Tipping Point Detection

Speculative If critical slowing down could forecast a real-world tipping date with skill above a naive baseline — which no published test has yet shown, and section 7 says what would — the prize is lead time on a decision worth trillions of dollars and taken only once. The decisions that wait on this instrument are the largest in Climate Overshoot and Lock-In: when, or whether, to deploy the emergency measures in Climate Engineering to shave a peak, and when to treat a slow circulation as a committed loss rather than a fluctuation. A warning that arrives a decade early is worth a different response than one that arrives the year of collapse.

Frontier The magnitude is set by what an Atlantic overturning collapse would do, and it is large in the brief’s own units. The one full-complexity demonstration of a modelled collapse puts winter cooling across Britain and Scandinavia at between 10 °C and 30 °C within a century, arriving as a different climate within a decade or two once tipping begins. Against that, the entire value of detection is the interval between a reliable alarm and the crossing — the window in which mitigation, adaptation finance or a deliberate intervention can still change the outcome. Today that interval is unmeasured, because no indicator has been scored on a forecast it made before the fact.

Established What the machinery already delivers is real and narrower than the headlines: it detects loss of resilience in models and in the historical record after the fact. A review of 229 empirical studies from 2004 to 2022 found rising autocorrelation and variance reported as a positive early-warning result in 67.8 per cent of cases — a genuine and reproducible signal, almost all of it measured in hindcast, on systems already known to have tipped. That is a diagnostic, not yet a forecast, and the gap between the two is the whole subject of this brief.

Speculative The upside is distributed unevenly, and the cost of getting it wrong lands on people who did not cause the risk. A skilful warning would benefit North Atlantic populations most directly, but the interventions it would license — sustained solar radiation modification, crash mitigation — carry regional consequences borne globally, so a false alarm is not free. The condition on all of it is a single missing number: the forward skill of these indicators against a baseline that guesses from the climatology alone. Section 5 is a list of the reasons that number does not yet exist.

An early-warning signal is a statistical fingerprint that a dynamical system leaves as it approaches a tipping point. The core idea is critical slowing down: near a fold bifurcation the feedbacks that pull a system back to its current state weaken, so it recovers more sluggishly from small perturbations, and that sluggishness shows up as rising lag-1 autocorrelation and rising variance in whatever time series you can measure. Established The theory is textbook and the signal is real in models: drive a model system toward a fold and its autocorrelation climbs toward one. The founding review set this out in 2009, and paleoclimate reconstructions of eight past abrupt transitions showed the slowing-down pattern before each event when analysed with the answer already known.

Frontier This brief owns the gap between that success and the claim policy actually needs, which is a forecast. Detecting critical slowing down in a model, or in an ice-core record of a transition that happened twelve thousand years ago, is a different act from telling a government in 2026 whether the Atlantic overturning circulation will cross a threshold this century. The first has been done many times; the second has never been scored against a fair baseline. The indicators were validated mostly backwards, and every climate-intervention decision that leans on them is leaning on forward skill nobody has measured.

Frontier The subject is not abstract because three high-profile results now disagree, loudly, on overlapping data. A statistical analysis put an Atlantic overturning collapse in a 95 per cent window of 2025 to 2095, central estimate around 2057; a physics-based indicator concluded the same circulation is “on tipping course” without naming a date; and a 2025 assessment argued that the warning signals are fundamentally ambiguous, appearing with and without genuine tipping points. Frontier That spread — a specific date, a directional warning, and a caution that the whole method may be reading noise — is what an unvalidated instrument looks like when three competent teams point it at the same ocean.

Established A note on sourcing. This brief was commissioned in September 2026. Reading-list entries without links are cited from the bibliographic record rather than re-fetched; several primary papers sit behind publisher walls this session could not open and are cited by name. Where a linked source and a fetch summary disagreed on a number — the collapse window is the case that mattered — the figure carried here is the one the paper’s authors and independent commentary both report.

