We are building many artificial intelligences. But are we, in effect, building variations of the same one?
Today’s leading systems, ChatGPT, Gemini, Claude, Llama, & Grok, are often framed as competitors: different companies, different architectures, different philosophies. The surface story is one of diversity.
Look closer, and a quieter pattern emerges: convergence.
Not total convergence, and not in every respect. But enough to raise a serious question about the kind of intellectual ecosystem we are creating.
A shared foundation
Most large language models (LLMs) are trained on broadly overlapping data: the public internet, large-scale crawls, reference works, forums, code repositories, and digitized books. The differences between AI labs lie less in what they learn from than in how that data is filtered, weighted, and refined.
This matters because training data is not just information, it is a distribution of perspectives, assumptions, and reasoning habits. When models learn from similar distributions, they inherit similar defaults: what counts as credible, what requires caution, what “balance” looks like.
This does not produce identical systems. But it does create a gravitational pull toward similar outputs, especially on familiar questions where the underlying material is dense and well-trodden.
Where convergence shows up
Use multiple systems side by side and a pattern becomes visible.
They differ in tone, verbosity, and style. But they often:
- converge on similar high-level conclusions
- hedge in similar places
- privilege similar kinds of evidence
- reflect similar assumptions about what a reasonable answer looks like
This is best understood as behavioral convergence, not a claim that models literally “think” the same way. Internally, their representations may differ. But from the user’s perspective, the outputs increasingly occupy the same intellectual neighborhood.
That distinction matters. But so does the outcome.
Why similarity at scale matters
Intellectual diversity is not just a philosophical preference. It is a functional requirement for error correction.
When multiple systems approach a problem differently, they expose each other’s blind spots. When they approach it the same way, agreement can look like validation, even when it is simply shared bias.
If AI systems converge on similar frameworks for interpreting evidence and uncertainty, then deploying more of them does not increase diversity of thought. It increases the reach of a particular pattern of thought.
Scale, in that context, is not a safeguard. It is amplification.
The feedback loop risk
A second dynamic makes this more consequential.
AI-generated content is rapidly becoming part of the internet. That content will, inevitably, become part of future training data. If current systems already exhibit partial convergence, their outputs may reinforce those patterns in the next generation.
The risk is not that models become identical. It is subtler: a gradual narrowing of the distribution of perspectives, as synthetic text – fluent, plausible, and statistically “typical” – accumulates.
Human discourse has always had biases and concentrations. But it also has a long tail: minority viewpoints, unconventional reasoning styles, culturally specific perspectives. Those tails are easy to under-sample and easier still to overwrite.
What diversity would actually require
If this trajectory is undesirable, it is not enough to rely on competition between companies. Structural similarity can persist across competing systems.
Meaningful diversity would likely require deliberate intervention:
- broader and more globally representative training data
- transparency around what is included and excluded
- variation in training objectives, not just architectures
- mechanisms that preserve minority viewpoints rather than averaging them away
In other words, diversity has to be engineered, not assumed.
A choice, not an inevitability
We are still early enough that these patterns are not fixed. The defaults are still being set by researchers, by companies, and by the data we choose to prioritise.
The question is not whether AI systems will agree sometimes. Of course they will. The question is whether we are comfortable building an ecosystem where agreement becomes the norm not because it is always correct, but because the systems producing it were shaped in similar ways.
That is not a technical constraint. It is a design choice. And it is one that is easier to examine now than to unwind later.
First dropped: | Last modified: April 29, 2026