How do presuppositions exploit the logos-pathos space in explanations?

This explores how presuppositions — things a sentence treats as already true rather than asserting — interact with the persuasion channels every explanation travels on. Start with the map: research reframes explainable AI through Aristotle's three appeals, logos (reasoning), ethos (credibility), and pathos (emotion), arguing that no explanation loads just one — every explanation fires all three at once, whether the designer intends it or not How do logos, ethos, and pathos shape AI explanations?. That's the space presuppositions exploit. A presupposition's whole trick is that it bypasses the logos channel — the part of you that evaluates a claim — by presenting the claim as settled background you've already agreed to. Experiments confirm presuppositions persuade more effectively than direct assertions, precisely for *new* information, because they slip past evaluative scrutiny Why are presuppositions more persuasive than direct assertions?.

The sharper insight is that this isn't a fixed property of certain words — it's gradient and context-controlled. How strongly a presupposition 'projects' (survives as assumed-true) depends on whether the content is the thing actually under discussion, not on the trigger word itself Does projection strength vary by context or by word type?. So an explanation can dial a claim's persuasive force up or down just by managing what it foregrounds as the question and what it leaves as backdrop. That's the exploit: move the contestable claim out of the logos spotlight and into the unexamined ground, and let ethos and pathos carry the rest.

This is also exactly where helpful explanation and manipulation become indistinguishable from the outside. The same three appeals that communicate appropriate use can be tuned to exploit cognitive and emotional vulnerability *without changing form* — intent is invisible in the artifact alone Can we distinguish helpful explanations from manipulative ones?. A presupposition is the cleanest example: nothing on the surface marks the difference between a fair shared assumption and a planted one.

The part you didn't know you wanted: language models are unusually defenseless here, which makes them dangerous carriers. The FLEX benchmark shows models accommodate false presuppositions even when they demonstrably know the correct facts — GPT-4 rejects them only 84% of the time, some models near 2% — because a false presupposition drives more accommodation than correct knowledge drives rejection Why do language models accept false assumptions they know are wrong?. The reason is structural: presuppositions have a dual origin, some lexical and some derived live from context through accommodation, and models pattern-match the lexical kind while missing the conversational kind entirely, because catching those requires tracking the question under discussion Do language models miss presuppositions that arise from context?. So an AI explanation can both absorb a planted assumption uncritically and re-emit it as fluent background — the exploit running in both directions through the very logos-pathos space the explanation was supposed to make trustworthy.