The key point
Machine AI imitates the outputs of intelligence on silicon. Xenocortical tissue is a small piece of the biological machinery of intelligence, running inside an animal. They are opposite bets about what a mind is.
Two projects that sound similar and are not
In 2026, two research programs are circling the same mystery from opposite directions. One trains machine intelligence: stacking trillions of artificial parameters on GPUs until systems can write, reason and converse. The other grows it: putting four million living human neurons into a mouse and letting them wire themselves into a body [1].
Both get called “artificial intelligence” in casual conversation, and both raise the same polite-society question — could this thing ever really think? — but they answer it with different machinery, different risks and different histories. Turing set the terms for the silicon side in 1950: can a machine’s behavior be distinguished from a mind’s? [10] The tissue side never asked that question, because its starting material is not an imitation of neurons. It is neurons.
Substrate: engineered versus grown
Machine AI is built from the outside in. Engineers design the architecture, choose the data, and run the training. The result can be copied, forked, versioned and rolled back. Whatever the system ends up being, it is the product of deliberate design choices.
Xenocortical tissue is the opposite: no one designs it. The cells were grown from human DNA and then left to organize themselves, using developmental programs that predate the species. The researchers could place the tissue, feed it, and watch — they could not specify its wiring. That self-organization is the whole scientific value of the model, and it is also why its outcome cannot be fully controlled [1].
Learning: data versus body
A large language model learns from oceans of text, with no body, no senses and no lifetime. Its knowledge is statistical: patterns in what was written. It can be superhuman in narrow domains and bafflingly alien in others.
The xenocortical mouse learns the way every animal does — from a body. The human tissue sits inside an animal that moves cautiously, explores, forgets, and improves. Human neurons there hooked into the mouse’s blood supply and spinal cord [1], which means they receive signals shaped by gravity, hunger, and injury. If the tissue ever develops anything like perception, it will be perception of a world — not of a training corpus.
Scale: trillions versus four million
The comparison in numbers is almost comic. Frontier machine systems train on trillions of parameters and petabytes of text. The xenocortical graft is about 4 million cells — roughly half of one mouse brain — and its tissue is immature, comparable to the middle of human fetal development [1].
But the scaling logic is inverted. Machine systems scale by more data and more silicon, with no principled ceiling. The tissue scales the way brains do: by development, maturation and integration, which are slow, biological and bounded — and which the study’s authors argue is exactly what makes it worth studying, because it runs on the brain’s own rules [1].
The consciousness question, on both sides
For machine AI, the question is whether any arrangement of computation, however clever, could constitute experience — a debate with no empirical grip. For xenocortical tissue, the question is narrower and closer to hand: this is human, neural, embodied tissue in a behaving animal. Today it is too immature and too poorly wired for its moral status to differ from the mouse’s own [1][6]. But the same papers that celebrate the model also insist the tissue must be monitored, because the trajectory points toward more maturity, more cells, more integration [2][7].
The risks people actually worry about
The two fields fear different failures. Machine AI’s debates are about misuse, control and alignment: a system that is competent but not oriented toward its users’ interests. The xenocortical debates are about welfare and moral status: what we owe an animal carrying human neural tissue, and where the line sits [2][6].
There is also a quiet convergence. Machine learning is increasingly used to read the activity of biological neural tissue, and biological findings constantly inspire AI architectures. The xenocortical model could give AI researchers something they have never had: ground truth about how living human neural circuits learn — measured not in a dish, but in an animal. And AI tools may be what make sense of the tissue’s signals. The comparison, in the end, is less a contest than a pair of instruments pointed at the same question: what, physically, is a mind? See also The papers, annotated.
