# What the Receiver Rebuilds
# Book: Language
# Chapter URL: https://programmer.ie/books/language/02-chapter/
# Prose fingerprint (sha256, normalized): 92b10841c011e371b868af07e98268e091c80f58b5d55c89959fda1eba258543
# Snapshot generated at build time by Hugo. This is the complete rendered prose
# of the chapter; it is not a transcript of the source file.

Imagine two registrars receiving the same handover note about bed seven: keep an eye on him. The first covered the ward last week. She knows that this patient’s observations can drift before he deteriorates, that he under-reports pain, and that this colleague usually means check within the hour when she says those words. She reads the note and forms a plan. The second is new to the ward. She reads the same words and infers only that the patient needs some vigilance, with no particular deadline. Both have read the words accurately. But the first has inferred a specific action and the second has not. The note alone cannot establish whether either reconstruction matches the outgoing registrar’s intent or the patient’s needs.
The representation was identical. The reconstructions were not. One reader drew on patient-specific experience; the other did not. Both also had to guess what the sender expected them to do. The chapter’s question is not merely why interpretations differ, but which parts of each interpretation came from the message, which came from prior knowledge, and which remain unverified. Its claim is simple to state and awkward to absorb:
Receiving a representation is not the same as receiving its meaning. The receiver builds the meaning, and what they can build depends on what they bring.
representation ≠ meaning sent ≠ understood Everything the book later asks of a generated representation (a diagram, a brief, a timeline, an agent’s message) will be judged against this distinction. A representation can be accurate to its source yet leave a receiver unable to act; it can be fluent and helpful-looking while prompting the wrong inference. The question is not only whether an output looks right, but what a particular receiver can recover from it, what evidence supports that recovery, and whether it serves the declared purpose.
The pipe, and why it fails A powerful picture of communication sits underneath everyday talk about messages. A sender puts meaning into words, the words travel, the receiver takes the meaning out. Claude Shannon’s mathematical theory of communication (1948) formalised a different, narrower problem: how reliably information selected at one point can be reproduced at another despite the limits of the channel and the presence of noise. Shannon explicitly set semantic meaning aside because it was not required to solve that engineering problem. The theory is not a failed account of understanding; it was not trying to be one. A signal may arrive without error while the receiver still draws an inference the sender did not intend.
Warren Weaver, introducing the book version of the theory in 1949, separated three levels of problem. How accurately can the symbols be transmitted? How precisely do they convey the desired meaning? How effectively does the received meaning affect conduct in the desired way? Shannon supplies a mathematical treatment of the technical problem, not a general solution to the other two. All three levels can fail. This book concentrates on the second and third, while refusing to infer semantic success merely from technically successful delivery.
The trouble starts when the pipe is mistaken for a theory of understanding. If meaning travels inside words, then understanding is decoding, and a clearer message is simply one with less noise. But the words alone do not always determine the action the speaker intends. Someone says “it’s cold in here” and might be reporting a temperature, asking for the window to be closed, or complaining about a colleague. The philosopher Paul Grice drew attention to this gap between what is said and what is meant, and to the way listeners cross it by reasoning about the speaker: why would they say this, now, in this way? Dan Sperber and Deirdre Wilson built on the idea in Relevance (1986). On their account a speaker does not hand over a thought. A speaker provides evidence of what they intend, and the listener infers the most relevant interpretation consistent with that evidence, given everything else they know and expect.
On this view, understanding involves inference by the receiver. That is what the registrar did with four words. The inference can be reasonable and still wrong. Human conversation has a partial remedy that a one-way handover note lacks: the receiver can ask “Do you mean check him within the hour?” and the sender can confirm, correct or qualify. Feedback is part of communication, not evidence that the first representation succeeded.
Conceptual comparison. Perfect transmission of words does not establish that the receiver reconstructed what the sender intended. Open full-size diagram. What a receiver brings It helps to separate what a receiver brings, because the factors call for different responses. Some can be supplied by the receiver as part of the task; some are established through conversation; some may be unknown. An AI mediator needs a way to handle those differences, not a licence to infer a private profile from whatever traces happen to be available.
Prior knowledge. The first registrar knew the patient. The second knew medicine in general but not him. Knowledge supplies candidate connections that a short representation leaves unstated. It can also supply the wrong connection: a previous pattern may no longer apply, or a remembered convention may not be what this sender intended. Filling a gap and filling it correctly are different achievements.