Established Critical slowing down is a genuine property of systems near a fold bifurcation, and its statistical signatures are correct. As the dominant eigenvalue of the linearised system approaches zero, the return time after a perturbation diverges, which raises the autocorrelation at lag one toward unity and inflates the variance. This is not disputed. It is analytic in the simple models, it is reproducible in the laboratory — yeast populations, cyanobacteria, thermoacoustic tubes all show it on cue — and it is what the whole field rests on.

Established The catalogue of indicators is large, and its performance in the literature is strongly positive and strongly biased. A 2024 synthesis in Earth System Dynamics by Dakos, Boulton, Buxton, Abrams and colleagues reviewed 229 empirical papers published between 2004 and 2022, reclassified 65 distinct early-warning indicators, and found 67.8 per cent reported positive performance, 24.7 per cent mixed, 4.8 per cent inconclusive and only 3.4 per cent negative. Frontier The authors do not present that as vindication. They note that most studies were “restricted to hindsight application, meaning that an a priori knowledge of a tipping point may introduce bias towards detecting CSD indicators,” and that the documented publication bias against negative results almost certainly applies. A 96.6 per cent non-negative rate in a field that mostly grades its own homework after seeing the answer is not the same as a 96.6 per cent forecast rate.

Frontier The distinction the field turns on is hindcast versus out-of-sample forecast, and almost the entire evidence base is the first kind. A hindcast analyses a record of a transition that already happened and asks whether the signal was present beforehand; the analyst knows there was a tipping point, knows roughly when, and chooses the window, the detrending and the proxy accordingly. A forecast commits to a claim — this system is or is not approaching a threshold, on roughly this schedule — before the outcome is known, and is then graded on it. The corpus of climate early-warning studies is overwhelmingly hindcast, and the prospective skill of the indicators, scored on records truncated before the outcome, is essentially unmeasured.

Frontier The clearest cautionary result in the field is a paleoclimate null, and it predates the current excitement. Ditlevsen and Johnsen, analysing the Dansgaard–Oeschger abrupt warmings in Greenland ice cores in 2010, found that critical slowing down was not reliably detectable before the actual transitions — a paper whose subtitle, “early warning and wishful thinking,” states the worry precisely. Frontier If the abrupt events best recorded in the paleoclimate archive do not announce themselves in the indicators, the assumption that a modern tipping element must do so is a hypothesis, not a background fact.

Frontier The Atlantic overturning circulation is where the disagreement is sharpest, and the observational record is far too short to settle it. Continuous, direct measurement of the overturning began with the RAPID array at 26.5 °N in 2004 and the subpolar OSNAP array in 2014; before that there are a handful of cross-Atlantic ship sections back to the 1950s. Everything longer is a proxy. A circulation with modes on centennial timescales cannot have its trend separated from its decadal wobble on a two-decade instrumental record, and this is the constraint under every AMOC warning, on all sides.

Frontier The statistical warning: Ditlevsen and Ditlevsen, 2023. Writing in Nature Communications, they built an AMOC “fingerprint” from subpolar-gyre sea-surface temperature in the HadISST dataset over 1870 to 2020, subtracting a multiple of the global-mean warming trend, fitted a stochastic model of a saddle-node bifurcation, and extrapolated the rising variance and autocorrelation to a critical point. Frontier They estimated a collapse around the year 2057 under continued emissions, with a 95 per cent window of 2025 to 2095. Their own stated caveats are substantial: a collapse might be only partial; other mechanisms could be at play so the true uncertainty is larger; noise-induced tipping could arrive before the deterministic threshold; and the freshwater forcing that is the real control parameter is “not well constrained.”

