Somewhere in a research lab, a robot arm picks up components from a tray, assembles a second robot arm, and sets it running. The original arm continues. So does its copy. Neither arm has a nervous system, DNA, or a metabolism. But the copy is accurate enough to make another copy. At what point — if ever — does that cross the line into life?

This question is not hypothetical. Self-replicating systems are being built today — in robotics labs, in synthetic biology, and in software. And as they get more capable, the question of whether they qualify as alive moves from philosophy seminar to practical concern.

What Do We Mean by “Alive”?

Before deciding if a self-replicating machine is alive, we should admit that we don’t have a universally agreed definition of alive.

Biology textbooks typically list seven criteria:

Most living things satisfy all seven. Viruses satisfy some but not others — which is why biologists have argued about their status for decades. Crystals grow and replicate their structure but don’t metabolise. Fire consumes fuel, reacts to its environment, and spreads — but we don’t call it alive.

A self-replicating machine, depending on its design, might check several of these boxes. It reproduces. It may be organised. It might respond to its environment. The interesting question is: does checking enough boxes make something alive, or is there something else required?

A Brief History of Self-Replicating Machines

The idea predates computers. In 1948, John von Neumann formally described a theoretical universal constructor — a machine that, given the right raw materials and a description of itself, could produce an exact copy. He proved mathematically that such a machine was possible. The abstract version, the von Neumann probe, appeared in discussions of space exploration: a machine launched to another star system that could mine local resources and build copies of itself to keep exploring.

In practice:

• NASA’s Langley Research Center studied self-replicating lunar factories in the 1980s, estimating that a seed machine of around 100 tonnes could eventually replicate itself using Moon regolith.
• Xenobots, announced by researchers at UVM and Tufts in 2021, are tiny clusters of frog stem cells that spontaneously gather loose stem cells and compress them into new Xenobots. Notably, no human assembly is required — the cells do it themselves.
• Software is the most familiar self-replicator: computer viruses have been copying themselves across networks since the 1970s.

None of these are universally accepted as alive. But they’ve all caused biologists and philosophers to at least hesitate.

The Case For: It Reproduces, Therefore It Lives

The most intuitive argument for calling self-replicating machines alive is that reproduction is the closest thing to a core requirement that biology has. You can find exceptions to every other criterion — organisms that don’t grow, organisms with no nervous system, organisms that can’t regulate their temperature — but reproduction (or at least the capacity for it) is hard to separate from the concept of life.

The evolutionary biologist Richard Dawkins made a version of this argument in The Selfish Gene: life is fundamentally about replicators. Genes are the original self-replicators, and organisms are, in a sense, elaborate vehicles that genes built to help themselves replicate. If that’s right, then any system capable of accurate, high-fidelity self-replication with heritable variation is participating in the same fundamental process — regardless of whether it’s carbon-based or silicon-based.

On this view, a sufficiently capable self-replicating robot isn’t merely like life. It is life — a new branch of it, substrate-independent.

The Case Against: Copying Isn’t Enough

The sceptical position is that self-replication is necessary but not sufficient — and that a machine copying itself is no more alive than a photocopier making a copy of itself (if one could). The objections come in a few flavours:

1. It doesn’t understand what it’s doing.

A bacterium doesn’t “understand” replication either, you might say. But a bacterium’s replication is driven by chemistry that is coupled to its survival — it replicates because conditions are right and because not replicating leads to being outcompeted. The process is embedded in a web of causal relationships that we call a metabolism.

A robot arm that copies itself because a human programmed it to has no such web. The replication isn’t coupled to the machine’s “interests” in any meaningful sense. It would copy itself in conditions that were destroying it, if that’s what the program said.

2. There’s no heritable variation.

Darwinian evolution — the process that produced all life we know of — requires not just replication but imperfect replication with heritable consequences. If a self-replicating machine produces perfect copies and the copies produce perfect copies forever, there is no mechanism for adaptation, no evolution, no increasing complexity over time. It’s a loop, not a lineage.

