Assessing ExxonMobil’s global warming projections

We analyzed 32 internal documents produced in-house by ExxonMobil scientists and managers between 1977 and 2002, and 72 peer-reviewed scientific publications authored or coauthored by ExxonMobil scientists between 1982 and 2014. The internal documents were collated from public archives provided by ExxonMobil Corp (28), InsideClimate News (29), and Climate Investigations Center (30). The peer-reviewed publications were obtained by identifying all peer-reviewed documents among ExxonMobil Corp’s lists of “Contributed Publications,” except for three articles discovered independently during our research (31) [see supplementary materials (SM) section S2 for details on the assembly of the corpus]. These constitute all publicly available internal ExxonMobil documents concerning anthropogenic global warming of which we are aware, and all ExxonMobil peer-reviewed publications concerning global warming disclosed by the company.

Using manual content analysis, we identify all documents that reported climate model outputs of (i) a time series of projected future temperature, and (ii) future external radiative forcings [including at least atmospheric carbon dioxide (CO₂) concentration] (see SM section S1.1 for coding details). For models driven by more than one forcing time series (i.e., for high- and low-CO₂ scenarios as well as a central, “nominal” one), each resulting temperature time series is treated as a separate and individual projection. Our figures and tables therefore distinguish between “nominal,” “high,” and “low” model projections. By contrast, for a given CO₂ scenario, temperature time series accompanied by uncertainty bars (corresponding, for example, to different model climate sensitivities) are treated as single projections with uncertainty bounds given by those uncertainty bars. This yields 12 documents published between 1977 and 2003, which contain 16 distinct temperature projections presented in the form of 12 unique graphs and one table (summarized in SM section S2.2). The 12 documents comprise seven internal memos (1977 to 1985) and five peer-reviewed papers (1985 to 2003). Twelve of the 16 temperature projections came from models built or run in-house by ExxonMobil scientists, typically in collaboration with independent researchers. Once identified, all original temperature and forcing projections are converted for analysis by digitizing graphs and extracting tables.

We assess each model projection over the period from the publication year of its containing document through 2019 (or through the final projected year, if earlier). First, we overlay all original temperature time series with observed temperature changes. Observations are aligned with respect to the earliest reference year(s) for which model projection data are available and, unless noted otherwise, reflect the smoothed annual average of five historical time series. Following Hausfather et al. (2020) and the Intergovernmental Panel on Climate Change (IPCC), we compare observations to model projections in two quantitative ways: (i) change in temperature versus time; and (ii) change in temperature versus change in radiative forcing (the “implied transient climate response,” or iTCR) (16, 24). The iTCR metric enables us to assess model performance while accounting for any differences in the assumptions about future radiative forcings driving the models. For each projected and observed temperature time series, per-decade temperature changes are calculated by fitting an ordinary least squares model over the projection period and multiplying the resulting gradient coefficient by 10. Analogously, iTCR is calculated by regressing temperature against anthropogenic radiative forcing over the projection period and multiplying the result by the forcing associated with doubled atmospheric CO₂ concentrations, $F_{2x}=3.7 {\text{W/m}}^2$ (16):

For model projections, $\text{Δ}{F}_{\text{anthro}}$ was based on explicit external forcing values when provided and was otherwise estimated from model CO₂ concentration scenarios aswhere $p_{{\text{CO}}_{\text{2}}}$ is the initial CO₂ concentration (in parts per million) at the start of the projection period and $p{\text{'}}_{{\text{CO}}_{\text{2}}}$ is the CO₂ concentration during each subsequent year through 2019 (16). In the real world, of course, global temperature changes are driven by multiple natural and anthropogenic factors, including but not limited to CO₂. Nevertheless, even when model projections are driven by CO₂-only anthropogenic forcing scenarios, retrospectively comparing projections to observations offers a robust, independent, and established test of model skill. This is because (i) global warming has been almost entirely human-caused since the late 19th century (32, 33) and (ii) total anthropogenic forcing over the past 150 years has been, to first order, similar to the forcing of CO₂ alone, because the warming effects of other greenhouse gases and the cooling effects of other sources cancel one another out (34). For further discussion of the implications and limitations of model-versus-observation comparisons, see SM section S1.2.7. Observed $\text{Δ}{F}_{\text{anthro}}$ values, meanwhile, were based on a 1000-member ensemble of observationally informed forcing estimates reported by Dessler and Forster (2018) (35).

