Calculation:
Observed slope = −2.891 m² per tonne CO2
CMIP5 selected median slope = −2.891 + 0.873 = −2.018
CMIP6 selected median slope = −2.891 + 0.982 = −1.909

CMIP5 uplift to loss sensitivity = 100 × (2.891 / 2.018 − 1) = 43.3%
CMIP6 uplift to loss sensitivity = 100 × (2.891 / 1.909 − 1) = 51.4%

Our rounded +50% case comes from that selected CMIP6 median. Its corresponding shortfall relative to observed sensitivity is 34.0%, illustrating again why “shortfall” and “uplift” differ. This is not a 2025-updated regression or a correction for every CMIP6 model; individual models differ, including some with stronger responses than observed. [23]

Calculation:
For an explicitly linear, comparable carbon-driven loss test:
S_test(t) = max[0, S_observed(t0) − 1.51 × (S_model(t0) − S_model(t))]

S = September Arctic sea ice area.
The expression in parentheses is the model’s future area loss.
Keep present ice fixed; increase the future loss, and impose a zero floor.

Suppose the model loses another 1.0 million km² under the specified cumulative-emissions increment. The test loses about 1.51 million km². It does not multiply the remaining ice area by 1.51. It does not add 51% to temperature. The example is invented to illustrate the formula.

The physical coefficient is tied to cumulative emissions and a particular metric, season and sample. Carrying it into the future assumes the fitted sensitivity persists. Prefer re-estimating the response for the actual model, current observations and chosen pathway. As ice approaches zero, a simple linear rule becomes less informative about seasonal duration and other physical changes.

13. Consequence timetables: where understatement can have a different shape

September Arctic ice

Wang and colleagues’ 2026 observationally constrained study places the first practically ice-free September sea ice area in 2037–2049 under SSP2-4.5 and 2033–2042 under SSP5-8.5, depending on the observational dataset. These are conditional projections, not guaranteed dates. [24]

In that study, area-based dates advance by roughly 20 years and 10 years, respectively, relative to its own unconstrained estimates. Extent-based dates are about ten years later. The threshold is less than 1 million km²; “practically ice-free” does not mean absolutely no ice. The dataset-dependent spans are not complete probability intervals. [24]

AR6 already assessed at least one practically ice-free September before 2050 as likely under all five illustrative scenarios. The newer dates can sharpen planning without proving the entire AR6 statement was wrong. [25]

An earlier ice mismatch

Stroeve and colleagues’ 2007 comparison found observed September ice-extent decline of about 7.8% per decade against a model average of about 2.5% over 1953–2006. The observed decline was roughly 30 years ahead of the ensemble mean. Much of that comparison evaluated historical simulations rather than a pure forecast made 30 years earlier. [26]

Calculation:
Historical ice-extent shortfall = 100 × (7.8 − 2.5) / 7.8 = 67.9%
Uplift to that modelled decline rate = 100 × (7.8 − 2.5) / 2.5 = 212%

That 212% is a historical ice-extent rate comparison. It is not interchangeable with the later 51.4% area-per-carbon calculation. Different periods, metrics and denominators produce different answers.

Sea level and impact severity

Rahmstorf and colleagues reported that 1990–2006 sea-level rise was about 50% faster than the TAR central projections. That corresponds to a 33.3% shortfall relative to observed rise. It is evidence of a historical sea-level mismatch, not a 50% global-temperature correction. [27]

AR4’s sea-level ranges did not encompass the full uncertainty in future rapid ice-flow changes. Later assessments improved their treatment. AR6’s impacts assessment also found many risks higher than previously assessed. Damage can worsen because exposure, vulnerability and nonlinear responses change, even when the global-temperature forecast is reasonable. [4] [28]

No single percentage can combine sea level in metres, ice in square kilometres, crop yields, heat deaths and ecological damage. For each consequence, define the region, threshold, duration, scenario and uncertainty range. A 2100 outcome cannot yet be scored as a completed observation.

Temperature threshold timing: a transparent illustration

Calculation:
Years to a specified warming threshold = (threshold − starting anomaly) / annual warming rate
From an illustrative 1.4°C to 2.0°C:
0.60 / 0.027 ≈ 22.2 years
0.60 / 0.031 ≈ 19.4 years
0.60 / 0.035 ≈ 17.1 years

Starting in 2026, those constant-rate examples reach 2°C around 2048, 2045 and 2043. They are scenario-free arithmetic, not new assessed crossing-date forecasts. A single year above a threshold differs from a sustained climate average. Nor does reaching 2°C prescribe the exact date of a particular ice-sheet or ecosystem transition.

