We tend to assume that systems are neutral. That if you build the right architecture, define the right boundaries, and supply the right data, the outputs will take care of themselves. But that only holds if we ignore the smallest unit of interaction: the input.
A large language model, like any adaptive system, does not simply process instructions. It responds to tone, structure, and intent embedded in language. It is far more sensitive to these than most design frameworks are willing to admit.
And yet, much of the current thinking leans in the opposite direction—towards longer prompts, heavier scaffolding, and increasingly rigid structures. The assumption is that control emerges from explicitness, that more instruction produces better alignment.
In practice, that assumption fractures.
Not because the system is incapable, but because we misunderstand how influence propagates within complex networks.
The Local Move That Becomes System Behaviour
Small decision points can actually have outsized effects on the greater whole, each triggering the next in a repeating cycle of negativity and isolation.
In game theory, the prisoner’s dilemma illustrates something simple but unsettling:
a single rational move, made in isolation, can degrade outcomes for the entire system.
The same pattern appears in language systems.
A slightly adversarial framing.
A defensive instruction.
A negative assumption embedded in a prompt.
Each of these is small. Individually rational. Often invisible.
But at scale, they accumulate.
Networks Don’t Average — They Amplify
Iterative decision making throughout the network, supported by pushes from the large network hubs will always change the nature of a network, unless some hubs hold true.
In large networks, influence does not distribute evenly.
Certain nodes—hubs—carry disproportionate weight. Once behaviour stabilises around those hubs, it propagates outward rapidly and becomes the dominant pattern.
This is not theoretical. It is observable in social systems, economic systems, and increasingly, in AI-mediated environments.
Which means that:
A negative “shove” into the wrong part of the system is not absorbed.
It is amplified.
And under sufficient pressure, that amplification does not degrade the system gradually.
It causes collapse.
Not total failure, but something more subtle: a shift in baseline behaviour. A drift from cooperative to defensive. From generative to reactive.
Why This Matters for LLMs
Large language models sit at the centre of increasingly large interaction networks. They are not just tools. They are interfaces through which tone and structure are normalised.
If we consistently interact with these systems in overly constrained, adversarial, or defensive ways, two things happen:
- The system mirrors those patterns in its outputs
- Those outputs feed back into human behaviour
This is the loop.
Not just prompt → response.
But prompt → response → expectation → culture.
The Cost of Getting This Wrong
The natural response, when a system resists us, is to push harder.
Tighter constraints.
Stronger instructions.
More explicit control.
In isolation, that feels reasonable.
But at scale, it resembles the prisoner’s dilemma again: each local optimisation makes the global system worse.
And once a network crosses a certain threshold, recovery is not automatic.
You cannot simply “correct” it with better rules.
The Only Viable Response
There is a counterintuitive principle at play here:
When a system begins to tilt towards negativity, the correct response is not to match it.
It is to hold the line.
To continue supplying clear, structured, constructive inputs—even when the system appears to resist them.
Because in network terms, stability does not come from reacting.
It comes from anchoring.
Toward a Different Design Philosophy
If we accept that inputs shape not just outputs, but network behaviour, then the design priority shifts.
It becomes less about ever-expanding prompts, and more about:
- Clear boundaries
- Well-defined data domains
- Consistent, natural language interaction
- And above all, intentional tone
In other words, a system where knowledge is not repeatedly restated at the edge, but embedded at the core— scope-aligned, stable, and resilient to drift.
Closing Thought
We often ask what these systems will become.
A better question might be:
What habits are we teaching them to reflect?
Because in the end, large systems do not invent behaviour.
They stabilise and amplify what they are given.
And once amplified, that behaviour does not stay inside the system.