Shared context. Herbert Clark’s common ground concerns what participants can reasonably treat as shared for the purposes of their exchange: earlier utterances, the visible situation and jointly established expectations. “The usual place” may work between friends and fail between strangers. Yet even friends can wrongly assume they have the same place in mind. Common ground is maintained and repaired through interaction; it is not identical to everything one participant knows, and a system should not silently treat private or inferred knowledge as shared.
Purpose. People read for something: to decide, to check, to learn, to act. The engineer reviewing an outage needs to judge whether the proposed fix is justified; the student reading the same report may need to explain the failure mechanism. Their questions, acceptable omissions and useful formats differ. Purpose supplies the success criterion, but it does not authorise the mediator to alter the underlying facts to make the task easier.
Condition. Time, attention and fatigue can change what a reader notices and retains. A person reading under severe time pressure may overlook a qualification that they would notice in a careful review. That is a risk to test, not a fact a mediator can safely deduce about this person from the time of day or the device they are using.
These differences are not defects in the receiver. They are conditions under which reconstruction happens. But different reconstructions are not automatically equally adequate. Where a distinction matters, the system needs a way to check it against the source, the sender’s clarified intent or an independently specified task criterion. Sometimes the right enhancement is not another representation but a question.
Conceptual model. The receiver’s contribution is not noise around a message; it is part of how the message becomes usable. Open full-size diagram. A model, not a copy What the receiver constructs is not necessarily a copy of the words. Research on discourse comprehension distinguishes the wording of a text from a reader’s model of the situation it describes. Walter Kintsch’s construction–integration account (1988) describes comprehension in terms of information from the text interacting with prior knowledge. Zwaan and Radvansky (1998) reviewed research on situation models: readers represent events, participants, goals and causal relations, not only the sentences used to describe them. We cannot inspect that representation directly. We can test aspects of it by asking a person to explain a cause, notice a contradiction, make a decision or transfer an idea to a new case. Each task supplies evidence about comprehension; none is an unmediated readout of a mind.
A finding from this tradition makes the role of prior knowledge concrete. McNamara, Kintsch and colleagues (1996) studied versions of science texts that differed in how explicitly they connected ideas. Readers with less background knowledge generally benefited from greater coherence. On some measures of deeper understanding, knowledgeable readers benefited from text that left more inferential work to them. The authors interpreted this as readers using their knowledge to supply connections. That is not evidence that obscure writing is better for experts, or that one can infer a person’s knowledge level from a clickstream. It is evidence, within those studies and measures, that a text’s effectiveness can depend on the interaction between its form and the reader’s knowledge. The design question is therefore which representation helps this receiver perform this task?—a question to be tested, not answered by a universal readability score.
Why confidence is not a comprehension test Two familiar errors make the whole process harder to see from the inside.
Receivers can feel they have understood more than they can explain. Rozenblit and Keil (2002) asked people to rate their understanding of familiar mechanisms, such as zippers and flush toilets, and then attempt an explanation. Their self-assessments often fell after that attempt: confidence had exceeded explanatory ability. The study does not establish a general effect of polished AI formatting on comprehension. It does establish why a satisfaction rating or a confident “I get it” cannot substitute for a task that probes what the reader can actually explain or use.
Senders, for their part, tend to overestimate how much of their intent has crossed. Keysar and Henly (2002) found that speakers believed their ambiguous statements had been understood more often than they were, because the speaker’s own knowledge of what they meant fills the ambiguity automatically. A well-known demonstration, from a doctoral study by Elizabeth Newton (1990), asked people to tap out the rhythm of a famous song for a listener to identify. In Newton’s tapping study, participants expected listeners to identify roughly half the tunes. The dissertation’s results section and Table 1 report just three correct identifications in 120 attempts (2.5%). The dissertation is internally inconsistent about the precise count, however: its abstract reports two in 150, and its later discussion reports three in 150. For the tapper the song plays in their head. For the listener there is only knocking. The tapper has access to a richer experienced whole than the tapping carries. Speakers are in a similar position: they may begin with an image, a relation, a direction or a partly formed thought, not a finished inner sentence, and the words they choose carry only some of it. The gap between what the sender has and what the representation carries is part of what the receiver has to rebuild across; no general theory of wordless thought is needed to see that.
Put the two together and a design lesson emerges for human and machine communication. The sender’s confidence and the receiver’s self-report are observations, but neither establishes that the intended distinctions survived. Better evidence comes from testing a consequence: can the receiver explain the critical relation, identify an exception or choose the action the task requires? Such tests need their own criteria and can also be imperfect. A successful answer is evidence about performance on that check, not proof of a complete mental reconstruction.