14. The next 40 years: scenarios, not a calendar with a thermostat

We retained the assessed AR6 Figure SPM.8 curves and replaced each curve’s 2026 starting level with an illustrative common anomaly of 1.4°C above 1850–1900. That anchor comes from 2025 human-induced warming of 1.37°C plus one year at 0.027°C, rounded to 1.4°C. It is an estimate of underlying warming, not a completed observed 2026 annual average. [11] [21]

Calculation:
Updated illustrative scenario in year y = 1.4 + [AR6 mean curve(y) − AR6 mean curve(2026)]

This constant-offset calculation preserves the archived future increments. It does not rerun climate physics or update the forcing pathways. All figures below are smoothed underlying warming above 1850–1900, rounded to 0.1°C. The decimal is a display choice, not a precision guarantee.

The rows take three-year steps through 2065; 2066 is added to complete 40 years. Lower pathways depend on substantial emissions reductions. The very high pathway is a conditional stress case, not an assigned probability or a claim that it is the most likely outcome.

The intermediate re-anchored curve reaches about 2.3°C in 2066. Constant rates reach about 2.5°C or 2.8°C. Neither a recent rate nor the selected +30% decade test should be treated as a permanent rate for four decades.

At 2066, shifting the source percentile curves by each scenario’s central re-anchoring offset gives approximate 5th–95th ranges of 1.4–2.3°C for SSP1-2.6, 1.9–2.9°C for SSP2-4.5, and 2.4–3.8°C for SSP5-8.5. These retain the old source spread and omit uncertainty in the new anchor and changed post-assessment forcings. They are not complete uncertainty intervals for a new forecast. [21]

The table does not predict each future year’s weather. ENSO, eruptions and other variability can move annual temperatures above or below a smoothed curve. Climate projections give conditional ranges and trends; Earth has not signed a delivery agreement for each decimal place.

15. What a stronger accuracy test needs

The next step is to separate the causes of the discrepancies. A publishable calibration needs matched forcings and metrics, several observational datasets, uncertainty estimates, and independent evaluation periods. It should distinguish projections made before an outcome from historical simulations and later reconstructions.

A useful assessment should report absolute errors in °C, rate errors, interval coverage and threshold-timing errors. It should identify the chosen central estimate and the full assessed range. Passing a range test is different from matching the central estimate.

Future validation should predefine its windows and choices rather than select them after seeing the answer. Refit the sea ice relationship with observations through 2025. For land, compare actual land forecasts with actual land observations. For the Arctic, compare spatial patterns and distributions, not just a global scalar.

16. How to use this without turning the numbers into folklore

For a near-term global planning exercise, label the selected baseline, re-anchor its starting point and run the +15%, +30% and +50% future-increment sensitivities with the disclosed persistence assumptions. For decisions on land, use regional projections and local heat information; the 1.44 translation is only a fallback illustration. For sea ice, test the relevant carbon response and threshold definition.

Keep these calculations in separate ledgers. Land/global and Arctic/global ratios describe spatial differences. A global persistence test describes a possible future trend. Sea ice sensitivity describes area loss per carbon emitted. They overlap physically, but their percentages cannot be summed into an “overall underestimate.”

The strongest observation from this project is that reasonable historical global-temperature performance and serious underestimation of a regional consequence can coexist. The evidence warrants better outcome-specific testing and updated planning. It does not warrant a permanent magic multiplier.

Ask what matters for your decision: how much additional warming, where, by when, under which emissions path, with which consequences? That question is more useful than choosing the largest percentage in the room. Percentages are helpful servants and exceptionally poor climate ministers.

17. Reader questions and AI review prompts

Reflection: Does your existing plan rely on the global mean when the decision concerns land, a coastal location or a polar ecosystem? Which quantity and threshold would actually change your decision?

Prompt 1: audit the arithmetic

Prompt 2: test the assumptions

Prompt 3: review a specific consequence

AI can help check calculations, but it can also invent a remarkably well-dressed number. Verify its sources and arithmetic before using its answer.

18. Frequently asked questions

Can I say the IPCC is 30% too low?

No. Say: “For a specified 2026–2036 SSP2-4.5 central-curve comparison, assuming 0.31°C per decade persists gives about 31% more future warming than that curve.” That is a conditional test, not a demonstrated overall bias.

Can I add 30% to today’s 1.4°C?

No. In this framework the multiplier changes future additional warming. The historical starting anomaly stays fixed. Multiplying 1.4°C itself would create a different and unsupported claim.

