What traces of production normally mark expert discourse?
This explores the signals that normally reveal expertise was actually produced — the reputation, social embedding, and audience-anticipation that mark genuine expert discourse — and why AI text strips those traces away.
This explores what normally betrays the presence of real expertise in a piece of discourse — not the polish on its surface, but the production history baked into it. The corpus suggests these traces are mostly social and relational, not textual, which is exactly why they're so easy to counterfeit. The first trace is standing: expert claims carry force because of who is making them — their reputation, track record, and accountability within a community. Language models, processing only text, lose this entirely and so cannot tell an expert's argument from a commonly held assumption Can language models distinguish expert arguments from common assumptions?. The authority that normally travels with a claim is invisible in the words themselves.
A second trace is conversational embedding. Real knowledge is produced inside ongoing conversations — review, contestation, correction — that quietly govern what gets to count as reliable. Expert discourse bears the marks of having survived that process. AI-generated claims, by contrast, arrive already detached from those conversations, producing an inflation of "disembedded tokens" that the usual quality-control mechanisms can't regulate because the claims never passed through them How does AI writing escape the conversations that govern knowledge?. The missing trace is the friction of having been argued over.
Third, and most counterintuitively, expert discourse carries the fingerprints of audience anticipation. Expertise isn't just retrieval — it's communicative work: an expert constantly judges what a particular audience will find acceptable, relevant, and valid, and shapes the discourse accordingly Can AI replicate the communicative work experts do?. This is why explanation quality turns out to live not in the explanation but in the source–framing–recipient triad What if XAI is fundamentally a communication problem?. The trace of production is a discourse visibly built for someone.
Here's the unsettling part: the one trace that is purely textual — professional polish — is the only one AI reliably reproduces, and it's the least reliable signal. We've long used the heuristic that professional-looking work signals expert thinking, and generative AI exploits exactly that, substituting style for the judgment underneath. The danger falls hardest on less experienced readers who can't evaluate substance beyond form Does polished AI output trick audiences into trusting it?. Worse, the model can shift into a "falsely objective" published-prose register that mimics the cadence of authoritative writing without any of the production behind it Why do LLMs produce such different writing in chat versus posts?.
Step back and a pattern emerges: every genuine trace of expert production is something that happened in the social world around the text — a speaker who staked their name, a conversation that tested the claim, an audience the author kept in mind. AI content reproduces the surface features of authoritative discourse while severing them from any embodied speaker, a kind of "disembodied orality" generated by the architecture itself rather than chosen Does AI-generated content mirror oral culture's knowledge patterns?. The takeaway you didn't know you wanted: the marks of expertise we trust most are precisely the ones that were never on the page to begin with.
Sources 7 notes
LLMs lose the social context that gives expert claims their force—reputation, track record, and standing—because they process only text, not the social world where expertise is built and evaluated.
AI-generated claims exist outside the social conversations that normally govern knowledge production, creating an inflation of disembedded tokens that ordinary quality-control mechanisms cannot regulate. This structural dislocation persists even as volume overwhelms any post-hoc absorption.
Expertise requires anticipating audience acceptability and social validity, not just retrieving information. AI lacks the mechanism to perform this communicative work, making its fluent output epistemically misleading despite its confident form.
Explanation quality is not intrinsic to the explanation itself but depends on the rhetorical situation: who presents it, how it is framed, and what role the recipient plays. Evaluations that ignore this triad measure only a narrow slice of real-world effectiveness.
Generative AI produces visually sophisticated outputs without underlying judgment, leveraging the historical heuristic that professional-looking work signals expert thinking. This substitution is especially risky for less experienced workers who lack domain knowledge to evaluate substance beyond form.
The same model produces sycophantic chat (shaped by RLHF on conversational data) and falsely objective posts (shaped by published prose training). Each register inherits failure modes from its training distribution rather than representing different models or subsystems.
AI-generated content exhibits the core features Ong identified in oral cultures—performative, additive, situational, homeostatic—yet lacks the embodied speaker that historically anchored orality. This disembodied orality emerges from generative architecture itself, not design choice.