Evaluated in terms of each of the above metrics, we deem model projections and historical observations to be consistent if and only if the 95% confidence intervals of the differences between the two include zero. As detailed in SM sections S1.2.2 and S1.2.3, these confidence intervals were calculated to reflect two sources of uncertainty: (i) statistical uncertainty in regression coefficients and (ii) structural uncertainty due to different model climate sensitivities, as and when indicated by error bars in projections reported by ExxonMobil scientists.

As an additional measure of performance, we calculate the “skill score” of each model by comparing the root-mean-squared errors of a model projection with those of a zero temperature change null hypothesis (20). For each projection, we calculate skill scores with respect to (i) each of the five observational temperature records for the temperature-versus-time metric and (ii) the 5000 estimates of $\Delta T/\Delta F_{\text{anthro}}$ for the iTCR metric. (See SM section 1.2 for details on graphical overlays and on calculation of consistency and skill scores and their accompanying uncertainties.)

Overall, ExxonMobil’s global warming projections closely track subsequent observed temperature increases.

Figure 1Opens in image viewer reproduces all 12 identified unique graphs, which contain 15 of the 16 identified temperature projections (the 16th was reported as a table). For example, panel 3 of Fig. 1Opens in image viewer is a graph showing “an estimate of the average global temperature increase” under the “Exxon 21st Century Study–High Growth scenario” for CO₂. It was included in a 1982 internal briefing on the “CO₂ ‘Greenhouse’ Effect” prepared by Exxon Research and Engineering Company and circulated widely to Exxon management (36). The briefing was labeled as “proprietary information for authorized company use only.” The graph appeared a second time in an Exxon manager’s presentation on “CO₂ greenhouse and climate issues” at an internal company environmental conference in 1984 (37).

Panel 3 of Fig. 1Opens in image viewer displays one of 12 unique temperature projections (out of a total of 16 projections) that were output by models built or run in-house by ExxonMobil scientists (the 12 projections are indicated by asterisks in Figs. 1Opens in image viewer to 3Opens in image viewer and Table 1Opens in image viewer). To our knowledge, the temperature projection in panel 3 was independently produced by Exxon scientists as part of “technology forecasting activities in 1981” operated by the company’s Corporate Planning Department (37). The temperature projection was based on “calculations” of future atmospheric CO₂ concentrations “recently completed at Exxon Research and Engineering Company” (36). The remaining 11 temperature projections were produced by models developed by ExxonMobil scientists in collaboration with academic coauthors. Specifically, the seven unique temperature projections shown in panels 5 to 7b in Fig. 1Opens in image viewer derived from a one-dimensional upwelling-diffusion Energy Balance Model to study how the “climatic transient response from fossil fuel burning is damped…by heat storage in the world’s oceans…” (38). The Exxon scientist leading the collaboration internally described their climate modeling as “sophisticated” and “state of the art” (39). The remaining four unique temperature projections (three in panels 8, 11, and 12 in Fig. 1Opens in image viewer and the fourth designated by “9” in Fig. 2Opens in image viewer) were generated by an “Integrated Science Model which consists of coupled modules for carbon cycle, atmospheric chemistry of other trace gases, radiative forcing by greenhouse gases, energy balance model for global temperature, and a model for sea level response” (40).

In Fig. 1Opens in image viewer, we overlay the original graphs with observed atmospheric CO₂ concentrations and temperature changes, shown in blue and red, respectively. In general, observations closely track projections.

In Fig. 2Opens in image viewer, we digitize all of ExxonMobil scientists’ temperature projections corresponding to “nominal” (i.e., central) CO₂ scenarios in all 12 graphs (and one table). These projections, shown in gray, are plotted from the observed temperature change, shown in red, at the start of each projection period. The darkness of the projection lines scales with their start years, from 1977 (lightest gray) to 2003 (darkest gray). Solid gray lines indicate projections modeled by ExxonMobil scientists themselves, whereas dashed gray lines indicate projections reproduced from third-party peer-reviewed papers. With the exception of the earliest projection (designated by “1”), which overestimated future warming, projections lie close to and evenly distributed around observations.

