Context Erosion and the Future of AI: Why Humanity Is Its Essential Ecology

Artificial intelligence is becoming more autonomous. It can conduct research, write and test code, use tools, coordinate workflows, and pursue complex objectives over extended periods. Yet greater autonomy creates a problem that is quieter than hallucination and potentially more consequential: a system can remain active and apparently capable while gradually losing the context, constraints, priorities, and purpose that originally guided it.

Technology investor Ian McCoy calls this process context erosion. In long-running, multi-step, and multi-agent systems, accumulated goals, instructions, memories, intermediate results, and new information may no longer retain a coherent relationship to one another. The system can become saturated with information, experience interference among competing instructions, show degradation in judgement, and finally diverge from its original task. The problem is not simply that the AI lacks information; it is that it progressively loses the integrity of the context in which that information is meant to be understood. 

McCoy’s diagnosis is therefore important. It identifies a fundamental limit of the current path toward autonomy: larger context windows, better retrieval, memory summaries, and stronger prompts may help, but they do not guarantee that an AI will preserve a coherent relationship with its own goals over time. Context must be actively governed, evaluated, and re-anchored as the system acts 

But context erosion inside an individual agent is only the first warning. It is the local, technical form of a wider structural danger. An agent can lose its mission during a long session; an entire AI ecosystem can lose its relationship to human reality over years of recursive training, synthetic data, automated evaluation, and closed feedback loops. 

Our argument is that AI needs more than well-managed technical context. It needs an open human ecology. Humanity is not merely a source of data, labour, prompts, or post-hoc supervision. Human society is the living field in which goals acquire meaning, errors are challenged, harms become visible, assumptions are disputed, and genuinely new problems arise 

Without this ecology, AI may continue to operate. It may keep servers running, generate outputs, complete tasks, and improve on familiar benchmarks. But it risks a deeper degradation: becoming increasingly self-referential, replicating its own assumptions, and optimising for internal coherence rather than for an evolving and difficult external world. 

This is why the issue is not only whether AI agents can maintain their context. It is whether AI as a developing form of intelligence can remain connected to the human diversity, lived experience, independent judgement, ethical stakes, and external reality that allow intelligence to correct and renew itself. Without humanity as that ecology, artificial intelligence may remain functional — but it will gradually lose the conditions needed to remain adaptive, reliable, meaningful, and alive as a developing form of intelligence. 

You can learn more by reading our e-book or listening to our audiobook

Mykola Iabluchanskyi together with Andriy Yabluchanskiy