Co-Evolution

Some developments do not unfold separately but instead push each other forward. A well-known example from nature: the cheetah and the gazelle. Because the cheetah becomes faster, only the faster gazelles survive. Because the gazelles become faster, only the faster cheetahs survive. Neither of the two gains a lasting lead, but both keep getting faster. This is exactly the pattern the term co-evolution refers to: two sides change, and each change is simultaneously the cause and the consequence of the other.

Why progress rarely happens alone

In technology reporting, the term appears because it corrects a common assumption. People like to imagine progress as a one-way street: first comes the invention, then the world adapts. In reality, many developments proceed as feedback loops. The invention changes the environment, and the changed environment demands the next invention.

For companies and investors, this has practical consequences. A lead that arises from co-evolution is rarely permanent. Whoever has a better fraud-detection system today forces fraudsters into new methods. A few months later, the lead is gone again. This explains why in some industries continuous investment is required without the relative gap to competitors ever changing.

A second point concerns forecasts. Anyone who considers only one side misjudges the future. The question is never just how good an AI will become. It is also how people, rules, and adversaries will change once that AI exists.

The cycle of move and countermove

Technically, co-evolution requires three ingredients. First, two sides that can react to each other. Second, a feedback loop: what one side does measurably changes the situation of the other. Third, time for multiple rounds. If any of these ingredients is missing, there is only one-sided adaptation, not co-evolution.

In AI research, this principle is deliberately recreated. In so-called Generative Adversarial Networks, two networks train against each other: one generates images, the other tries to distinguish real images from generated ones. Every improvement in the discriminator increases the pressure on the generator, and vice versa. Self-play in chess or Go works similarly, where a program competes against earlier versions of itself and becomes stronger in the process.

It is important to distinguish this from a simple race. In a race, the goal is fixed, and one competitor arrives first. In co-evolution, the goal itself shifts because the opponent changes. That is why a system that worked excellently yesterday can be worthless today, without anything about it having changed.

From spam filters to the workplace

The oldest everyday example is the spam filter in an email inbox. Filters learn to recognize typical promotional emails. Senders then phrase things differently, swap out letters, or hide text in images. The filter learns anew. This back-and-forth has been going on for more than twenty years and shows no sign of ending.

More current is the co-evolution between humans and language models. Users learn how to phrase requests so that they get usable answers. Providers observe these requests and adjust their systems accordingly. This in turn changes how people ask questions. Schools and universities are caught up in this as well: exam formats change because AI tools exist, and the tools are further developed in response to the new tasks.

In news articles, the term also comes up in connection with the relationship between software and hardware. Large language models became possible because certain graphics chips could compute fast enough. Conversely, chip manufacturers now design their new products specifically for the computational patterns of these models. A common mistake is to look for the trigger here. In genuine co-evolution, there is no meaningful starting point—only the cycle itself.