Current atmospheric CO2 concentration is known to be higher than it has been during the past ∼800 k.y. of Earth history, based on direct measurement of CO2 within ice cores. A comparison to the more ancient past is complicated by a deficit of CO2 proxies that may be applied across very long spans of geologic time. Here, we present a new CO2 record across the past 23 m.y. of Earth history based on the δ13C value of terrestrial C3 plant remains, using a method applicable to the entire ∼400 m.y. history of C3 photosynthesis on land. Across the past 23 m.y., CO2 likely ranged between ∼230 ppmv and 350 ppmv (68% confidence interval: ∼170–540 ppm). CO2 was found to be highest during the early and middle Miocene and likely below present-day levels during the middle Pliocene (84th percentile: ∼400 ppmv). These data suggest present-day CO2 (412 ppmv) exceeds the highest levels that Earth experienced at least since the Miocene, further highlighting the present-day disruption of long-established CO2 trends within Earth’s atmosphere.

Knowledge of atmospheric CO2 concentration is vital for understanding Earth’s climate system because it imparts a controlling effect on global temperatures across recent (Hegerl et al., 2006) and geologic (Foster et al., 2017) time scales. Proxies (Breecker et al., 2010) and models (Royer et al., 2014) indicate that CO2 has varied widely during the geologic past. Direct measurement of CO2 has been performed at the Mauna Loa Observatory (Hawaii, USA) for the past 60+ yr, and historical CO2 has been sampled continuously from ice-core bubbles recording the past 800 k.y. (Petit et al., 1999; Lüthi et al., 2008), allowing for trends in CO2 during the latter portion of the Quaternary to be evaluated in detail. Direct observations of atmospheric greenhouse gases are also now available from discontinuous ice up to 2 m.y. old from East Antarctica (Higgins et al., 2015; Yan et al., 2019).

For time periods older than the Pleistocene, many CO2 proxies have been applied, including the proportion of epidermal cells that are stomatal pores (Kürschner et al., 1996, 2008; Beerling et al., 2009; Grein et al., 2013; Wang et al., 2015; Reichgelt et al., 2016); the stable carbon isotope composition of paleosol carbonate (Breecker and Cerling, 1992; Ekart et al., 1999; Retallack, 2014; Da et al., 2015, 2019); alkenones derived from marine phytoplankton (Seki et al., 2010; Badger et al., 2013a, 2013b; Zhang et al., 2013); and the pH of ocean water as derived from boron isotopes (Bartoli et al., 2011; Foster et al., 2012; Greenop et al., 2014; Martinez-Boti et al., 2015; Stap et al., 2016). Each of these proxies provides robust results for specific time periods (Foster et al., 2017; Hollis et al., 2019); however, a CO2 proxy for use across the entire history of vascular land plants (i.e., the past ∼400 m.y.) is lacking.

Here, we present a method for calculating CO2 that is based on a ubiquitous substrate, is sensitive across a wide range of CO2, and is rooted in a fully understood mechanism of response to changing CO2. We illustrate its efficacy by presenting a novel, high-resolution record of CO2 for the Neogene through the Quaternary (i.e., the past 23 m.y.), a period that lacks a continuous record of CO2 from any single proxy.

Our approach is centered upon the δ13C value of C3 vascular land plants (hereafter δ13Cp), which is available from terrestrial sediments for most of the Phanerozoic (Nordt et al., 2016). Our calculations of CO2 assumed that global changes in atmospheric composition affect the plant tissues of all terrestrial C3 plants via the universally shared mechanism of photorespiration. Because CO2 is well mixed in Earth’s atmosphere, and diminished photorespiration with increasing CO2 is fundamental to the biochemistry of photosynthesis, this mechanism is recorded globally (Keeling et al., 2017). Our previous growth chamber experiments, in combination with meta-analyses, established that the effect of CO2 on δ13Cp value is consistent across a wide range of species and environments (Schubert and Jahren, 2012, 2018). Recent works have also shown that the influence of CO2 on δ13Cp value is not affected by water availability (Lomax et al., 2019) or atmospheric O2 levels (Porter et al., 2017), and it is recorded within multiple organic substrates (e.g., cellulose and collagen [Hare et al., 2018], hair [Zhao et al., 2019], and n-alkanes [Wu et al., 2017]) and inorganic substrates (e.g., speleothems [Breecker, 2017] and cave air [Bergel et al., 2017]). Consequently, researchers now correct δ13C values for changes in CO2 across a myriad of fossil (e.g., ungulate teeth [Luyt et al., 2019; Sealy et al., 2019], soil-respired carbon [Caves Rugenstein and Page Chamberlain, 2018], soil carbonate [Basu et al., 2019], pyrogenic carbon and n-alkanes [Zhou et al., 2017], and pollen [Bell et al., 2019]), and modern (e.g., fungi [Hobbie et al., 2017] and leaves [Tibby et al., 2016]) substrates, and recent experiments have shown that the δ13Cp value can produce accurate estimates of paleo-CO2 concentration (Porter et al., 2019).

