Nature is complex and chaotic

Before complexity theory covered the bases, now we have to add chaos theory

Nature has always been complex (Bak and Paczuski 1993). Now it's chaotic (Bernardini et al. 2025, Rockström 2009, Sakschewski 2025).

"More than three-quarters of the Earth's support systems are not in the safe zone. Humanity is pushing beyond the limits of a safe operating space, increasingly the risk of destabilising the planet." — Johan Rockström, 2025

We are now in ‘black swan’ territory (Sornette 2009), with chaotic weather events increasingly likely (IPCC 2021).

AI does extend our cognitive horizon. That is a real tool in dealing with this problem, but it will not save us. It will not solve the problem for us.

Machine learning (ML) is pattern recognition AI. Which means it is particularly error-prone in this context, for the same reasons described in the Nobel-prize-winning work that is the basis of the book Thinking Fast and Slow about why human intuition breaks down on some structured problems (Tversky & Kahneman 1974). ML recognizes patterns, which means it doesn't work when the pattern breaks.

Long-term prediction made with AI may fail catastrophically regardless of the data available (Fan et al. 2020). In this world, in our lifetimes, in your lifetime if you are reading this, predictions based on the past, no matter how expensive or detailed, are increasingly unreliable.

Figure J. Chaos and complexity theory producing errors in natural systems data and statistical analysis

When data is linear, points fall neatly on a line. Like this.

Nonlinear data our stats start to fail. The line goes through the middle. But misses both ends. The data isn't linear.

Sometimes linear models work near zero. The tail is where reality lives. And reality is bigger than the line.

Nature doesn't even pretend to follow these rules. This is real data, from a real system. There IS a pattern. It's just smarter than us.

Zoom in and the chaos has order inside it. Zoom in on that order and there's more chaos. All the way down.

Some basics of complex systems

Complex systems aren't just hard to measure — they're impossible to fully quantify (Anderson 1972). That's a property of complexity itself, not a tooling problem.

• Emergence — system-level properties don't exist at the component level; the whole has behaviors the parts don't (Anderson 1972).
• Non-linearity — outputs aren't proportional to inputs; small perturbations can cascade, large ones can dissipate (Strogatz 2015).
• Feedback loops — components shape each other; cause and effect run in circles, not lines (Meadows 2008).
• Sensitivity to initial conditions — tiny differences in starting state diverge exponentially over time; this is the formal definition of chaos (Lorenz 1963).
• Tipping points — systems sit in stable regimes until a threshold is crossed, then flip abruptly to a new one (Scheffer et al. 2009).

The vast majority of Indigenous scientists we have talked to have scoffed at the reductionist mindset toward Nature. This isn't a lack of understanding of reductionism; it's a clear dismissal of the approach as the wrong tool for the job.

"With due respect, but without baggage, we express that we understand, but we lament the logic that those who hold technological power and economic power wield control over the world; because through this path, power has become an obsession of powerful nations that have blinded their essence as children of the earth and have become masters of the planet. In this utilitarian logic, they have objectified the world and commodified everything that exists, and this is now the backbone of the system that states have adopted today." - Miguel Chindoy, The voice of the Indigenous Peoples of the world on the planetary environmental emergency

Some basics of chaotic systems

The simple distinction between complex and chaotic is that complex systems can be stable. Chaotic ones can't. Complexity is structure; chaos is what happens when that structure tips into instability.

• Deterministic but unpredictable — the rules are fully known, the future still isn't (Lorenz 1963).
• Sensitive dependence on initial conditions — the butterfly effect; tiny causes, vast consequences (Lorenz 1963).
• Bounded but never-repeating — chaotic trajectories don't fly to infinity, but they never exactly retrace their path. They trace a "strange attractor" — a shape in possibility space that the system orbits forever without repeating (Ruelle & Takens 1971).
• Prediction horizons are finite by mathematical necessity — past a certain point, no amount of measurement precision or computing power helps. The error compounds faster than the data informs (Palmer 2000).
• Statistics survive; trajectories don't — you can characterize the shape of the chaos (averages, ranges, frequencies of regimes) even when you can't predict any specific path (Eckmann & Ruelle 1985).

That's why ML breaks on chaotic systems specifically: ML learns from trajectories, but in chaos, the trajectory is the part that can't be generalized. Only the statistical envelope can.

How to work in complex and chaotic systems

The point of working in these systems is to work smarter, not harder. In other words, we need to listen to the world's best scientists (above) and work within the constraints.

• Don't try to measure everything (even with AI, especially with AI!). Choose dimensions that are orthogonal and tangible; sample, don't enumerate. Completeness is impossible. Signal that can't be gamed is the goal.
• Don't try to predict the future far in advance. Prediction horizons are mathematical, not technical — no amount of data closes the gap past a certain point. Use short feedback loops instead of long forecasts.
• Optimize for butterfly effects. In a chaotic system, the leverage is at the tipping points — small inputs in the right place produce disproportionate outcomes. The #SexyTrees payment structure ($1 plant / $0.50 alive at 6mo / $0.50 alive at 12mo) is a butterfly: a tiny per-tree payment that flips farmer behavior at scale.
• Don't assume you know what's going to happen. Look at what actually did. Ex-post measurement only. Models are maps; when they conflict with the territory, trust the territory.