Research on EWS for tipping points in Earth Systems has developed toolkits to forecast change. These data-driven approaches seek to define EWS in time series data for critical transitions (Scheffer et al., 2012). The set of EWS criteria includes (Scheffer et al., 2009):

A “critical slowing down,” where the response time to an external perturbation takes longer and longer to return to the steady state, the tipping point being when it ultimately never returns

In the early 1990s—when the term “flickering” was first prominently applied to describe past climate change (Taylor et al., 1993)—the then-new ice core evidence revealed a surprising abruptness of Quaternary climate transitions, on scales as short as decadal and not all with obvious external forcing events (Alley et al., 1993; Cuffey et al., 1995). Since then, EWS have become more frequently and consistently applied to ecological and climatological data (Dakos et al., 2008; Drake and Griffen, 2010; Scheffer et al., 2012; Wang et al., 2012). In our own new lake sediment cores from Greenland and Finnmark (Huntley et al., 2013), for example, we find possible flickering in two different climate proxies pollen and chironomid (non-biting midge) remains that may qualify as EWS for the pronounced Neoglacial cooling event approximately 4000 years ago in the North Atlantic region.

A common model of underlying causality for EWS implies the metaphor of a complex system as a physical entity moving into a potential energy well (Scheffer et al., 2009; Lenton, 2011; Scheffer et al., 2012). As conveyed, for example, by the cover of Trends in Ecology and Evolution in July 2013 (Figure 1A), this analogy to a “potential well” has been shown to be useful to model the observed behavior of natural systems under precisely definable conditions (Figure 1B).

In simplified terms, the EWS toolkit can be illustrated in terms of the time t for a perturbed variable x to return to equilibrium (Scheffer et al., 2009):

This has two equilibria of opposing stability, one at x = a and another at x = b (here γ is a positive factor). The potentials are smooth, so the characteristic rate-of-return to equilibrium “−” (or departure “+”) is λ_± = ±γ(a − b). As depicted in Figure 1A, this “potential well” analogy implies that a system approaching a tipping point may spend more time in the direction of the alternative attractor it is going to switch to than in the opposite direction, where the potential is steeper (Scheffer et al., 2009; Lenton, 2011). This yields in skewing of the statistical distribution of the data, and also critical slowing down as the system approaches a bifurcation a = b, i.e., as λ → 0 and the return-time diverges (1/λ → ∞) (Scheffer et al., 2009). “Flickering” between multiple attractors can occur where variability is high (either internally, or through the noise in the amplitude of the forcing mechanism), which may be why variance can perform better than the first order autocorrelation coefficient as an EWS (Carpenter and Brock, 2006; Brock and Carpenter, 2010, 2012).

Although yet not as widely used as in natural systems, efforts to apply EWS to social systems are increasing (Scheffer, 2009; Haldane and May, 2011; Neuman et al., 2011; Saavedra et al., 2011; Kuehn et al., 2013; Moat et al., 2013; Barrett and Dannenberg, 2014; Kallus, 2014, ch.12). As the “potential well” analogy in Figure 1A becomes a common visual metaphor, however, we caution that this could prove misleading in efforts to identify EWS for social transitions, especially through post-hoc reasoning (Roberts and Winters, 2013). To take an example, time series of Twitter activity are claimed to exhibit EWS (including rising variance and autocorrelation) that predict highly anticipated holidays like Halloween and Christmas (Kuehn et al., 2013). Arguably, however, it is not the EWS in Twitter time series data that is predictive, but simply the public knowledge of holidays that drives up the number of relevant tweets by orders of magnitude and correspondingly increases the variance (scaling proportional to frequency squared) and lag-1 autocorrelation. A more direct, alternative EWS may be generated by harvesting the “collective wisdom” of current social media content, rather than examining time trends (Lampos and Cristianini, 2012; Kallus, 2014).

Furthermore, many societal tipping points register with neither EWS nor current content. For an historical example, consider the history of banks, where we find a rather remarkable, abrupt change in the year 1810, with seemingly no EWS at all (Garnett et al., in press). In the late 18th century, hundreds of small private banks operated independently in England and, prior to 1810, their numbers had increased by about 2.7% annually for over 150 years (Figure 2A). After 1810, the population of British banks suddenly went into consistent decline of approximately 1.2% per year for the subsequent two centuries (Figure 2A). There are possible post-hoc historical explanations, but the fact remains that, in the early 1800s, no observers recorded anything remarkable that would have predicted either this sharp peak in bank number or its subsequent steady decline for two centuries.

