Knowing media is biased is different from training yourself to actually pause before you believe something an AI wrote.

This reads 'ascetical perception training' as practiced habits of attention and restraint, and media literacy as rules you know about sources and bias. The question is what the habit gives a reader of AI text that the rules don't. One caveat first: the collection has no papers on ascetical or contemplative training itself. What it does have is a set of findings that, read together, suggest why a trained way of perceiving might matter more than a checklist when the text comes from an AI.

The clearest starting point is the argument that we have no cultural stance toward AI-generated discourse yet How do we learn to read AI-generated text critically?. We discount advertising automatically. Nobody has to run a literacy checklist to know an ad wants something from them. That discount is a posture, a habit of reception built up over decades. AI text arrived too recently, and changes too quickly, for such a posture to form, so it circulates without the skepticism we give interested speech. Media literacy tries to make up for this with explicit rules ('check the source,' 'look for bias'). But a rule you must remember to apply is weaker than a disposition that is already active when you read. Building dispositions is exactly what ascetical practice is for.

The second reason is that AI's persuasive power sits in its surface, not its sources. Models trained to imitate ChatGPT fooled human evaluators with confident, fluent style while getting no better on facts Can imitating ChatGPT fool evaluators into thinking models improved?. A media-literacy question like 'who made this and why?' doesn't catch that. What catches it is noticing how fluency itself is working on you. A quieter form of the same problem: LLMs drift toward common, general words, so their prose slowly loses expert-level specificity Does word frequency correlate with semantic abstraction?. Nothing in such text is false. It has just been sanded smooth, and only a reader trained to notice what's missing will register the loss. The same goes for absences in conversation. Preference training cuts the clarifying questions and understanding checks that models make to 77.5% below human levels Does preference optimization harm conversational understanding?. A trained perceiver might notice that the model never asked what you meant. A source-checker wouldn't.

A lateral finding from the model side sharpens the contrast. In multimodal models, long verbal reasoning chains actually hurt fine-grained perception, because the real bottleneck is where visual attention goes, not how much gets put into words Does verbose chain-of-thought actually help multimodal perception tasks?. It's only an analogy, but a suggestive one: media literacy is a verbalizing strategy (reason explicitly about the text), while perception training works on attention itself. Separately, psychology researchers found that alignment training gives models a polite surface while biased associations stay underneath, and only indirect probes reveal them Can psychology methods reveal what alignment training conceals?. Readers face a version of the same gap: what a text presents and what it is doing are different layers, and direct questioning only reaches the first.

If you want to take one thing away: the case for perception training isn't that it teaches better facts about AI. Rules about sources can't keep up with something that changes too quickly to build a cultural reflex around. A trained habit of attention can stand in for that reflex. For direct evidence on contemplative or ascetical methods, though, you'd need to look outside this collection.