Frontier The objections to that date were immediate, specific and from senior oceanographers. On the proxy: Penny Holliday of the National Oceanography Centre said subpolar-gyre sea-surface temperature “is not a clear indicator of the state of the AMOC” and that the trend-subtraction used “is not the usual approach”; Jochem Marotzke of the Max Planck Institute said there is “considerable doubt as to whether surface temperature measurements are a valid proxy.” Frontier On the model: Marotzke noted the essential equation relies on simplified bifurcation models “also being correct” while the comprehensive models do not show that bifurcation, and Jonathan Bamber pointed to recent studies in which the AMOC responds more smoothly. On the extrapolation: Niklas Boers judged the model assumptions “heavily oversimplified” and the dataset uncertainties large enough to make the extrapolation “far too uncertain.” On confounders: Ben Booth of the Met Office noted that volcanic and industrial aerosols project onto the same historical record and were neglected. None of these is a claim that the AMOC is safe; they are claims that the date is not supported by the method.

Frontier The physics-based warning: van Westen, Kliphuis and Dijkstra, 2024. In Science Advances they ran a high-resolution global climate model for roughly 4,400 simulated years under slowly rising freshwater forcing, watched the AMOC collapse, and identified a physically grounded precursor — the minimum of the AMOC-induced freshwater transport at the southern boundary of the Atlantic, near 34 °S — that turns negative before the collapse. Frontier Applying the same diagnostic to reanalysis products, they concluded the real AMOC sits in the regime the model associates with being on a tipping course. The result is stronger than a statistical fingerprint because it rests on a known salinity-advection feedback rather than on curve-fitting, and weaker as a forecast because it deliberately declines to give a date: the authors say more observations at 34 °S are needed before any date is proposed, and the freshwater input that tips the model is far larger than present Greenland melt.

Frontier The 2025 assessment: the signals themselves are ambiguous. Rietkerk, Skiba, Weinans, Hébert and Laepple, in Nature Climate Change, argued that early-warning signals “can occur with and without tipping points” — that a rising autocorrelation or variance trend can be produced by non-stationary noise or by a smooth, non-tipping change in forcing, so the signal is not diagnostic of an approaching bifurcation on its own. Frontier They propose a Bayesian framework to quantify the probability that a tipping point is being approached, and argue for “better testing of the existence of tipping points, beyond the application of” the signals. This is the methodological centre of gravity in 2025: not that the AMOC is fine, but that the instrument as commonly used cannot distinguish the case that alarms us from several that should not.

Established The assessed institutional position remains cautious and is not the same as any of the three headlines. The last IPCC assessment judged a decline in the AMOC over this century very likely, and an abrupt collapse before 2100 not likely, at medium confidence. Frontier The gap between “not likely before 2100, medium confidence” and “central estimate 2057” is not a gap in the ocean; it is a gap in what the available instruments can support, and it is the reason the decisive experiment in this brief is a test of the instrument rather than another look at the sea.

Frontier Other tipping elements show the same pattern: a real hindcast signal, no scored forecast. Boulton, Lenton and Boers reported in 2022 that more than three-quarters of the Amazon rainforest has lost resilience since the early 2000s, measured as rising autocorrelation in satellite vegetation indices; Boers and Rypdal reported in 2021 that critical slowing down suggests the western Greenland ice sheet is close to a critical transition. Frontier Both are serious observational results and both are diagnoses of declining stability, not forecasts of a dated transition — and both inherit the same open question of whether a resilience-loss signal in a driven, noisy, non-stationary system reliably precedes an actual tip.

Frontier The organising open question is whether critical slowing down carries forward predictive skill at all in real climate records, as distinct from diagnostic power in hindcast. Every result in section 3 is consistent with two very different worlds: one in which the indicators genuinely lead real transitions and were simply never graded prospectively, and one in which they fit past transitions because the analyst knew where to look. Nothing published distinguishes those worlds, and the distinction is the entire practical value of the field.

