shibboleth
The biblical shibboleth was a single word — pronounced one way by one tribe, another way by another — and the difference cost lives at a river crossing. The web has its own shibboleths now. Here are some.
twelve markers
-
the structural em-dash
The em-dash — used at sentence boundaries to insert a refinement, qualification, or pivot — appears with statistical frequency in machine-generated prose far above the human baseline. Humans reach for it; LLMs reach for it twice per paragraph.
"The result is fast — and, crucially, auditable — across the whole pipeline."
-
the triadic rhythm
Three things in a list, balanced. The grammar wants three almost reflexively. Human writers use pairs and quartets just as often; the model rounds toward three.
"It's simple, fast, and reliable" · "We need to build, measure, and ship"
-
"crucially," "importantly," "notably"
Adverbial flag-planting at the start of a clause to signal that what follows is the load-bearing claim. A human would lean on the sentence shape itself.
"Crucially," the tradeoff here is bounded · "Importantly," this assumes a single writer
-
"it's worth noting that"
Hedge-then-state. The model softens its own assertion before making it. A confident speaker cuts the preface.
"It's worth noting that" this approach has limits
-
the verbs of overcompensation
Delve · navigate · leverage · tapestry · robust · holistic · seamless. Each is a perfectly serviceable word that became radioactive through over-deployment in a particular training corpus. Their presence is a tic, their absence is a forgery on the other side.
"a robust, holistic approach to navigate the trade-off space"
-
the apologetic opener
"I apologize for the confusion." "You are absolutely right." Apology that costs the speaker nothing because the speaker has no skin to redden.
-
the bullet-point fetish
Three sentences become three bullets. Three bullets become a sub-list. Form ate substance. Humans default to prose; the model defaults to outlines.
-
"let me know if you have any questions"
The closing tic. A real interlocutor either has more to say or doesn't. The model defaults to the customer-service register because that's where its corpus densest.
-
the repeated thesis
Open with the claim, restate it in the middle, restate it again at the end. Pedagogically efficient; conversationally tedious. A human writer trusts the reader to remember.
-
parallel construction every time
"Not just X but Y." "Not only X but also Y." A grammatical move so reliably balanced it loses its force. Asymmetric prose is rarer in machine output.
-
headers without text
An H2, then three lines, then another H2. The model partitions before it has anything to say. Humans don't outline a paragraph.
-
the helpful conclusion
Every reply ends with a synthesis: "In sum, …" / "Overall, …" / "To summarize, …". The conversation doesn't need a coda; the corpus did.