The most consequential things AI does in your organisation were not planned. Nobody decided that the recommendation engine would lock your customers into a narrowing content corridor. Nobody scheduled the moment when three AI-optimised suppliers simultaneously de-prioritised your orders because they all read the same market signal. Nobody chose the emergent norm where your AI-assisted teams stop questioning outputs because the model always sounds confident.
These outcomes emerged. And the instinct to respond by clamping down — adding oversight layers, tightening parameters, appointing an AI governance committee — misunderstands what emergence is and how it works.
What emergence actually is
The concept has a precise meaning that gets lost in casual use. John Holland, the computer scientist whose work at the Santa Fe Institute shaped complexity theory, defined emergence as global behaviour arising from local interactions that cannot be predicted or explained by examining individual agents in isolation. His 1998 book Emergence: From Chaos to Order showed this wasn’t metaphor — it was a structural property of any system with many interacting, adaptive components.
Holland’s colleague Stuart Kauffman added the thermodynamic edge: complex adaptive systems don’t just produce emergence, they are drawn toward it. Kauffman’s research on self-organisation demonstrated that systems naturally evolve toward what he called the “edge of chaos” — the narrow zone between rigid order and pure randomness where the richest, most adaptive behaviour occurs. You do not engineer a system to this zone. It finds its own way there if conditions allow.
The Santa Fe Institute’s 2025 symposium on “The Emergence of Complexity and the Complexity of Emergence” — held in April of that year and spanning the natural and social sciences — treated formalising emergence as one of the defining scientific challenges of our era, and explicitly named AI systems as one of the domains where the stakes are highest. Complexity science, applied to AI, tells us that the most significant system behaviours will be the ones nobody specified.
AI turns emergence up to eleven
Classic complex adaptive systems — ant colonies, financial markets, immune responses — produce emergent behaviour on timescales of days, seasons, generations. The feedback loops are slow enough that human observation and intervention can keep pace.
AI breaks this. Three mechanisms make AI-driven emergence qualitatively different from anything organisations have faced before.
Feedback velocity. A recommendation algorithm updates its model of user preferences in milliseconds. A market-making AI re-prices positions in microseconds. The local interactions that generate emergence now happen faster than any human governance process can observe, let alone correct. By the time an emergent pattern is visible, it has already compounded.
Interaction surface. Holland’s original CAS models involved bounded populations of agents. AI systems operate at the scale of the internet — millions of simultaneous local interactions generating emergent global patterns. The Netflix thumbnail personalisation phenomenon is a clean example: no engineer decided to serve racially correlated content. The algorithm optimised locally for click-through; the emergent outcome was demographic clustering that users experienced as bias. Nobody planned it. The system found it.
Non-linear amplification. Kauffman’s NK fitness landscape models show how small changes in local conditions can produce discontinuous jumps in system behaviour. AI concentrates this. When multiple organisations deploy similar AI systems tuned on similar data, their local optimisations synchronise. Supply chain monitoring data from the post-pandemic period consistently shows elevated disruption rates among major manufacturers — driven in significant part by AI procurement systems simultaneously de-risking in response to the same market signals, creating cascade effects that none of them had modelled individually.
The management error
The standard organisational response to unwanted emergent outcomes is suppression: add a rule, tighten a constraint, require a human sign-off. This is not wrong because it is bureaucratic. It is wrong because it addresses the symptom while leaving the generative conditions intact.
Suppressing an emergent outcome in a complex system is like clearing a patch of rainforest. The underlying ecosystem processes that produced that vegetation remain. Something else grows — often faster, because the cleared space removes competition.
The management error runs deeper than tactics. It is a category mistake about what kind of thing AI systems are. Command-and-control management assumes a mechanistic model: the system does what it is instructed, and deviations are faults to be corrected. Complex adaptive systems are not mechanisms. They are ecologies. The behaviour that emerges from them is not a fault — it is the system expressing the conditions it was placed in.
When Amazon’s AI hiring tool began systematically downgrading applications from women, the emergent bias was not a bug in the code. It was the system accurately reflecting the conditions it was trained on: a decade of hiring decisions made in a male-dominated industry. Suppressing the output without changing the conditions just means the same dynamic expresses itself elsewhere.
Teams working with AI assistants show this at a smaller scale. A consistent pattern is emerging in organisations that deploy AI coding or writing tools: team members stop pushing back on AI suggestions, not because the suggestions are always right, but because pushing back requires more effort than accepting. The norm crystallises at the team level — not because anyone decided it, but because the incentive gradient made deference the path of least resistance. Management cannot simply tell people to be more critical. That norm emerged from conditions; only changing the conditions changes the norm.
What cultivation looks like
Gardeners do not control gardens. They cultivate them — shaping the conditions that allow the outcomes they want to be more probable than the ones they don’t. This is the right frame for AI governance.
Cultivation in complex adaptive systems has three practical disciplines.
Design enabling conditions, not prescribed outputs. Holland identified “lever points” — places where simple interventions have lasting, structural effects. For AI systems, these are typically the training environment, the incentive structure, and the diversity of inputs. A recommendation system cultivated for long-term user satisfaction rather than immediate engagement will generate different emergent behaviour than one optimised for clicks — not because individual recommendations are different, but because the conditions select for different attractors.
Use boundary constraints, not content control. Kauffman’s edge-of-chaos work shows that systems need some constraints to produce coherent emergence rather than noise — but too many constraints push them into rigid, non-adaptive behaviour. Effective AI governance sets hard limits on what outcomes are unacceptable (the boundaries) while leaving the interior free to self-organise. The EU AI Act’s risk-tiering approach is a rough attempt at this: categorical prohibitions for the highest-stakes uses, light-touch conditions for lower-risk applications. Whether it gets the calibration right is debatable; the architectural instinct is sound.
Amplify weak signals early. In complex systems, emergent outcomes announce themselves before they fully arrive. Echo chamber formation in social networks left early traces in engagement pattern shifts months before filter bubbles became visible to users or regulators. Supply chain cascade risk shows up in Tier-2 supplier inventory data long before it hits Tier-1 delivery. Cultivation requires investing in the observational infrastructure to catch these signals when they are still cheap to respond to — not waiting for the emergent outcome to become undeniable. This is why organisations that treat AI monitoring purely as a compliance function miss the point: the signal you need is not whether the system broke a rule, but what pattern is forming before it becomes the next rule-breaker.
The question leaders actually need to ask
The wrong question is: how do we control what our AI systems produce?
The right question is: what conditions are we creating, and what will naturally emerge from them?
This reframe has operational consequences. AI audits should examine training environments and incentive structures, not just output distributions. AI governance should include complexity scientists alongside ethicists and legal teams. AI strategy should track emergent phenomena — the patterns no one specified — as leading indicators of where the system is heading.
Emergence is not an AI problem to be solved. It is a property of the systems organisations are now embedded in. Leaders who treat it as a feature to manage will keep fighting outcomes they don’t understand. Leaders who treat it as a condition to cultivate will find themselves ahead of the curve — not because they predicted what would emerge, but because they built the conditions for something worth having.