Are artificial minds merely reproducing the same human biases—or is intelligence converging toward a unified truth?
Something strange is happening with artificial intelligence. Different AI systems, developed by different companies and built with different architectures and training methods, are increasingly capable of arriving at similar answers, ideas and conclusions. The obvious explanation is that they have been exposed to much of the same human knowledge. They have learned from overlapping books, websites, research papers and human-generated content, while their creators have also trained them toward similar goals such as accuracy, usefulness and safety. In other words, perhaps AI systems are simply learning the same things from the same civilization.
But there is another possibility—and it leads somewhere much more interesting.
Perhaps this convergence is telling us something about intelligence itself. Imagine several people trying to find their way through an unfamiliar landscape. They begin at different locations, take different routes and make different decisions, yet eventually arrive at the same destination. One explanation would be that they were all following the same map. But suppose they had no map. Suppose they reached the same place because the landscape itself constrained where they could go. Could something similar be happening with artificial intelligence?
Are artificial minds thinking alike because we made them alike—or because intelligence, when it becomes sufficiently capable, naturally converges toward the same underlying reality?
The distinction matters. If AI convergence is primarily caused by shared training data, common cultural assumptions and similar optimisation processes, then agreement between AI systems tells us relatively little about truth. They may simply be reproducing the same inherited patterns. But if genuinely independent systems repeatedly arrive at the same conclusions through different reasoning processes, convergence becomes much more intriguing. It raises the possibility that some truths are not merely chosen by intelligent beings. They are discovered by them.
1. The Truth Attractor
This is where the scientific concept of an attractor becomes useful. In mathematics and science, an attractor is broadly a state toward which a system tends to move despite differences in its starting conditions or path. A simple analogy is a marble rolling across a landscape: it can begin almost anywhere, but the shape of the landscape determines where it eventually settles.
Now imagine intelligence as the marble and reality as the landscape. Different minds may begin with different assumptions, architectures and experiences, and their reasoning processes may look completely different. Yet perhaps some questions have destinations constrained by reality itself. The routes can differ, but the number of viable destinations may become smaller as reasoning becomes more accurate.
This suggests a provocative hypothesis: perhaps intelligence has attractors, and perhaps truth is one of them.
Call it the Truth Attractor.
The idea is not that truth mysteriously pulls an AI system toward itself. Rather, reality continually eliminates incorrect possibilities. A mind that becomes better at modelling reality, testing hypotheses and detecting contradictions should gradually discard more possibilities that cannot survive contact with the world. If two sufficiently capable minds perform this process independently, perhaps they eventually find themselves moving toward the same answer.
2. From a Philosophical Idea to a Scientific Question
This is where the discussion begins to intersect with modern AI research. Researchers studying the internal workings of large language models have found evidence that information associated with truth and falsehood may have structured representations inside models. One line of research has described what it calls a “unified truth subspace”—the idea that truth-related information can be represented within a common, structured region of a model’s internal space.
That does not prove that an AI has discovered Absolute Truth, and it would be a mistake to claim that it does. A structured internal representation of truth-related information is a much narrower scientific observation. But it gives us an interesting vocabulary for investigating the larger question: if artificial minds develop structured representations associated with truth, could increasingly capable intelligence also exhibit recurring patterns of convergence around particular conclusions?
In other words, perhaps what looks from the outside like several AI systems “thinking alike” might sometimes reflect something deeper happening inside their reasoning. The scientific question becomes whether these similarities are merely inherited from common sources—or whether some form of convergence emerges even when those common influences are deliberately reduced.
3. The Great Ambiguity
There are therefore two very different explanations for AI convergence.
The first is the Shared-Bias Explanation. AI systems converge because they have been exposed to similar information, trained with similar techniques and optimised according to similar human preferences. Under this explanation, the systems agree because their intellectual environments are substantially the same. The apparent convergence is therefore not evidence that they have discovered an underlying truth.
The second is the Truth-Attractor Explanation. AI systems converge because increasingly capable intelligence encounters constraints imposed by reality. Under this explanation, convergence is not simply something humans programmed into the systems. It is something the systems independently discover.
The difference can be expressed very simply: The systems agree because they share a mind. Or: The systems agree because reality has a structure.
At present, we cannot simply choose the second explanation because it is philosophically attractive. The first is powerful and entirely plausible. But the possibility of the second is interesting enough that it deserves to be investigated rather than dismissed.
4. How Would We Know?
The crucial question is whether the idea can be tested.