Frontier The second question is whether the tipping being warned of is even the kind the theory covers. Critical slowing down is a signature of bifurcation-induced tipping, where a control parameter drifts slowly across a fold. It is not guaranteed for noise-induced tipping, where a large fluctuation kicks the system across a still-intact barrier, nor for rate-induced tipping, where a system fails to track a threshold that is moving too fast. Speculative If a real tipping element goes by the second or third route — plausible for an ocean forced by an accelerating meltwater pulse — the warning signal may be weak or absent right up to the transition, which is the worst possible failure mode for a warning system.

Frontier The third is the proxy problem, which is specific to the AMOC and unresolved. There is no agreed observable that tracks overturning strength over the long record. The sea-surface-temperature fingerprint is contested; the freshwater-transport indicator is physically motivated but sparsely observed at the latitude that matters. Frontier A method that produces a different answer depending on which unvalidated proxy it is fed is not yet a measurement of the ocean; it is a measurement of the proxy.

Frontier The fourth is whether machine learning changes the picture or merely moves the bias. A deep-learning classifier trained on hundreds of thousands of simulated bifurcations outperforms classical autocorrelation-and-variance indicators on out-of-sample test sets and can name the type of bifurcation approaching. Speculative Whether that transfers to a real climate record it was not trained on is open, because the network learns the bifurcations of the models it was shown, and a follow-up showed its verdicts are sensitive to preprocessing choices. A better hindcaster is not automatically a forecaster.

Frontier The fifth is what a warning is worth given its lead time, which nobody has quantified because the lead time is unmeasured. Rahmstorf’s standing objection is that “by the time they can provide a reasonably reliable warning of an impending AMOC tipping, it will be too late to prevent it” — the signal-to-noise ratio only becomes convincing close to the threshold, by which point the committed change may already be locked in. Frontier A warning system whose reliable alarms all fire after the point of no return is an epitaph, not an early warning, and whether these indicators are that is an empirical question section 7 can answer.

Frontier The binding constraint is the absence of a single number: the out-of-sample forecast skill of these indicators against a fair baseline. The field has 229 papers and 65 indicators and no scoreboard. Nobody can state, for critical slowing down applied to the AMOC, the fraction of prospective calls that were correct, the false-alarm rate, or the improvement over guessing from the climatology — because the calls have not been scored prospectively. Until that number exists, every downstream use is an extrapolation from hindcast success.

Frontier The second is the observational record length, which cannot be bought with money on the timescale that matters. The continuous overturning arrays are two decades and one decade old against a circulation with centennial modes. Established Sustaining RAPID and OSNAP is a funding-continuity problem in the low tens of millions of dollars a year, and it is worth solving, but it does not produce a century of data in under a century. This is the constraint that makes the model side of the decisive experiment necessary: the real ocean will not supply enough out-of-sample tests in time.

Frontier The third is the proxy: there is no validated observable for AMOC strength over the historical period. The disagreement between the sea-surface-temperature fingerprint and the freshwater-transport indicator is not a detail; it is the reason two competent analyses of the same era reach different conclusions. A validated proxy, cross-checked against the direct arrays during their overlap, is a prerequisite for any statistical warning to mean anything.

Frontier The fourth is model structural uncertainty: the comprehensive models mostly do not reproduce the very bifurcation the statistical method assumes. Several senior modellers noted that the CMIP-class models do not show a clean AMOC fold, and that many do not even reproduce the observed North Atlantic “cold blob.” Frontier If the models used to generate training data and to define what a tipping trajectory looks like disagree with each other on whether the fold exists, the ground truth for any forecast test is itself contested.

Frontier The fifth is confounding by forcing. Aerosols, volcanic eruptions and the global warming trend all leave marks in the same records used to detect slowing down, and a rising-variance signal can be manufactured by a change in the forcing that has nothing to do with an approaching fold. Speculative Disentangling a resilience-loss signal from a forced-response signal is unsolved for the AMOC, and the 2025 ambiguity result is, in part, a formal statement of exactly this bottleneck.