We reconstructed CO2 across the past 23 m.y. using a compilation of 700 δ13C measurements gathered from 12 previously published studies of terrestrial organic matter (TOM; n = 441) and plant lipids (n = 259) that spanned at least 1 m.y. of the Neogene (Table S1 in the Supplemental Material1). We chose these substrates because both TOM and plant lipid δ13C values have been shown to respond similarly to changes in CO2 (Schubert and Jahren, 2012; Wu et al., 2017; Chapman et al., 2019); these substrates also represent an integrated signal with multiple photosynthetic inputs, which has been shown to improve the accuracy of the proxy (Porter et al., 2019). For studies that reported δ13Cp data for multiple n-alkanes (e.g., n-C27, n-C29, n-C31), we selected only one record, or the weighted mean values (if reported), thus avoiding redundancy in our compiled data set. The δ13Cp data set used for input exhibited a large range in δ13Cp values (∼8‰), sampled from a wide range of environments; plant lipids generally exhibited lower δ13Cp values than TOM of the same age, as is commonly observed (e.g., Chikaraishi and Naraoka, 2003). We limited our literature compilation to records with δ13Cp values of TOM ≤ −22.0‰ and plant lipids ≤ −27.0‰, thus avoiding δ13Cp values that reflected C4 ecosystems (O’Leary, 1988). Less than 2% of all compiled δ13Cp values fell above these thresholds, and these were determined to be statistical outliers (all values are reported in Figure 1 and in Table S1).

The approach used here to reconstruct atmospheric CO2 concentration based on changes in δ13Cp value was first described by Schubert and Jahren (2012) and then demonstrated by Schubert and Jahren (2015). This approach calculates CO2 based on changes in δ13Cp value (i.e., δ13Canomaly) between two points in time, time t = 0 (for which CO2 is known) and time t (for which CO2 is not known):
where A, B, and C are curve-fitting parameters (A = 28.26, B = 0.22, C = 23.9; Schubert and Jahren, 2012, 2015; Cui and Schubert, 2016). When calculating δ13Canomaly, it is necessary to correct for (1) changes in the δ13C value of atmospheric CO2 between time t13Catm(t)] and time t = 0 [δ13Catm(t= 0)], and (2) any biosynthetic fractionation when comparing across different plant tissues (e.g., TOM and lipids). Therefore, δ13Canomaly represents the change in δ13Cp value, after correcting for changes in the δ13C value of atmospheric CO213Catm) and any systematic δ13Cp offset between plant tissues (ε), such that
We can then rewrite Equation 1 in order to solve for CO2 at any time t (CO2(t)), as a function of δ13Cp and δ13Catm:

Descriptions of the inputs are provided in the Supplemental Material.

Figure 1 shows a continuous record of CO2 across the past 23 m.y. based on changes in δ13Cp value (i.e., δ13Canomaly). We calculated that the median CO2 value was lower than that of today across the entirety of the past 23 m.y., and it likely never fell below levels experienced during Pleistocene glacial advances (∼170 ppm; Petit et al., 1999; Kawamura et al., 2007).

Our record commences at the start of the Neogene, when CO2 was at a local high for the entire record (∼350 ppmv; 23.0–22.4 Ma; Fig. 1C). During the middle Miocene (i.e., 17.1–15.4 Ma), CO2 reached a maximum and then steadily decreased to below the threshold for Northern Hemisphere glaciation (∼280 ppmv; DeConto et al., 2008) at the end of the Miocene. The middle Pliocene (ca. 5–3 Ma) experienced CO2 levels that might have approached early 21st century levels (∼400 ppmv; 84th percentile). This time period corresponds with elevated global temperatures as inferred from multiple models (Haywood et al., 2013), and sea levels up to 25 m higher than today (Miller et al., 2012; Grant et al., 2019). CO2 declined to near or just-below pre-industrial levels during the late Pliocene, while Northern Hemisphere glaciation increased (Balco and Rovey, 2010; Bailey et al., 2013). Low CO2 continued across the Quaternary glacial-interglacial cycles (Fig. 2) until the anthropogenic disruption in carbon cycling via the widespread use of fossil fuels (Keeling et al., 2001). Our overall record of the past 23 m.y. reveals a significant linear CO2 decline equal to an average of 5 ppmv per million years (p < 0.0001). This contrasts with an average increase of 5 ppmv per decade experienced across the past 270 yr that has more than offset the CO2 decline of the past 23 m.y.