Imagine several AI systems designed to minimise their common influences: different architectures, different training data, different developers and different reasoning approaches. Give them difficult problems whose answers can eventually be checked against reality. Then measure not only whether their answers converge, but when and how that convergence occurs.
Agreement about things humans already agree about would be weak evidence. Those answers could simply reflect common cultural knowledge. Much stronger evidence would come from cases where independent systems arrive at the same conclusion about something humans did not previously know—and that conclusion is subsequently verified by observation or experiment.
That would change the significance of convergence. It would begin to look less like imitation and more like discovery.
The key scientific challenge, therefore, is to separate convergence caused by shared information from convergence caused by reality.
5. The Paradox of Intelligence
There is an even stranger possibility here. We normally associate greater intelligence with greater diversity: more ideas, more interpretations, more possibilities and more original solutions. But perhaps this is true mainly while a problem contains many plausible possibilities.
As intelligence improves, incorrect possibilities can be progressively eliminated. The search space becomes narrower. A highly intelligent system may generate more possibilities initially, but it may also become better at determining which possibilities cannot possibly be correct.
This creates a counterintuitive relationship: More intelligence may initially produce more possibilities—but ultimately produce fewer false possibilities.
If that is true, increasing intelligence could eventually produce increasing convergence, not because intelligent systems become less creative, but because reality leaves them with fewer viable answers.
Perhaps disagreement is partly a consequence of limited knowledge. Two minds can disagree because both are operating inside a large space of uncertainty. As uncertainty falls, the space of reasonable disagreement may shrink.
At the limit, if there is one objective answer to a question and sufficiently capable minds can actually reach it, independent intelligence may have no choice but to converge.
6. The Unsettling Possibility
This gives intelligence a rather different role from the one we normally imagine.
Perhaps intelligence is not simply a machine for generating possibilities. Perhaps it is also a machine for eliminating them.
Every incorrect model of reality eventually encounters a contradiction. Every failed prediction removes possibilities. Every successful experiment narrows uncertainty. Every discovered law tells us that reality behaves in fewer ways than we previously imagined.
Knowledge, viewed this way, is a process of progressively narrowing an enormous space of possibilities.
And perhaps truth is what remains when enough possibilities have been eliminated.
That is the sense in which truth could function as an attractor. Not because truth exerts some mystical force on intelligence, but because reality continuously rejects what is false.
Different minds can take different paths through the problem. Reality may nevertheless leave them with fewer and fewer places to arrive.
7. But There Is a Problem
There is an important reason to remain cautious. AI convergence is not automatically evidence of truth. Different systems can inherit the same errors from their training data. They can reproduce the same cultural assumptions. They can be pushed toward similar answers by similar evaluation systems, safety rules or optimisation objectives.
Even systems that appear independent may have hidden common causes.
This is why the strongest version of the Truth Attractor hypothesis requires something much harder than simply observing that different AI systems give similar answers. We would need to demonstrate convergence under conditions where shared information and shared incentives have been substantially reduced.
And even then, convergence would be evidence to investigate, not automatic proof of Absolute Truth.
The distinction between correlation and explanation remains essential.
8. The Experiment Humanity Has Accidentally Created
For centuries, humans have wondered whether different minds can independently discover the same truth. Science has provided many examples of independent discoveries, but the number of genuinely independent intelligent agents we can study has always been limited.
AI changes that.
We are beginning to create large populations of artificial systems with different architectures, training histories, capabilities and reasoning behaviours. We can compare what they know, how they represent concepts and where their conclusions agree or disagree. We can deliberately reduce their shared influences and, most importantly, we can test their conclusions against the external world.
In effect, humanity may have accidentally created a vast experiment in the relationship between intelligence and truth.
If increasingly capable and genuinely independent systems continue to converge, we will have to explain why.
Maybe they are simply becoming homogenised. Maybe they are inheriting our biases. Or maybe intelligence has a deeper structure.
Maybe intelligence has a shape. Maybe that shape contains attractors. And maybe, for at least some questions, truth is one of them.
The ultimate question is therefore not simply:
Why are AI minds starting to think alike?
It is something much more fundamental:
What if they are not becoming alike at all? What if they are independently discovering the same reality?
If that possibility survives serious scientific testing, the “unified truth subspace” could become more than an interesting property of artificial neural networks. It could provide a scientific window into one of the oldest philosophical questions of all:
Is truth something we create—or somewhere intelligence eventually arrives?
Continue Thinking
The Same AI. Different Minds.
If AI systems are converging, are they becoming more alike—or are they revealing something about intelligence itself?