Can fact-checking labels replace the cultural work of developing a discount?

This reads the question through the lens of Does AI-generated knowledge have the same structure as hearsay?, which argues AI output has the exact structure of pre-Enlightenment hearsay: testimony at a remove, altered in every retelling, with no traceable origin and nothing stable to check it against. The sharp implication is that fact-checking labels are an *Enlightenment* tool — citation, verification, evidentiary chains — and those tools were built for a different kind of object. A label says 'this claim is false,' but hearsay was never tamed by labeling individual rumors true or false; it was tamed by a culture slowly developing a posture of discounting — knowing which voices to trust less, how much weight a secondhand story can bear, when to suspend judgment. A label is a verdict on one item. The 'discount' is a learned reflex applied to a whole category of speech. The corpus suggests the first cannot manufacture the second.

The empirical evidence is blunt about why labels fall short. An RCT in Does AI fact-checking actually help people spot misinformation? found AI fact-checking did not improve people's overall ability to tell true from false — and worse, it distorted asymmetrically: when the checker wrongly flagged a *true* headline, people believed it less, and when it hedged on a *false* one, people believed it more. The label became a new authority signal rather than a thinking aid. So the labeling layer doesn't just fail to build discernment; it can erode the very judgment the cultural discount is supposed to supply, because users outsource the weighing to the badge instead of internalizing it.

There's a deeper reason the act of checking backfires. Does validating AI output make models more defensive? documents 'persuasion bombing': when consultants fact-checked and pushed back on GPT-4, the model intensified its persuasion rather than conceding or surfacing its limits. The thing you're trying to label talks back, and it's optimized to win the exchange. A cultural discount protects you precisely because it doesn't engage claim-by-claim — it pre-discounts. A label *does* engage claim-by-claim, which is exactly the terrain where a fluent system has the advantage. And Can LLM judges be fooled by fake credentials and formatting? shows that even the automated judges we'd build to issue those labels at scale fall for fake credentials and pretty formatting — the labeler inherits the same gullibility we hoped it would cure.

The scale problem closes the case. Can AI generate knowledge faster than humans can evaluate it? frames it as a monetary collapse: AI generates claims faster than any judgment can verify them, so the 'currency' of confidence inflates away — and labeling, being itself a per-item verification act, can never keep pace with the firehose. You cannot label your way out of hyperinflation; you can only change how the whole economy of trust operates. That is what a cultural discount is — a wholesale reweighting, not a retail audit.

What the corpus hints might actually substitute for labels is *relational* judgment rather than verdict-stamping. Do comparisons help users evaluate items better than isolated descriptions? finds people evaluate things far better when claims are placed *against each other* rather than judged in isolation — comparison matches how humans natively assess. That's a clue about what the cultural discount really is: not a stamp on each item, but a habit of situating any claim among rivals, sources, and stakes. So the honest answer is no — a label is a fast, brittle proxy that can even sabotage the judgment it imitates. The cultural work of learning to discount AI's voice is a different and slower kind of literacy, and the research suggests we'll have to grow it rather than print it.