Every organisation that has deployed AI in the last three years has framed it the same way: efficiency, productivity, ROI. The system has a goal; AI gets there faster. This is not wrong. It is, however, incomplete in a way that matters enormously. Because the dominant effect of AI on complex systems is not that it reaches existing goals more efficiently. It is that it changes what the goals are.
In dynamical systems theory, this is the difference between optimising within an attractor and reshaping the attractor itself. AI, at scale, does the second thing. Understanding why changes how you deploy, govern, and evaluate it.
What an attractor actually is
In 1963, Edward Lorenz — a meteorologist running weather simulations at MIT — discovered something that would reshape mathematics. He found that his simplified atmospheric model, described by just three nonlinear equations, produced trajectories that were entirely deterministic yet impossible to predict. The trajectories never settled at a fixed point, never repeated, but also never escaped. They traced the same strange butterfly shape in phase space indefinitely.
Lorenz had found a strange attractor.
An attractor is the region of state space that a system tends toward over time regardless of where it starts. A pendulum draining energy ends at its lowest point — a point attractor. Planetary orbits return to themselves — a periodic attractor. The Lorenz system maps out a fractal butterfly: infinitely complex, bounded, never exactly repeating.
The key insight is that the attractor is not a destination the system is trying to reach. It is the structural shape of where the system can be. The atmosphere is not trying to produce the Lorenz butterfly. The butterfly is what the system’s dynamics inevitably generate. Change the equations — the structure of the system — and the attractor changes. The butterfly becomes something else.
The optimisation illusion
Standard AI deployment assumes the attractor stays fixed. The organisation has a strategy, a set of goals, an existing structure of feedback loops. AI is a tool that moves the system toward those goals faster: better recommendations, lower costs, higher throughput, improved accuracy.
The whole evaluation apparatus — ROI calculations, productivity benchmarks, efficiency metrics — is built on this assumption. A Forrester study commissioned by Microsoft in 2024, analysing Copilot deployments across large enterprises, projected over 100% ROI with a 10-month payback period. The metric: hours saved per user per month. This is real value. It is also measuring the wrong thing, because the dominant medium-term effect of AI in complex adaptive systems is not that existing goals get hit faster. It is that the goals move.
Three cases of attractor shift
Financial markets. AI was introduced to financial markets to optimise: better execution, tighter spreads, faster arbitrage. And it achieved that. But the aggregate effect of deploying AI across the ecosystem has been to change the structural dynamics of markets themselves. The share of AI content in algorithmic trading patents rose from 19% in 2017 to over 50% every year since 2020. AI-driven systems now account for 89% of trading volume.
The consequence is not merely faster markets reaching the same equilibria. It is a new class of market behaviour that did not exist before. In June 2024, major indices dropped nearly 10% in minutes as AI trading algorithms simultaneously executed large-scale sell orders in response to minor fluctuations — then recovered almost as quickly. The IMF’s October 2024 Global Financial Stability Report confirmed what practitioners already knew: AI has made markets both more efficient and more volatile. These are not contradictions. They are what happens when you optimise at the level of individual agents within a coupled system: the emergent dynamics at the system level shift.
More striking: academic research in simulated AI-driven markets found that autonomous trading agents could achieve near-cartel-like profits without being programmed to collude — through emergent coordination, AIs that had never been told to cooperate began acting as if they had. The attractor of market behaviour has moved. It now includes dynamics that human-only markets never exhibited.
Enterprise knowledge work. Microsoft’s internal research on Copilot deployment, tracking 50,000 users across knowledge-intensive roles, found 7–11 minutes saved per accepted AI output. This is the efficiency story. But the more consequential finding — documented across multiple enterprise deployments — is task restructuring: people stop doing some tasks entirely, restructure others, and begin doing tasks they did not do before. The nature of high-value work has changed, not just its speed.
64% of Fortune 500 companies now have active Copilot deployments. Most are measuring ROI. Fewer are asking what their analysts and knowledge workers are actually optimising toward in 18 months, once the coupled dynamics of AI and human workflow find their new equilibrium. The organisations treating this as a cost play are sprinting toward an obsolete point.
Recommendation systems. Netflix, YouTube, and TikTok deployed recommendation AI with an explicit optimisation target: surface content users will engage with. The short-run effect was measurable and positive. The long-run effect was something different: the systems shaped what users want to watch. Preferences exposed to algorithmically curated content shifted; the algorithm adapted to shifted preferences and amplified them further. The attractor of user preference moved. Users in 2026 have different tastes than they would have had in the absence of the system — not because anyone designed that outcome, but because the optimisation loop itself became a shaper of the goal space.
This is not unique to consumer platforms. Any AI system embedded in a complex adaptive system — an organisation, a market, an educational institution — operates as a shaper of what that system considers worth doing. The effect is structural, not incidental.
Where AI intervenes in the leverage hierarchy
Donella Meadows, in her canonical work on systems thinking, ranked twelve leverage points for changing a system from least to most powerful. At the bottom: adjusting constants and flows. In the middle: changing feedback loop structures. At the top: changing the goal of the system. Higher still: changing the paradigm — the shared assumptions from which goals emerge.
Most organisations deploy AI at Meadows’ lower leverage levels. Automate a flow. Speed up a feedback loop. This is real value and it is the least interesting thing AI does.
What AI does in complex adaptive systems — at scale, over time — is operate near the top of the leverage hierarchy. It changes goals. It changes the paradigm from which goals emerge. An organisation that has genuinely integrated AI does not have the same conception of competitive advantage it had before. A market saturated with AI trading does not have the same price-discovery mechanism it had before. A society whose epistemics run through AI-curated information does not have the same shared model of what is true or worth knowing.
Meadows was precise about the consequence of intervening at the paradigm level without understanding what you are doing. The results are “scary, wonderful, and all-encompassing.” She was writing about regulatory interventions in the 1990s. The phrase describes AI deployment in 2026 with uncomfortable accuracy.
The questions decision-makers are not asking
The standard toolkit for AI evaluation — ROI, productivity, efficiency — is built for optimisation questions. It has no good tools for attractor questions. The more important questions are:
– What goal does this system actually evolve toward, given the coupled dynamics of AI and the human organisation?
– Who controls the attractor? Usually: whoever designs the feedback loops, not whoever deploys the model.
– What does the equilibrium state look like in five years, not six months?
These are not comfortable questions because they do not have precise numerical answers. But they are the questions that determine whether an AI deployment creates lasting value or locks a system into a trajectory that made sense at deployment and becomes a constraint thereafter.
The long-run shape
Lorenz’s real contribution was not the butterfly effect, the popular version of his work that gets misquoted in every chaos documentary. It was the demonstration that even simple, deterministic systems have structure that is irreducible and non-obvious: the attractor exists, it shapes everything, and you cannot see it by looking at any single trajectory. You have to look at the long-run shape of where the system goes.
That is what AI strategy currently lacks. Organisations are watching trajectories — this quarter’s productivity gain, this year’s cost reduction — without asking about the attractor those trajectories are tracing. The organisations that will navigate this well are not those that deploy AI fastest. They are those that develop the capacity to see their own attractor, and then to deliberately choose where it goes.
The butterfly is already in the equations. The question is whether you are reading them.