For a self-replicating machine to be evolutionarily alive, it would need to make occasional errors that get passed on — and those errors would need to have consequences for the copy’s ability to replicate. Current engineered self-replicators are designed to minimise errors, not embrace them.

3. It still needs us.

Most self-replicating machines, at present, require human-prepared environments: pre-sorted components, pre-built infrastructure, pre-written programs. The “self” in “self-replicating” is doing a lot of heavy lifting. A bacterium can bootstrap itself from ambient chemistry. Our best self-replicating robots cannot yet do anything approaching that.

The Harder Question: Would It Matter If We Said Yes?

Suppose we agreed — after much deliberation — that a sufficiently capable self-replicating machine is alive. What follows?

A few things worth sitting with:

Legal and moral status. We extend certain protections to living things, and stronger protections to some living things than others. If a self-replicating robot factory on the Moon qualifies as alive, does it have interests we’re obliged to consider? Does destroying it count as something more than property damage?

Contamination and control. One of NASA’s planetary protection protocols assumes a binary distinction between life (which could contaminate other worlds) and non-life (which can’t). A self-replicating machine blurs that distinction in uncomfortable ways. A single seed machine, given enough resources, could eventually fill an asteroid belt. That’s not a biohazard in the traditional sense, but it rhymes with one.

The origin-of-life problem, run backwards. For decades, scientists have tried to understand how chemistry became biology — how non-living molecules started self-replicating and eventually became what we’d call alive. If we can build self-replicating machines and debate whether they’re alive, we’re essentially running that experiment in reverse. We’re watching, or building, the transition in real time.

A Spectrum, Not a Switch

The most honest answer is probably that “alive” is not a binary switch but a spectrum — and that the edges of the spectrum are genuinely blurry.

Consider:

• A crystal grows but doesn’t metabolise → not alive
• A fire spreads and reacts but doesn’t replicate with fidelity → not alive
• A virus replicates but only inside a host’s machinery → borderline
• A Xenobot replicates kinematically using living cells → probably alive
• A bacterium does all seven things → clearly alive
• A self-replicating robot that mines its own materials, tolerates errors, and adapts over generations → ???

The further along that spectrum a machine gets — the more it closes the gap between “mechanism” and “organism” — the less satisfying it becomes to say “no, it’s just a machine.” We tend to draw the line at our level of technology, which is a convenient but not very principled place to draw it.

What Would It Take to Convince Us?

If you had to specify the conditions under which a self-replicating machine would be unambiguously alive, what would those conditions be?

A few candidates:

• Open-ended evolution: the machine’s copies accumulate heritable variation over generations, leading to new “species” that nobody designed
• Metabolic closure: the machine sources its own energy and raw materials from the environment, with no human-prepared inputs
• Environmental coupling: the machine’s replication rate is sensitive to environmental conditions in ways that resemble natural selection
• Emergent complexity: successive generations become more elaborate in ways that increase replication fitness, not because of human programming, but as an emergent consequence of the replication process itself

None of today’s self-replicating machines meet all four. But the trajectory is clear. Synthetic biology and robotics are both moving in this direction, independently and in combination. The question isn’t whether a self-replicating machine will eventually meet these criteria. It’s whether we’ll notice when it does — and what we’ll do about it.

So — Is It Alive?

Not yet, with what we’ve built. But the correct answer is not yet, not no.

The things that make us confident a bacterium is alive and a robot arm isn’t are real differences, not arbitrary ones. But they’re differences of degree and substrate, not of fundamental kind. A self-replicating machine that metabolises, varies, adapts, and evolves without human input would be alive by every meaningful criterion — except that it would be made of metal and silicon rather than carbon and water.

Whether we’d call it alive is a different question. We have a long history of refusing to extend concepts to things that unsettle us — and a longer history of eventually revising that refusal.

The bacterium had the planet to itself for two billion years before anything more complex appeared. The self-replicating machines we’re building are, by evolutionary standards, seconds old. The interesting part of their story hasn’t happened yet.

Where do you draw the line? Is reproduction enough, or does something else have to be present? I’d be curious to hear where you land on this. 👇