Why is the historical uplift average 46.6% while the near-term test is about 30%?

They answer different questions. One averages selected older trend mismatches over overlapping windows. The other compares a particular ten-year curve increment with an explicitly assumed rate. Neither is a universal error estimate.

Does the land figure show that the global average is underweighted?

No. The global mean includes land and ocean with area weights. A land/global ratio is useful for understanding terrestrial exposure; it does not replace a direct test of land forecast accuracy.

Can I multiply land damages by 1.44?

No. Temperature changes and consequences use different relationships. Exposure, thresholds, humidity, adaptation and location matter. Calculate the relevant damage response separately.

Can I add the global, land, Arctic and sea ice percentages together?

No. They describe different quantities and may include overlapping physical effects. A combined result requires a consistent coupled calculation, not percentage addition.

Did the Hausfather study prove every feedback was correctly modelled?

No. It evaluated selected historical global-temperature responses. It did not separately validate all carbon-cycle feedbacks, regional impacts or abrupt transitions.

Do forcing adjustments mean changing ECS to fit observations?

No. Driver assumptions and physical sensitivity are separate. Comparing response per unit forcing asks whether the original model reacted appropriately to its assumed energy-balance change.

Does a later synthesis provide a new forecast to score?

Not automatically. A synthesis usually integrates the current assessment’s results. Verify whether there is a new model generation, projection or target period before counting it as an independent test.

Does practically ice-free mean no Arctic ice remains?

No. This article uses a specified September threshold below 1 million km². A first threshold crossing differs from repeated crossings, a longer ice-free season or a year-round ice-free ocean.

Does showing one decimal place mean accurate forecasts to 0.1°C?

No. The table is rounded for readability. Scenario and model uncertainty are much larger, and individual annual weather differs from the smoothed climate trend.

What would justify a true transferable correction?

A replicated, like-for-like test showing a persistent physical-response bias across appropriate periods, observations and scenarios, with uncertainty and independence handled. A correction must then work on outcomes not used to choose it.

Can recent acceleration be important while still inside an assessed range?

Yes. Central estimates and uncertainty ranges are different tests. Faster warming within a range can still change planning, while giving no basis to say that the entire range has failed.

Why not calculate one percentage for all consequences?

Because metres of sea level, square kilometres of ice, crop output and human mortality are not one interchangeable metric. Outcome-specific errors and dates are more informative.

19. Glossary

Anomaly

Temperature change relative to a stated reference average, rather than absolute air temperature.

Area weighting

Giving each region a contribution proportional to its share of the measured surface.

Arctic amplification

The ratio of Arctic warming to global warming over a specified region, period and temperature metric.

Baseline

The reference period from which an anomaly is measured.

Blended temperature

An observation-like combination of surface air temperature over land and sea surface water temperature over oceans.

CMIP5 / CMIP6

The fifth and sixth phases of the Coupled Model Intercomparison Project, collections of model experiments used by assessments.

Critical carbon transition level

Here, a carbon-related threshold beyond which a climate or ecosystem response can change sharply or become difficult to reverse. The broader scientific term “tipping point” is also used.

ECS

Equilibrium climate sensitivity: eventual warming associated with doubled atmospheric CO2 under the defined feedback and equilibrium assumptions.

Emissions-driven / concentration-driven

A simulation may calculate gas concentrations from emissions, or take concentrations as prescribed inputs. These test different parts of the system.

Ensemble

A collection of simulations, often with several models or initial conditions.

Feedback

A response that reinforces or opposes an initial change.

Forcing

A change in the climate energy balance caused by a driver such as greenhouse gases or aerosols.

Global mean

An average for the defined global surface; it is not the temperature experienced everywhere.

IGCC

Indicators of Global Climate Change, an annually updated scientific assessment of key climate indicators.

Internal variability

Climate fluctuations generated within the system, including ocean–atmosphere variability.

Nonlinear response

A response that is not proportional to its driver over the relevant range.

OLS trend

An ordinary least-squares straight-line fit to values over time.

Practically ice-free

Here, a September Arctic ice metric below 1 million km²; specify area or extent.

Projection

A conditional future result under stated assumptions.

RCP / SSP / SRES / IS92

Scenario frameworks used in different assessment generations; they are not interchangeable probability forecasts.

Sea ice area / extent

Area counts the ice-covered part of each cell; extent counts qualifying cells, typically with at least 15% ice cover, including open water within them.

Shortfall

In this article, observed minus projected change divided by observed change.