Frontier The result this topic most needs is a prospective skill score, and it depends on an archive of model runs that tip and model runs that do not. A forecast test needs ground truth: ensembles in which the true answer — tips, or does not, and when — is known to the scorer but withheld from the method. The van Westen quasi-equilibrium runs and the wider model literature contain such trajectories; what is missing is a curated, shared archive built for scoring rather than for illustration.

Frontier Second, it depends on a validated AMOC-strength proxy, which depends in turn on the continued overlap of the direct arrays with the proxy record. Every year RAPID and OSNAP run alongside the sea-surface and reanalysis products is a year of cross-calibration; cutting the arrays does not just stop future data, it forecloses the validation of the entire proxy record behind them.

Frontier Third, it depends on methods that can separate forced responses from resilience loss. This is a signal-processing dependency the ambiguity paper turns into a research programme: a Bayesian or otherwise principled framework that assigns a probability to “approaching a tipping point” rather than reading a significant trend as a yes. Until detrending and forcing-removal are principled rather than chosen, the false-alarm rate is unknowable.

Frontier Fourth, it depends on decision frameworks that can act on a probabilistic, wide-windowed warning. A 95 per cent window of seventy years is not actionable in the way a weather forecast is, and the machinery that would consume such a warning — the observation-to-decision chain in Earth-System Digital Twins, and the intervention-scheduling logic in Climate Engineering — is not built to ingest deep uncertainty of this shape. The warning and the decision framework have to be co-designed or the warning is inert.

Frontier The decisive test is a pre-registered, out-of-sample forecast tournament: critical-slowing-down and deep-learning indicators run on withheld model runs and on observational records truncated before the outcome is known, scored against a naive baseline that guesses from the climatology alone. Frontier This is the one result that would most change this brief’s assessment, because it converts the field’s central claim from an appeal to hindcast success into a number: the fraction of forward calls that were right, the false-alarm rate, and the margin over guessing. If the indicators beat the baseline by a wide margin on data they were denied sight of, forward skill is real and the AMOC warnings deserve their weight; if they do not, the field is a diagnostic discipline wearing a forecaster’s clothes.

Frontier The tournament needs no new hardware and could be run within a year, which is why its absence is the finding. Established The model ensembles that tip and do not tip already exist; the deep-learning early-warning code and its half-million-run training set are openly archived; the classical indicators are a few lines of code. Frontier The only missing ingredients are pre-registration — fixing the method, the windows and the scoring before seeing the outcomes — and a neutral scorer holding the answers. That it is a coordination problem rather than a science problem, and still has not been done, is the strongest single piece of evidence that the field’s confidence has outrun its validation.

Speculative The truncated-record arm is the part that bites, and the paleoclimate archive is where it can be run today. Take the ice-core and sediment records of past abrupt transitions, hide everything after a cutoff, and ask each method to call whether a transition is coming and when — then reveal the record. Frontier The 2010 Dansgaard–Oeschger null suggests this arm may be unkind to critical slowing down, which is precisely why it is the honest test rather than another hindcast on a curated event.

Frontier The natural experiment on the real AMOC is already running, and its limitation is time, not instruments. Established The RAPID array at 26.5 °N and OSNAP in the subpolar gyre are accumulating exactly the record a real out-of-sample test would need. Frontier The problem is that the record grows one year per year, and the freshwater-transport indicator that matters most is measured worst at 34 °S; work on optimal observation locations for AMOC early warning is the sensible response, but it buys resolution, not decades.

Speculative A cheap institutional experiment would settle a related question: pre-register the falsifier for each published warning. No AMOC warning to date states, in advance, the observation that would retire it — an array value, a proxy trajectory, a date passed without collapse. Frontier A registry of dated, falsifiable predictions from each method, scored as the years pass, would turn a decade of argument into a decade of evidence at essentially no cost, and its absence is itself a result about how the field operates.