The changes in CO2 that we have constructed are corroborated by contemporaneous changes in various Earth cycles at the sub-epoch scale. The most important change is the long-term global cooling in progress across the Neogene, as determined by Zachos et al. (2001) based on the δ18O value of foraminifera, that coincides with increased reactivity of the land surface (Caves Rugenstein et al., 2019), and our long-term decrease in CO2.

In comparing our record to the sparse data available from other proxies (Fig. 3), we see that alkenone- and stomata-based reconstructions generally estimate higher CO2 across much of the past 23 m.y., although with overlapping uncertainties, while the δ11B- and paleosol-based reconstructions do not show any consistent biases relative to our data set. In addition, the lack of continuous proxy data precludes identification of unequivocal, long-term changes in CO2 over the past 23 m.y. (Figs. 3A–3C), except perhaps for a downward trend within the data set generated using stomatal indices (Fig. 3D).

Two key intervals of the past 23 m.y. have been cited as potential analogs for anthropogenic climate change (IPCC, 2013): the middle Miocene and Pliocene. A corresponding CO2 increase across these two warm intervals, however, remains enigmatic (Fig. 3). For example, stomatal indices suggest CO2 above pre-industrial levels during much of the middle Miocene (Fig. 3D), while paleosol carbonate data indicate very low CO2 and no apparent trends (Fig. 3C). The δ11B-based reconstructions do not show any clear trends during the middle Miocene, with estimates ranging from ∼200 to 600 ppmv (Fig. 3B). High-resolution CO2 data are generally lacking for the late Miocene, which makes inference of CO2 trends during global cooling difficult to establish. In contrast, our reconstruction allows for a nearly continuous record of CO2 that links the mid-Miocene and Pliocene warm intervals by a long-term CO2 decline (Fig. 1C). Finally, our record reveals a CO2 increase within the early Pliocene that is not evident when examining any single proxy, but that corresponds with mid-Pliocene warming and an inferred CO2 increase (e.g., IPCC, 2013, their figure 5.2).

One of the most pressing messages that climate scientists attempt to convey to the public is that current CO2 (2019 CE = 412 ppmv; Keeling et al., 2001) is elevated compared to the geologic past. The fact that current CO2 is higher than it was at any time during the past ∼800 k.y. is a straightforward claim based upon direct CO2 measurements from ice cores (Petit et al., 1999; Kawamura et al., 2007) and the Mauna Loa Observatory (Keeling et al., 2001); claims associated with the more distant geologic past have been variable, partially based on a lack of consensus within the paleoclimate community. Statements addressing values from 3 m.y. ago (Willeit et al., 2019) to 15 m.y. ago (Tripati et al., 2009) can be found, contributing to public confusion and skepticism.

Our results support the claim that CO2 has been lower than present-day values at least across the past 7 m.y., and potentially during the entirety of the past 23 m.y.; however, CO2 likely never fell below levels experienced during the greatest ice-sheet advances of the Pleistocene (∼170 ppm; Petit et al., 1999). Our results also indirectly imply that the major reorganizations of plant (e.g., Salzmann et al., 2008), animal (e.g., Stebbins, 1981), and hominid (e.g., White et al., 2009) ecosystems were not driven by large-amplitude changes in CO2. More meaningful, perhaps, is the inference that these reorganizations could have impelled, or been impelled by, relatively small-amplitude changes in CO2.

Our CO2 record differs from that gained by prior proxies in that it was produced from substrates that span 23 m.y. of uninterrupted Earth history. Our results also show good agreement with discontinuous marine and terrestrial CO2 proxies, suggesting that the validity of the proposed mechanism underlying the effect of CO2 on δ13Cp values (Schubert and Jahren, 2018) may be comparable to those of these previously confirmed CO2 proxies. Compared to these methods, however, our proxy has the advantage of relying upon a substrate (terrestrial fossil organic carbon) that is widely available both spatially and temporally (Strauss and Peters-Kottig, 2003; Nordt et al., 2016), allowing the possibility for a near-continuous reconstruction of CO2 across the entire evolution of C3 land plants.

We thank Peace Eze and Bryce Landreneau for assistance with data compilation. This manuscript benefited from the comments of three anonymous reviewers. This work was supported by U.S. National Science Foundation (grant EAR-1603051); the Research Council of Norway through its Centers of Excellence funding scheme (Project 223272); and the National Science Foundation of China (Grant #41888101).

1Supplemental Material. Description of the inputs used to calculate atmospheric CO2 concentration, uncertainty in CO2(t), Figure S1, and Tables S1 and S2. Please visit to access the supplemental material, and contact with any questions.
Gold Open Access: This paper is published under the terms of the CC-BY license.