TCR

Transient climate response: warming around CO2 doubling in a standard gradual-increase experiment. An implied response from historical temperature and forcing is a related diagnostic.

Thermokarst

Ground and landscape changes caused by thawing ice-rich permafrost.

Uplift

In this article, the increase required relative to the projected quantity, using that projected quantity as denominator.

W/m²

Watts per square metre, a unit of energy flow used for forcing.

20. Bibliography, intellectual lineage and reproduction

Primary assessment documents, original research and public data are listed below. The numerical extensions and sensitivity formulas are this project’s calculations, not new peer-reviewed findings. Links identify the scientific sources; organisational action links appear separately at the end.

Original sources and data

1. IPCC (1990, 1992). First Assessment and supplementary assessment. Original scenario rates and report sequence.
   Read source 1
2. IPCC (1995). Working Group I Summary for Policymakers, Second Assessment. University-hosted copy of original text.
   Read source 2
3. IPCC (2001). Working Group I Summary for Policymakers, Third Assessment.
   Read source 3
4. IPCC (2007). Working Group I Summary for Policymakers, Fourth Assessment; temperature projections and sea-level qualifications.
   Read source 4
5. IPCC (2013). Working Group I Summary for Policymakers, Fifth Assessment, E.1 and Table SPM.2.
   Read source 5
6. IPCC (2021). Summary for Policymakers. In: Climate Change 2021: The Physical Science Basis, pp. 3–32. DOI: 10.1017/9781009157896.001. See A.1 and Table SPM.1.
   Read source 6
7. IPCC special and methodology reports. Includes SRES, SREX, Global Warming of 1.5°C, Climate Change and Land, and Ocean and Cryosphere.
   Read source 7
8. IPCC. Sixth assessment cycle and the seventh assessment in preparation.
   Read source 8
9. NASA GISTEMP v4. Global annual land–ocean temperature anomalies, J-D column; frozen inputs retrieved October 12, 2026, complete years through 2025.
   Read source 9
10. Hausfather et al. Public OldModels code and projection archive. Notebook: obs_temp_forcing_analysis.ipynb. See exact workbook selections below.
    Read source 10
11. Forster, P. M., et al. (2026). Indicators of Global Climate Change 2025: annual update of key indicators of the state of the climate system and human influence. Earth System Science Data, 18. DOI: 10.5194/essd-18-3889-2026.
    Read source 11
12. Hausfather, Z., Drake, H. F., Abbott, T., and Schmidt, G. A. (2020). Evaluating the Performance of Past Climate Model Projections. Geophysical Research Letters, 47, e2019GL085378. Published online December 4, 2019. DOI: 10.1029/2019GL085378.
    Read source 12
13. IPCC (2021). AR6 Working Group I, Chapter 7: The Earth’s Energy Budget, Climate Feedbacks, and Climate Sensitivity.
    Read source 13
14. Hansen, J. E., et al. (2023). Global warming in the pipeline. Oxford Open Climate Change, 3, kgad008. DOI: 10.1093/oxfclm/kgad008. NASA publication record.
    Read source 14
15. Hansen, J., et al. (1984). Climate sensitivity: Analysis of feedback mechanisms. Climate Processes and Climate Sensitivity, pp. 130–163. NASA publication record.
    Read source 15
16. Hansen, J., et al. (1988). Global climate changes as forecast by Goddard Institute for Space Studies three-dimensional model. Journal of Geophysical Research, 93, 9341–9364. DOI: 10.1029/JD093iD08p09341.
    Read source 16
17. Manabe, S., and Stouffer, R. J. (1993). Century-scale effects of increased atmospheric CO2 on the ocean–atmosphere system. Nature, 364, 215–218. DOI: 10.1038/364215a0.
    Read source 17
18. IPCC (2021). AR6 Working Group I, Chapter 5: Global Carbon and other Biogeochemical Cycles and Feedbacks; especially permafrost representation.
    Read source 18
19. IPCC (2021). AR6 Working Group I, Chapter 4: Future Global Climate—Scenario-Based Projections and Near-Term Information. Regional warming and land–ocean differences.
    Read source 19
20. Rantanen, M., et al. (2022). The Arctic has warmed nearly four times faster than the globe since 1979. Communications Earth & Environment, 3, 168. DOI: 10.1038/s43247-022-00498-3.
    Read source 20
21. Fyfe, J., Fox-Kemper, B., Kopp, R., and Garner, G. (2021). AR6 WGI Figure SPM.8 data, version 20210809. CEDA. DOI: 10.5285/98af2184e13e4b91893ab72f301790db. Panel a assessed temperature curves.
    Read source 21
22. Foster, G., and Rahmstorf, S. (2026). Global Warming Has Accelerated Significantly. Geophysical Research Letters, 53, e2025GL118804. DOI: 10.1029/2025GL118804.
    Read source 22
23. Diebold, F. X., and Rudebusch, G. D. (2023). Climate models underestimate the sensitivity of Arctic sea ice to carbon emissions. Energy Economics, 126, 107012. DOI: 10.1016/j.eneco.2023.107012. Author-hosted full paper, Tables 1 and 2.
    Read source 23
24. Wang, Q., Li, W., Shan, L., and Liu, X. (2026). Observationally-constrained projections of September ice-free Arctic by selecting optimal fingerprint strategy. Climate Dynamics, 64, 109. DOI: 10.1007/s00382-026-08088-1.
    Read source 24
25. IPCC (2021). AR6 Working Group I Technical Summary: Arctic September ice assessment.
    Read source 25
26. Stroeve, J., et al. (2007). Arctic sea ice decline: Faster than forecast. Geophysical Research Letters, 34, L09501. DOI: 10.1029/2007GL029703. NSIDC’s original research explanation.
    Read source 26
27. Rahmstorf, S., et al. (2007). Recent climate observations compared to projections. Science, 316, 709. DOI: 10.1126/science.1136843. Accompanying author explanation of the sea-level comparison.
    Read source 27
28. IPCC (2022). AR6 Working Group II Summary for Policymakers: impacts, adaptation and vulnerability.
    Read source 28
29. Kim, Y.-H., et al. (2023). Observationally-constrained projections of an ice-free Arctic even under a low emission scenario. Nature Communications, 14, 3139. DOI: 10.1038/s41467-023-38511-8. Complementary sea ice research examined in this project.
    Read source 29