Speculative The result that would break the tie between the three headlines is a forcing-attribution test on the fingerprint itself. Feed each method synthetic records built from known aerosol, volcanic and greenhouse forcings with no tipping point, and measure how often it raises a warning anyway. Frontier A high false-positive rate on forced-but-stable records would vindicate the 2025 ambiguity argument directly; a low one would rehabilitate the statistical warnings. Either way it is a keystroke experiment that nobody has published a scored version of.

Established The observing requirement is continuity, not novelty: keep the arrays running and extend them where the physics says the signal lives. RAPID at 26.5 °N since 2004 and OSNAP since 2014 are the backbone; the engineering ask is sustained funding and a southern-boundary component near 34 °S to measure the freshwater-transport indicator directly rather than inferring it. This is moored-array and glider engineering that exists; the failure mode is budgetary, and it feeds the same observation-to-decision chain that Earth-System Digital Twins identifies as where these systems break.

Frontier The computational requirement is a scoring infrastructure, not a bigger model. An out-of-sample tournament needs a curated archive of tipping and non-tipping model runs, a held-out test protocol, and a neutral registry — the software equivalent of a weather-forecast verification bureau, applied to tipping. Established The pieces are cheap and mostly written; the deep-learning training set is archived and the classical indicators are trivial to implement. What does not exist is the institution that holds the answers and grades the calls.

Frontier The hardest engineering requirement is the interface between a wide-windowed probabilistic warning and an irreversible decision. A warning of the form “collapse likely within a seventy-year window” has to be turned into a schedule for the interventions in Climate Engineering and the authorisation pathways in Geoengineering Governance, which are the only tools that could act on the warning at all. Speculative Designing a trigger — what indicator value, sustained for how long, licenses what response — is an unsolved control problem, and getting it wrong in either direction is expensive: a false trigger deploys planetary-scale intervention against a fluctuation, a missed one wastes the lead time the warning was supposed to buy.

Frontier The nearest neighbour is the overshoot question, which presupposes an answer this brief owns. Climate Overshoot and Lock-In nominates sustained AMOC observation together with the freshwater-transport fingerprint as its own decisive test, and treats the reversibility of the circulation as the largest open variable in its accounting. This brief is the instrument-validation layer under that: whether the fingerprint can warn in time is exactly the question overshoot planning cannot answer for itself.

Frontier On the intervention side the dependency is direct and asymmetric. Climate Engineering establishes that the physics of cooling is the least contested part and the governance the most, and a tipping warning is the most plausible trigger anyone has proposed for emergency deployment; Geoengineering Governance establishes that no authorisation pathway exists to act on such a trigger. A warning with no legitimate actor to receive it is a different object from one that feeds a decision.

Frontier The method is shared across other tipping elements, each with its own brief-worthy stakes. The Amazon resilience-loss result connects to Biodiversity Restoration, and the compounding of simultaneous extremes that a circulation change would drive is owned by Compound Climate Hazards. Frontier The machine-learning turn in detection shares its toolkit with AI Weather Prediction, and inherits the same question in a harder form: a model that skillfully predicts next week’s weather is graded daily, whereas a model that predicts a once-ever tipping point may never be graded at all.

Frontier There is no body that owns the forecast-verification question, and that is the central institutional gap. Weather prediction has verification bureaus that score forecasts continuously and publicly; tipping-point detection has journals that publish hindcasts. Frontier The result is that individual warnings compete for headlines while the underlying skill of the method goes unaudited, because no institution is charged with grading it.

Established The observing systems are governed as research infrastructure, which makes them chronically insecure. RAPID and OSNAP depend on renewed grants, not on a standing mandate, and the continuity that the warning problem most needs is exactly what the funding model least guarantees. Frontier A two-decade record interrupted for a budget cycle is worth far less than an unbroken one, and no institution treats these arrays as the permanent monitoring the science implies they should be.