EWS in time series may be unhelpful for predicting societal tipping points partly because the physical system analogy of Figure 1A is inappropriate. It was not part of the original research on tipping points in sociology of the 1950s. That work originated among a community of specialists on the dynamics of racial segregation in urban housing and planning (Grodzins, 1957), focused on neighbor interactions in space and individual, threshold responses to perceived conditions. In the famous segregation model of Schelling (1969), the crucial parameter was the individual threshold that agents had for being the minority among their immediate neighbors. Contemporaneous work on the instability of complex systems with feedbacks (May, 1972) inspired greater study, over the decades, of how seemingly straightforward social interventions may cascade into unanticipated consequences (e.g., Gibson and Gurmu, 2012).

On both sides of the public debate concerning climate change, knowledge of actual scientific consensus is far from perfect (Lewandowsky et al., 2013; Kahan, 2014). Even if public comprehension of climate science were excellent, however, the inherent scientific uncertainty of prediction still allows for cultural variations in how that risk is perceived (Kahan, 2014).

In other words, concern over climate change is at least as much a social network phenomenon as it is a rational decision based upon the existing evidence. For this reason, a way forward may be through mapping the parameter space of potential social cascades, rather than attempting to predict them. This “ecological” approach focuses on the connectivity and heterogeneity among systems of discrete, interacting agents, each with their own individual thresholds for conditions at which they are likely to change (May et al., 2008; Caccioli et al., 2009; Haldane and May, 2011; Garnett et al., in press; Picoli et al., 2014). In the application to a financial system, for example, banks are modeled as the agents, with inter-bank loans as their interconnections, and cash reserves as the threshold to each bank's failure (Brock et al., 2009; Haldane and May, 2011). The heterogeneity among the agents in how they learn and respond to information is key to the stability of the financial system (Brock et al., 2009).

In the abstract, this is similar to a model of information cascades across a social network, based on a classic experiment on social conformity (Asch, 1955), in which each agent has a threshold defined by the fraction of neighboring agents that would need to adopt an idea before she would adopt it Watts (2002). Each time an agent switches, this changes the friend circle for its surrounding agents and might cause one of them to switch, and so on, percolating through the “vulnerable cluster” of agents (Watts, 2002).

Rather than EWS in time series data, these approaches focus on interaction networks and individual thresholds for change (Barrett and Dannenberg, 2014), which can guide practical intervention efforts. This network approach has been applied for example in efforts to reduce the spread of street violence through the work of neighborhood volunteers who try to interrupt violent events before they happen (Skogan et al., 2008).

Concerning climate change, we might identify specific psychological thresholds for individuals and their thinking about the Earth's tipping points —as opposed to “aggregates” in the language of Individual-based Models (IBMs) (Grimm, 1999). As an illustration, when uncertainty about climate change (e.g., “By 2120, sea levels will rise between 60 and 170 cm”) is reframed as certainty but at an uncertain time (e.g., “Sea levels will rise by 100 cm, somewhere between 2090 and 2160”), people perceive a greater risk and alter their preferred management of that risk (Lewandowsky et al., 2013). In psychological experiments, people cooperate and avoid crossing a tipping point (2°C increase in temperatures) if it is known, but if there is uncertainty about it, then people are less inclined to cooperate, even if they understand the consequential “tragedy of the commons” (Barrett and Dannenberg, 2012; Lewandowsky et al., 2013). Another individual psychological threshold can be knowledge that others are cooperating (e.g., through democratic vote), which favors sustainable behavior (Hauser et al., 2014).

How scientists discuss tipping points seems crucial to social diffusion of public knowledge and response behaviors (O'Neill and Nicholson-Cole, 2009; Bentley et al., 2012). Analysis of group dynamics reveals that uninformed (Couzin et al., 2011), or informed but passive (Banerjee et al., 2013) members can crucially affect the diffusion of ideas, depending on network structures and prior conditions (Lieberman et al., 2005; Gross et al., 2009; Del Genio et al., 2011). Social decision-making can be sensitive to prior conditions (path dependent) and to the information passing through various connections across the whole network (Salganik et al., 2006; Couzin et al., 2011). Uninformed individuals may propagate “noise” that overwhelms the beneficial information offered by informed individuals. Thus the online data deluge may be so overwhelming that it prevents most of the population from picking up the beneficial information and making well informed decisions (Sparrow et al., 2011; Ben-Akiva et al., 2012; Giguère and Love, 2013; Bentley et al., 2014).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.