The language of artificial intelligence is the language of construction. We build models and ship them; we speak of architectures, pipelines, foundations, stacks. The metaphor flatters us, because construction implies control: buildings stay where you put them, do what the blueprint says, and hold their shape in a storm. Nothing about modern AI behaves this way. We have named the work wrongly, and the wrong name is shaping how we do it.
What the work actually resembles is gardening.
The daily disciplines are the same. A gardener does not make plants grow. She prepares ground, chooses what goes where, prunes, waits, watches the weather, and accepts that what comes up was never fully hers to decide. A large model is not assembled from parts; it is trained, which is to say cultivated. Much of what comes up is what you hoped for. Some of it is ivy.
Start with the soil. A model is grown from data the way a garden is grown from ground, and ground has a history. Whatever was spilled there decades ago is still in it, invisible, waiting to come up through the roots. This is not speculation; it is by now one of the best-replicated findings in the field. Models that pass explicit bias tests still carry implicit associations that surface in their open-ended behaviour: in the stories they tell, the candidates they imagine, the risks they quietly weigh. Nobody plants these biases deliberately. Nobody plants bindweed deliberately either. It arrives with the soil, and the honest response is the gardener's: know your ground, test it often, and keep pulling — knowing you will never be done.
Then there is pruning, the least celebrated of the garden's disciplines and the most important, because it is the art of saying no to growth. In the current AI economy, growth is the whole religion: more parameters, more users, more capability, shipped sooner. Evaluation, red-teaming, the unglamorous meeting where somebody asks "should we, though?" — all of it is treated as a brake on the real work. A gardener knows better. An unpruned tree is not a greater tree. It is a tangle that shades out everything beneath it and splits in the first serious wind. The companies that skip the secateurs do not avoid the pruning; they defer it to the public, who encounter it as a scandal, recall, or an act of parliament.
There is a soil lesson too in what the internet is becoming. Plant a single crop across an entire landscape and you get an efficient harvest and a fragile one: a single blight takes everything. The parallel is becoming uncomfortably literal. In 2024, researchers writing in Nature showed that when models are trained repeatedly on the output of other models, they degrade: the tails of the distribution vanish, the strange and rare disappear, and the result curdles into a bland average of itself. They called it model collapse. Old gardeners had a name for this too: exhausted soil. The remedy has been understood for centuries. Rotate the crops. Leave fields fallow. Plant unlike things side by side — the old gardens grew roses next to garlic and let each protect the other. In our terms: guard the human and the heterodox in the training mix, fund the strange interdisciplinary work, and treat intellectual diversity not as a courtesy but as a soil amendment.
And over everything, the weather. You can prepare the ground, prune with discipline, plant with judgement, and a late frost will still take the lot. The frontier of this field runs on emergence: capabilities appearing that no one explicitly trained, behaviours arriving unannounced at scale. The builders themselves say plainly that they do not fully understand the systems they are growing. This is often reported as a confession of failure. It is better understood as the basic condition of working with living complexity, and it calls for the gardener's posture of attention rather than the engineer's of mastery. It is fair to ask what a metaphor actually buys. Gardens will not pass an audit, and no regulator will accept "late frost" as a root-cause analysis. But names set expectations, and expectations set budgets. An engineer expects, one day, to be finished; the word "launch" contains a small promise that the hard part is over. A gardener knows there is no finished, only tended and untended — because in gardens, as in models, catastrophe usually announces itself first as a small strangeness in the leaves.
None of this argues for slowness as a virtue in itself, or for wonder as a substitute for rigour. Gardens are rigorous. They simply locate the rigour differently: in preparation rather than prediction, in maintenance rather than launch, in the long habit of care rather than the single act of creation. The tall things will keep being built, and should be. But if the next decade of artificial intelligence goes well, I suspect it will be because enough of the people growing it learned to think less like engineers on a deadline and more like gardeners in March: patient, attentive, respectful of the ground, and entirely unsentimental with the secateurs.