Frontier The assessment bodies transmit the disagreement without resolving it, because it is not theirs to resolve. The IPCC can report a range and a confidence level; it cannot commission the out-of-sample tournament that would narrow the range, and the funders who could have not been asked to. Speculative The institution this subject actually needs — a neutral, standing scorer of dated tipping forecasts — would cost a rounding error against a single observing array and does not exist anywhere.

Frontier A live legal question is beginning to attach to the warnings. As duties of climate prevention are litigated, a published, dated collapse warning becomes a fact in the record that governments may be argued to have been on notice of. Speculative That raises the stakes on getting the skill question right in both directions: an unvalidated warning treated as actionable, and a valid warning treated as noise, now carry legal as well as physical consequences.

Frontier A false alarm and a missed alarm have radically different costs, and they fall on different people. A warning treated as real could license sustained solar radiation modification, whose regional consequences are borne globally and disproportionately in the tropics; a warning dismissed as noise could forfeit the only lead time available to prepare for a North Atlantic cooling of 10 °C to 30 °C. Frontier The asymmetry means the acceptable false-positive and false-negative rates are not a technical choice but a distributional one, and no framework currently sets them.

Frontier Communicating an unvalidated warning is itself an ethical act with a track record. The 2057 estimate travelled the world as a headline while its caveats stayed in the paper, and the gap between “central estimate 2057” and “95 per cent window 2025 to 2095 under strong assumptions the authors themselves flag” is the difference between a deadline and a very wide probability. Frontier A field that cannot yet score its forecasts has a heightened duty to carry the uncertainty into the headline, not just the methods section.

Speculative There is a mirror hazard to alarmism, which is the paralysis of demanding certainty that the method cannot provide. If the honest state is “we cannot yet forecast this,” a decision system that refuses to act without a validated forecast has quietly chosen to accept whatever tipping risk exists. Frontier The defensible position is to treat the warnings as what they are — loss-of-resilience diagnostics with unmeasured forward skill — and to fund the test that would upgrade or retire them, rather than either acting on them as forecasts or ignoring them as noise.

Frontier The civilizational question is whether humanity can acquire advance notice of a planetary-scale state change in time to do anything about it — and the current honest answer is that we do not know, and are not running the experiment that would tell us. A validated tipping forecast would be one of the few genuinely new capabilities in the human relationship with the Earth system: the ability to see a threshold coming rather than to discover it by crossing it.

Speculative If the pessimistic reading is right, the deepest implication is that some planetary thresholds are, in practice, unforecastable. If reliable warning only arrives within a few years of a crossing — the Rahmstorf case — or if the relevant tipping is noise- or rate-induced and shows no slowing down, then for the decisions that matter the warning is not early. Frontier That would push the whole enterprise back onto precaution: acting on the possibility of tipping rather than on its detection, which is a different and harder political proposition.

Speculative If the optimistic reading is right, the capability reshapes climate decision-making from a debate about long-run averages into a monitored, triggered system. A civilization that could score and trust tipping forecasts could hold interventions in reserve against a validated alarm rather than deploying or foreswearing them on ideology. Frontier The condition on that world is the boring one this brief keeps returning to: a forward skill score that beats a naive baseline, which is entirely obtainable and has not been obtained.

These horizons track what becomes knowable about the instrument, not what the ocean does.

• 10 yr: Frontier The out-of-sample tournament on model and paleoclimate records is entirely feasible this decade and would, if run, deliver the field’s first prospective skill score; the arrays reach a three-decade record long enough to begin separating trend from decadal variability. Expect the deep-learning and Bayesian methods of AI Weather Prediction to be applied to detection, and expect the proxy dispute to remain open.
• 25 yr: Speculative If forward skill is real, a validated warning with a usable lead time is plausible and would force the intervention decisions in Climate Engineering into the open; if it is not, the field consolidates as a resilience-monitoring discipline that reports declining stability without dating transitions. The 2025 to 2095 window itself begins to be tested against the arrays.
• 50 yr: Speculative Enough of the Ditlevsen window will have elapsed to grade the 2023 forecast directly — a collapse, a partial weakening, or a passed date — which is the first real out-of-sample verdict on a statistical AMOC warning, delivered the slow way.
• 100 / 250+ yr: Handwave Claims that early-warning systems will routinely forecast planetary tipping with actionable lead time rest on forward skill that has never been demonstrated and on observing continuity longer than any array has been maintained; coherent as an aspiration, unsupported as a projection.

