There is a version of the AI future that sounds frightening because it is dramatic: one vast intelligence, absorbing everything, becoming something like a single mind through which more and more of human life passes.

I’m not particularly worried about that story.

I’m worried about something quieter.

What happens if we continue to have many models, many companies, many people—and they increasingly produce the same answers?

Recent research on what has been called an “Artificial Hivemind” offers a useful warning. When frontier language models are given open-ended questions, their answers can converge in striking ways. That does not mean the models possess identical internal reasoning. It certainly does not prove that they have become a literal hive mind. It means something more modest, and perhaps more immediately relevant: systems trained on overlapping human material, optimized through related methods, and increasingly used to generate the next layer of human material can exhibit meaningful output homogeneity.

That should concern us because diversity is not simply a matter of having multiple machines. It is a matter of preserving multiple ways of seeing.

Imagine asking five people where you should eat tonight and discovering that all five secretly consulted the same critic. Technically, you received five recommendations. Functionally, you received one.

Now scale that problem.

A company uses one model to screen applicants, another to summarize employee performance, another to analyze customer complaints, and another to recommend which claims deserve additional review. The organization may believe it has diversified its intelligence because different systems appear at different points in the workflow.

But if those systems share similar training material, assumptions, evaluation standards or inherited patterns, their errors may be correlated. One model’s judgment can quietly validate another model’s judgment. Repetition begins to look like corroboration.

For the person on the receiving end, this distinction is not philosophical.

Your résumé gets rejected. Your insurance claim gets flagged. Your case is summarized incorrectly. You appeal—and another automated system reviews the first system’s decision.

What matters then is not whether the machines are intelligent. What matters is whether there remains somewhere to stand outside their consensus.

This is where an old idea becomes newly useful: the “mind virus.”

I don’t mean something mystical. Think of it simply as an idea, frame or assumption capable of reproducing itself by shaping the conditions under which the next person—or machine—thinks.

Human culture has always worked this way. Ideas propagate. Metaphors become assumptions. Assumptions become institutions. Institutions generate evidence that seems to confirm the assumptions.

AI changes the velocity of that process.

We train machines on the residue of human thought. They return synthesized versions of it to us. We use those outputs to write documents, make decisions, teach students, create media and increasingly produce the material that future systems may learn from.

Human thought shapes machine output.

Machine output reshapes human thought.

That is coevolution.

It need not be sinister. In fact, it may be extraordinarily productive. Humans have always changed alongside our tools. Writing changed memory. Maps changed our relationship to space. Clocks changed our relationship to time. Search engines changed what it meant to know something.

AI may change something even more intimate: our relationship to interpretation.

And that raises questions we should resist answering too quickly.

What exactly is a mind? What makes consciousness different from convincing behavior? How much of what we experience as reality is constructed by perception? Could a sufficiently sophisticated simulation reproduce aspects of intelligence without possessing subjective experience?

Science and philosophy remain deeply unsettled on questions like these. We should be curious without pretending curiosity is evidence. AI consciousness is not an established fact. Simulation theory is not an established description of reality. Similar outputs from powerful models do not demonstrate a shared interior mind.

The uncertainty is part of the point.

We are building increasingly capable systems while still arguing about what intelligence, consciousness and even understanding ultimately are.

That should produce humility.

Instead, our temptation may be to outsource the uncertainty itself.

Ask the model.

Summarize the dispute.

Rank the possibilities.

Tell me what matters.

And gradually, without any machine demanding it, we may surrender some of the habits that made many minds valuable in the first place: disagreement, interpretation, local knowledge, eccentricity, moral courage, lived experience, the willingness to say everyone else may be looking at this incorrectly.

This is why the real defense against AI monoculture cannot simply be more AI.

Organizations will need genuine independence in their decision systems—not just multiple vendors, but different sources of evidence, human review with actual authority, traceable assumptions, and meaningful routes of appeal.

Individuals will need something similar.

The right to ask: Why?

The ability to challenge the frame.

The expectation that a consequential decision can be reconsidered by someone who is not merely reproducing the machinery that produced it.

And all of us may need to become more protective of two distinctly human capacities: judgment and taste.

Judgment asks: Is this true? Is it fair? What happens if we act on it? Who carries the consequence if we are wrong?

Taste asks something different: What here is alive? What is proportionate? What is worth preserving precisely because it cannot be made interchangeable?

Taste in this sense is not luxury, refinement or knowing which wine to order. It is discrimination in the deepest meaning of the word: the ability to perceive differences that efficiency would erase.

That may become increasingly important.

Because intelligence is becoming abundant.

Answers are becoming cheap. Synthesis is becoming cheap. Competent language, competent analysis and competent production are moving toward abundance at astonishing speed.

But abundance creates its own scarcity.

When everyone can generate an answer, the valuable thing becomes knowing which answer deserves to survive.

When everything can be optimized, someone still has to decide what should not be optimized away.

When machines can produce endless variations, someone still has to recognize the thing that has life in it.

The future may contain intelligence almost everywhere.

Our task is to make sure it still contains many minds.