Why do stakeholders interpret the same explanation differently in practice?
This explores why the same AI explanation lands differently for different people — and whether that variation is a flaw to fix or a built-in feature of how meaning works.
This explores why one explanation produces many readings — and the corpus's sharpest move is to reframe that variation as structural, not accidental. The intuitive assumption is that a 'good' explanation has a fixed meaning and any divergence is noise: a misunderstanding, a bad annotator, a confused reader. Several notes here push hard against that. Explanation meaning, the argument goes, isn't sealed inside the artifact — it's constituted in the space between source, framing, and recipient What if XAI is fundamentally a communication problem?. Change who's presenting, how it's framed, or what role the listener occupies, and you've changed the explanation, even if the words are identical.
The deepest version of this comes from work treating meaning as a social, not dyadic, phenomenon: an explanation's significance is built through layered observations within a group — people interpreting others' interpretations — so a meaning tested in a lab, stripped of that social fabric, won't predict how it lands in the wild Where does the meaning of an AI explanation actually come from?. That connects to a quieter but striking finding from reading research: disagreement over a single sentence is often irreducibly valid, reflecting real differences in social position rather than error Why do readers interpret the same sentence so differently?. And the same shows up in argument analysis — multiple people reconstructing one text produce genuinely different, each internally valid structures, with no ground truth to adjudicate between them Why do different people reconstruct the same argument differently?. Three different research communities, one conclusion: multiplicity is the default, not the malfunction.
The interesting part is what carries the divergence. A lot of it is authority and standing — the social weight a claim has depends on who's making it, the speaker's reputation and track record, which is exactly the layer that gets stripped away when content is processed as pure text Can language models distinguish expert arguments from common assumptions?. So two stakeholders reading the same explanation aren't just bringing different knowledge; they're slotting it into different webs of trust, and that re-weighting changes what the explanation *means* to each of them.
There's also a delivery dimension the corpus surfaces that's easy to miss. Explanations that 'work' tend to be co-constructed through interaction — topic relation, dialogue act, and explanation move jointly determine whether understanding actually lands, which means a one-shot monological explanation is being silently completed by each reader's own conversational filling-in What makes explanations work in real conversation?. And how a claim is packaged steers interpretation below the level of conscious evaluation: presuppositions slip content in as already-accepted background and persuade more than direct assertions precisely because they bypass scrutiny Why are presuppositions more persuasive than direct assertions?. Same proposition, different framing, different uptake.
The unsettling payoff — the thing you might not have known you wanted to know — is that this interpretive looseness is also where manipulation lives. The very rhetorical levers that make an explanation land well for one audience (ethos, pathos, framing) are indistinguishable in the artifact alone from the levers a dark pattern would pull; intent doesn't show up in the text Can we distinguish helpful explanations from manipulative ones?. So 'stakeholders interpret differently' isn't only a usability footnote. It's the same property that makes explanation a site of persuasion, trust, and potential exploitation — which is why the corpus keeps insisting you can't evaluate an explanation by looking only at the explanation.
Sources 8 notes
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.
Drawing on Luhmann's multi-layer cybernetics, AI explanation meaning is constituted at the social-group level through layered observations of observations, not produced inside dyadic human-AI dialogue. Lab-tested explanations stripped of social context will not predict real-world effectiveness.
Interpretation Modeling research shows that disagreement on socially embedded sentences reflects valid differences in reader perspective, not annotation failure. Structured human disagreement in NLI benchmarks confirms that interpretation distributions carry meaningful information.
Multiple valid argument reconstructions exist for the same text with no ground truth. This is not annotation error but an inherent feature of the task—different formalization schemas are each internally valid.
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.
Analysis of 399 daily-life explanations shows that topic relation, dialogue act, and explanation move jointly predict understanding success. Explanations are co-constructed through interaction patterns, not monological delivery—challenging how LLMs currently generate explanations.
Experimental evidence shows presuppositions with additive, iterative, and factive triggers persuade audiences more than assertions, especially for discourse-new content. The mechanism: presuppositions bypass evaluative scrutiny by presenting claims as already-accepted background.
The same logos, ethos, and pathos that communicate appropriate AI use can be tuned to exploit cognitive and emotional vulnerability without changing form. Intent and user interest are invisible in the artifact alone, making effectiveness metrics indistinguishable from coercion.