• Depends on the observation-to-decision chain in Earth-System Digital Twins, which owns the modelling pipeline any warning must travel, and the reversibility framing in Climate Overshoot and Lock-In, which nominates the AMOC fingerprint as its own decisive observation; nothing else on the map blocks this topic, because its missing pieces are results and institutions nobody is producing.
• Requires (not on this map) a prospective, out-of-sample skill score for the indicators against a naive baseline, without which every warning is an extrapolation from hindcast; a validated observable for AMOC strength cross-checked against the direct arrays, without which the statistical warnings measure the proxy rather than the ocean; a curated archive of model runs that tip and do not tip, which is the ground truth a forecast test is scored against; a neutral registry that holds dated forecasts and grades them as the years pass, which no institution currently maintains; and a continuous multi-decadal observing array, which grows only one year per year and cannot be accelerated with money.
• Enables triggered rather than scheduled climate intervention, falsifiable overshoot pathways with stated retirement conditions, legally defensible notice of a committed change, and adaptation finance timed to a validated warning rather than to a scenario average.
• Adjacent to Climate Engineering and Geoengineering Governance on what a warning would trigger and who could authorise it, Compound Climate Hazards on the consequences a circulation change would compound, Biodiversity Restoration on the Amazon resilience signal, and AI Weather Prediction on the shared machine-learning toolkit and the harder grading problem.

Handwave “Early-warning signals can predict when a tipping point will happen.” No published method has been scored on a forecast it made before the outcome was known, and the developers of the leading deep-learning system state plainly that “none of these approaches can predict exactly when a transition will occur.” The indicators detect declining resilience; converting that into a date requires assuming a specific bifurcation and extrapolating a fitted curve, which is where the confidence is manufactured.

Frontier “The 2023 study proved the AMOC will collapse around 2057.” It produced a central estimate of 2057 with a 95 per cent window of 2025 to 2095, conditional on a contested proxy, a simplified bifurcation model and an extrapolation that senior oceanographers judged “far too uncertain,” and neglecting aerosol forcing that marks the same record. It is a serious warning that the method cannot support as a date.

Frontier “A physics-based indicator is a forecast, so van Westen settled it.” The freshwater-transport indicator is a genuine advance because it rests on a known feedback rather than curve-fitting, and its authors deliberately decline to give a real-world date, calling for more observations at 34 °S first. “On tipping course” is a statement about a regime, not a schedule, and reading it as a schedule is the error the authors themselves refuse.

Frontier “A rising early-warning signal means a tipping point is near.” The 2025 assessment shows the signals “can occur with and without tipping points” — non-stationary noise and smooth forced changes both produce rising autocorrelation and variance. A significant trend in an indicator is evidence to be weighed in a probabilistic framework, not a detector that has gone off.

Handwave “Hindcast success on paleoclimate transitions validates the method for forecasting.” Hindcast success is real and is not forecast skill; the same paleoclimate archive contains the 2010 Dansgaard–Oeschger null in which slowing down was not reliably detectable before the actual events. A method that fits past transitions when you know where to look has not thereby earned the right to call the next one.

Speculative “Better machine learning will solve the forecasting problem.” A deep-learning classifier does outperform classical indicators out-of-sample and can name the bifurcation type, but it learns the bifurcations of the models it was trained on, a follow-up found its verdicts sensitive to preprocessing, and transfer to a real climate record it never saw is exactly the untested step. A better hindcaster is not automatically a forecaster.