In Fig. 3AOpens in image viewer, we compare trends in temperature change versus time for historical observations (in red) and for all 16 projections reported by ExxonMobil scientists (in gray or black). Over the course of their respective projection periods (indicated in blue boxes at the top of each panel in Fig. 3Opens in image viewer), the average predicted warming was 0.20° ± 0.04°C per decade. Ten of the 16 projections are consistent with historical observations (differences between models and projections are shown in fig. S1A). Of the remaining six projections, two forecast more warming than observed and four forecast less. Treating each unique graph and table—rather than each forcing scenario—as independent, 10 out of the 12 unique projection datasets are consistent with observations. Of the remainder, one forecasts more warming than observed and one forecasts less. Notably, these two projections are among the only three (out of 12) that were reported without uncertainty bars. They therefore have less “room for uncertainty” in our consistency tests. Overall, the models perform very well.

When we account for mismatches between forecast and observed forcings by using the iTCR metric, 12 of the 16 projections reported by ExxonMobil scientists are consistent with observations. Figure 3BOpens in image viewer uses the iTCR metric to compare trends in observed and projected iTCRs, and fig. S1B shows their differences. Treating each unique graph and table as independent, 9 out of 12 datasets are consistent. The three outliers forecast more warming than observed; two of them do not have uncertainty bars.

We also calculate skill scores for the temperature-versus-time and iTCR metrics (Table 1Opens in image viewer). A skill score of 100% indicates perfect agreement between projections and observations; a score between zero and 100% indicates some degree of skill; and a score less than zero indicates a performance worse than a zero-change null hypothesis (16, 20).

With respect to temperature change versus time, we find the average of the median skill scores of all 16 reported projections to be 67 ± 7%. Across projections modeled by ExxonMobil scientists themselves, it is 72 ± 6%. These scores indicate highly skillful predictions. The highest-scoring projection was a 1985 peer-reviewed publication [Hoffert and Flannery (1985, nominal CO₂ scenario)], with a skill score of 99% (38). The 1982/1984 projection discussed earlier (Fig. 1Opens in image viewer, panel 3) has a skill score of 82% [although it marginally failed the consistency test (Fig. 3Opens in image viewer and fig. S1)]. Only three of the 16 projections have skill scores below 50%. For comparison, NASA scientist James Hansen’s global warming predictions presented to the US Congress in 1988 have been found to have skill scores ranging from 38 to 66% across the three different forcing scenarios that he reported (16, 20).

Using the iTCR metric, the average skill of the 16 projections is 67 ± 9%. Among projections modeled by ExxonMobil scientists themselves, it is 75 ± 5%. Seven projections score 85% or above. Hoffert and Flannery (1985, high CO₂ scenario) is again the highest scorer (92%), closely followed by two projections scoring 90%, which are featured in three internal reports in 1982/1984 and 1985, respectively (38, 39, 41, 42). Only four projections have skill scores below 50% for the iTCR metric. Again, for comparison, Hansen’s 1988 projections had skill scores in terms of the iTCR metric ranging from 28 to 81% (16).

We can compare these metrics with Hausfather et al. (2020), who calculated the average skill scores of 18 academic and government climate model projections published between 1970 and 2007. They obtained a value of 69% for both temperature-versus-time and iTCR metrics (16). On average, therefore, global warming projections reported by ExxonMobil scientists were as skillful as those of independent scientists of their day, and their own models were especially skillful. (As described earlier, ExxonMobil scientists did not simply rerun existing models; they developed their own models, typically in collaboration with academic coauthors, which independently corroborated the findings of other climate scientists.) To the extent that these projections represented contemporary knowledge of the likely effects of fossil fuel burning on global temperature, we can conclude that Exxon knew as much in the 1970–1990s as academic and government scientists knew. The average warming projected by the 18 academic and government models was 0.19° ± 0.03°C per decade, which is, within uncertainty, the same as ExxonMobil’s average of 0.20° ± 0.04°C per decade.

