Every organisation deploying AI is making a version of the same mistake. They assess what AI can do today, decide whether to adopt it, and build a strategy around that assessment. The mistake is not being wrong about current capabilities. The mistake is treating capability as a fixed property of the technology.
It is not. AI systems change in response to how they are used. The organisations using them change in response to what AI makes possible. Competitors adapt to what you are doing. The whole landscape shifts — and your strategy, built on last quarter’s snapshot, is now a map of a territory that no longer exists.
This is co-evolution. Understanding it is not optional for serious AI strategy.
What Co-Evolution Actually Means
In 1973, evolutionary biologist Leigh Van Valen proposed what became known as the Red Queen hypothesis. The name comes from Lewis Carroll: in Wonderland, you have to run as fast as you can just to stay in the same place. Van Valen observed that species do not evolve in isolation — they evolve in response to each other. Predator adaptations pressure prey to evolve defences; prey defences pressure predators to become more sophisticated. The fitness landscape never stabilises. Standing still is falling behind.
Stuart Kauffman, in his work on complex adaptive systems, formalised a related concept: the adjacent possible. At any moment, the state-space of what can happen next is shaped by what currently exists. As new combinations emerge, they open doors to combinations that were previously unreachable. The adjacent possible expands with each step taken. You cannot predict which doors will open from outside the current moment.
Both concepts describe the same dynamic: change begets change, in directions that cannot be fully anticipated in advance.
This is the situation every organisation is now in with AI.
The Feedback Loop in Practice
When an organisation deploys a large language model into its workflows, the model does not stay static. It learns from interactions — through fine-tuning, through retrieval-augmented patterns, through the implicit feedback of which outputs are accepted and which are overridden. The way the organisation uses the AI shapes what patterns the AI reinforces. The AI changes what the organisation knows how to do. What the organisation knows how to do changes what it asks the AI to do next.
This is not metaphor. It is the operational mechanics of modern AI deployment.
Consider AI coding assistants. Workers who use them stop practising certain classes of problem-solving. A 2024 analysis by GitClear, examining code changes across organisations with high GitHub Copilot adoption, found elevated rates of code churn alongside declining complexity in newly produced code. The proposed mechanism: when AI automates the mechanical parts of coding, developers’ tolerance for complexity narrows, and the AI learns from the simpler code they accept. The loop feeds itself. The AI shapes the developer; the developer shapes what the AI is asked to produce.
The same dynamic operates, at greater scale, in recommendation systems. When YouTube deployed its recommendation engine, it created a feedback loop with content creators: the algorithm rewarded certain content patterns, creators optimised for those patterns, the algorithm updated on the new content distribution, which shifted the reward signal, which changed creator behaviour further. This is not a design flaw. It is co-evolution operating as the theory predicts. The ecosystem of creators and the algorithm are now jointly producing each other’s behaviour. Neither can be understood independently of the other.
A third example is less discussed but strategically significant: AI procurement co-evolution. As organisations adopt AI tools, their suppliers adopt AI to serve them more efficiently, changing the terms on which procurement decisions get made, the criteria for evaluation, the timeline expectations, the cost structures. Early-adopting buyers shape the tools their suppliers build. Late-adopting buyers face a supplier landscape shaped by their competitors’ choices. The adjacent possible in supplier relationships has been rearranged around them without their participation.
The Strategy Failure
The standard approach to AI strategy is a capability assessment: what can this technology do, and what is the ROI of adopting it? This is a sensible question to ask about a static tool. A spreadsheet does not change its capabilities in response to how you use it. A hammer does not evolve.
AI does.
The capability assessment captures a point-in-time snapshot of a co-evolving landscape. By the time the assessment is complete, the deployment is live, and first-year results are in, the landscape has shifted. The organisation has changed. The AI system has changed. Competitors have moved. Suppliers have adapted. A strategy built on the initial assessment is now partly a strategy for a system that no longer exists.
Researchers Juan Perdomo and colleagues formalised this in their 2020 paper on performative prediction: ML model predictions actively change the distribution they are trying to predict. A credit scoring model shapes who applies for credit. A content recommendation model shapes what content gets produced. The model is co-producing the world it models.
This is the structural reason why treating AI as a controllable instrument — something you point at a problem and measure the result — systematically underestimates what you are actually dealing with.
What Changes About Governance
Static AI governance is the organisational equivalent of fighting the last war. A policy written for today’s AI capability distribution will be wrong for tomorrow’s. Rules designed for current human-AI skill distributions will not survive co-evolutionary skill drift.
The implication is not that governance is useless. It is that governance must be adaptive.
Adaptive governance treats AI policy as a living document updated on a rhythm that reflects the pace of co-evolutionary change. It assigns explicit ownership of monitoring co-evolutionary drift — the divergence between the world the AI was deployed into and the world it is now operating in. It creates feedback circuits: mechanisms for detecting when the co-evolutionary loop is becoming harmful and intervening before lock-in occurs.
Co-evolutionary lock-in is the scenario to avoid. It occurs when mutual dependencies between the organisation and its AI systems become so deep that either cannot function without the other in its current form — even if the current form is producing bad outcomes. The recommendation platform that cannot de-optimise for engagement without triggering a content creator exodus is in co-evolutionary lock-in. The development team that cannot review code without AI scaffolding, on a codebase written for AI to manage, is approaching it.
Practical Disciplines
Three disciplines follow from taking co-evolution seriously.
Monitor for drift. Establish baseline measurements at deployment and track them over time: task completion profiles, skill distribution across the team, decision patterns that humans override or accept from the AI. Drift is not a failure signal. It becomes a problem when it is unmonitored and the strategy is not updated to reflect it.
Design feedback circuits with intent. Every AI deployment creates feedback loops. The question is not whether to have them but which ones to cultivate. Loops that reinforce human skill, surface novel patterns for human judgment, and feed exceptions back into model improvement are worth building deliberately. Loops that automate human oversight out of existence are worth suppressing early.
Treat AI strategy as a negotiation, not a decision. A point-in-time decision — “we are deploying AI for X use case” — is the beginning of a co-evolutionary relationship, not the resolution of a strategic question. That relationship requires ongoing attention: periodic renegotiation of what the AI does, how the organisation adapts, and whether the co-evolutionary trajectory is heading toward a place the organisation actually wants to be.
The Red Queen runs in one direction. You cannot stop running. What you can do is choose your direction with more care than you have been.