Exact inputs for the temperature comparisons

In the public OldModels repository, use data/raw/model_temp_data/Model timeseries.xlsx, sheet Individual papers. The selected columns are FAR_EBM_Best_T, SAR_EBM_Best_T, TAR_EBM_Best_T and AR4_MMM_Best_T. The AR4 line is the A1B multi-model mean. For AR5, use references/Model data spreadsheet.xlsx, sheet IPCC AR5, column T blended RCP4.5. These are representative archived curves, not every report’s full assessed distribution.

Use the NASA annual J-D values, not monthly values or incomplete 2026 data. Differences use the same two years. Trends use all annual values in each inclusive window. Data revisions may slightly change a later reproduction; the frozen observation values below preserve the numerical inputs used here.

Reproduce a trend and its percentage

Calculation:
For year x and temperature T:
Slope per year = Σ[(x − mean x) × (T − mean T)] / Σ[(x − mean x)²]
Slope per decade = 10 × slope per year
Fit projected and observed series separately over the same inclusive years.
Shortfall = 100 × (observed slope − projected slope) / observed slope
Uplift = 100 × (observed slope / projected slope − 1)

For the smoothing check, compute five-year trailing averages independently for both series, then subtract the endpoint averages. For the future table, use the official Figure SPM.8 panel-a CSV files, version 20210809, columns Year, Mean, 5% and 95%. The data archive is linked in source 21. Preserve numeric precision until final display.

Sea ice units require care: the regression converts million km² per teratonne CO2 to m² per tonne CO2, giving the same numeric slope. The Table 2 biases are signed additions to the observed negative slope. We calculate the ratio of loss magnitudes only after recovering the model slope. Arctic area weighting uses a spherical cap north of 66.5°N.

21. From forecast arguments to better decisions

A useful next step is to choose one real decision—heat preparation, crop planning, infrastructure design, or coastal adaptation—and specify the regional metric, scenario, and threshold it depends on. Then test whether the plan remains workable under the relevant higher case. Accurate labels make precaution stronger, because others can inspect what you assumed.

For readers using Job One for Humanity, continue to the Job One climate action and preparation plan. That is an organizational action resource, separate from the scientific evidence and calculations cited here.

Return to the executive summary · Review the global calculation · Review LAND · Review sea ice

---

Attribution and disclosure: Prepared for the IPCC Forecast Accuracy project with AI-assisted research synthesis, source checking and arithmetic. The original scientific findings belong to the cited researchers and institutions. The new calculations are exploratory; the conditional adjustments are disclosed planning tests. This article is not an IPCC publication or a peer-reviewed calibration. Research cutoff: October 12, 2026. Complete annual observational year: 2025.