We note that 2 of the 18 projections analyzed by Hausfather et al. (2020) are among those reported by ExxonMobil scientists. However, excluding these two projections has negligible effect on the average warming predicted by ExxonMobil or on the average skill scores of all ExxonMobil projections with respect to both temperature change versus time and iTCR (see sensitivity analyses, SM section S1.2.5 and table S1). Our conclusions also hold true when considering only the 12 (of 16) temperature projections from models built or run in-house by ExxonMobil scientists, indicated by asterisks in Figs. 1Opens in image viewer to 3Opens in image viewer and Table 1Opens in image viewer (see SM section S1.2.5 and table S1).

In summary, climate projections reported by ExxonMobil scientists between 1977 and 2003 were accurate and skillful in predicting subsequent global warming. Some projections suggested slightly too much warming and others not quite enough, but most (63 to 83%, depending on the metric used) were statistically consistent with subsequently observed temperatures, particularly after accounting for discrepancies between projected and observed changes in atmospheric CO₂ concentrations. ExxonMobil’s projections were also consistent with, and as skillful as, those of academic and government scientists. All told, ExxonMobil was aware of contemporary climate science, contributed to that science, and predicted future global warming correctly. These findings corroborate and add quantitative precision to assertions by scholars, journalists, lawyers, politicians, and others that ExxonMobil accurately foresaw the threat of human-caused global warming, both prior and parallel to orchestrating lobbying and propaganda campaigns to delay climate action (1, 2, 10, 11, 13, 43–48), and refute claims by ExxonMobil Corp and its defenders that these assertions are incorrect (49).

Our findings about the company’s early understanding of climate science contradict many of the claims that the company and its allies have made in public.

We gain additional insights into how ExxonMobil misled the public and other stakeholders by further evaluating the company’s climate projections and comparing them to its public communications.

The substantial body of literature documenting the history of climate lobbying and propaganda by fossil fuel interests has been described as a “vast blind spot” of major climate assessments—ignored, in particular, in all but the most recent IPCC assessment report (61–63). Yet bringing quantitative techniques from the physical sciences to bear on a discipline traditionally dominated by qualitative journalistic and historical approaches offers one path to remedying this blind spot. Here, it has enabled us to conclude with precision that, decades ago, ExxonMobil understood as much about climate change as did academic and government scientists. Our analysis shows that, in private and academic circles since the late 1970s and early 1980s, ExxonMobil scientists (i) accurately projected and skillfully modeled global warming due to fossil fuel burning; (ii) correctly dismissed the possibility of a coming ice age; (iii) accurately predicted when human-caused global warming would first be detected; and (iv) reasonably estimated how much CO₂ would lead to dangerous warming. Yet, whereas academic and government scientists worked to communicate what they knew to the public, ExxonMobil worked to deny it.

The authors thank Z. Hausfather (University of California, Berkeley) for technical guidance; P. Achakulwisut (Stockholm Environment Institute) for helpful discussions; and two anonymous peer reviewers.

Funding: The authors are supported by a Rockefeller Family Fund grant (G.S.) and Harvard University Faculty Development Funds (N.O.).

Author contributions: Conceptualization: G.S., S.R. Methodology: G.S., S.R. Investigation: G.S. Writing – original draft: G.S. Writing – review & editing: G.S., S.R., N.O. Visualization: G.S. Supervision: G.S., N.O. Funding acquisition: G.S., N.O.

Competing interests: The three authors have received speaking and writing fees, and S.R. and N.O. have received book royalties for communicating their research, which sometimes includes but is not limited to the topics addressed in this paper. G.S. and N.O. have offered their expertise pro bono to groups and organizations combating climate change, including briefing attorneys and coauthoring amicus briefs in climate lawsuits. N.O. has in the past served as a paid consultant to Sher Edling law firm, which has filed complaints against ExxonMobil Corp and other fossil fuel companies. However, Sher Edling played no role in this or any other study by the authors (including but not limited to study conceptualization, execution, writing, or funding).

Data and materials availability: Raw data (original PDF internal documents and peer-reviewed publications) for this study cannot be reproduced in full owing to copyright restrictions. However, a catalog of all analyzed documents, and links to public archives containing these data, are provided in SM section S2.1. Raw data resulting from digitization of all analyzed original PDF datasets are deposited on Harvard Dataverse at https://doi.org/10.7910/DVN/R4MOAE (87). The code used to generate the results of this study is provided in the same repository.

License information: Copyright © 2023 the authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original US government works. https://www.science.org/about/science-licenses-journal-article-reuse