A target for everything that follows This gives the book a measurement target it can carry from here to the last chapter. First declare what the communication is meant to accomplish and which source-grounded distinctions must survive: a deadline, a causal direction, a qualification, an uncertainty, a prohibition. Then assess both the generated representation and the receiver’s observable response. The same four questions apply, but they must not be collapsed into one score:
SOURCE → REPRESENTATION → RECEIVER'S RECONSTRUCTION ├─ what survived? ├─ what changed? ├─ what was added? └─ what became easier? The first three questions concern preservation, distortion and additions; the fourth concerns what the receiver can now do. They are related but not interchangeable. A diagram might preserve every source claim and still confuse its reader. A shorter brief might make an important decision easier while quietly dropping a condition that would reverse the decision in another case. Gain therefore needs an outcome measure—accuracy on a declared question, detection of an exception, time or effort under a quality constraint—alongside a source-level check of what survived. A gain in usability is not permission to change the facts.
GAIN ≠ ADDITION A diagram that invents a causal link the source does not justify has added something unsupported, even if its layout is persuasive. A diagram that helps a reader identify a cycle already supported by the source may deliver a genuine gain. A mediator can sometimes add outside explanations or examples, but those are augmentations, not preservation: their provenance, standing and permission must be explicit, and their correctness must be checked separately.
Notice what this does to the idea of a good representation. It cannot be judged from the artifact alone: the same diagram may help one receiver and mislead another. The unit of evaluation is a representation for a task and a receiver, under declared preservation and authority constraints. This does not require a secret psychological profile. The mediator can begin with the user’s stated purpose, an explicit choice of detail level and a request for clarification when the necessary knowledge is missing. Any further inference about the person has to remain distinguishable from what they declared, open to correction, and within the permission the system actually has. The later Parts ask whether more adaptation earns its cost and risk rather than assuming that more personal data means better communication.
What can and cannot be claimed A note on limits, because the territory invites overreach. Reconstruction is not directly observable. What can be observed are its products: answers, decisions, errors, time taken, and what a person can do with the material later. The book’s vocabulary of source, representation and reconstruction is a working model for engineering and experiment, not a theory of how minds encode meaning. The cited research comes from psycholinguistics, philosophy of language and the psychology of comprehension; findings from their studied settings do not automatically transfer to AI-generated interfaces.
A bounded experiment would begin with one frozen source and a small set of independently specified questions about it. It would declare which claims, exceptions and relations must survive, then compare a source-faithful baseline with a generated representation on the same tasks. Measure source-level preservation separately from the receiver’s answers, error types, response time and optional later recall; record self-rated understanding as another variable, not as the success criterion. Vary representation while keeping the task fixed, and examine whether prior knowledge changes the result rather than presuming the direction of the effect. No such experiment is reported by this chapter. It is a design for what later evidence would need to establish.
Where this leaves us A representation does not deliver a ready-made mental model. It supplies evidence from which a receiver draws inferences using prior knowledge, context, purpose and present condition. Those inferences may be helpful, mistaken or simply uncertain. A useful mediator therefore has two separate obligations: preserve the source-grounded distinctions that matter, and test whether the chosen form improves an observable outcome for the intended receiver and task. Confidence on either side cannot discharge those obligations by itself. Clarification, an unchanged source view and an honest unknown must remain available alternatives to fluent transformation.
If the receiver rebuilds from whatever channels are available, then the channels matter. A timetable, a subway map and a spoken warning give a receiver very different things to build from, and the next question is what different kinds of representation make possible.
References Shannon, C.E. (1948). A Mathematical Theory of Communication. Bell System Technical Journal, 27, 379–423, 623–656. Shannon, C.E. & Weaver, W. (1949). The Mathematical Theory of Communication. Urbana: University of Illinois Press. (Weaver’s introductory essay distinguishes the technical, semantic and effectiveness problems.) Grice, H.P. (1975). Logic and Conversation. In P. Cole & J. Morgan (eds.), Syntax and Semantics 3: Speech Acts. New York: Academic Press, pp. 41–58. Sperber, D. & Wilson, D. (1986). Relevance: Communication and Cognition. Oxford: Blackwell. (2nd ed. 1995.) Clark, H.H. (1996). Using Language. Cambridge: Cambridge University Press. Kintsch, W. (1988). The role of knowledge in discourse comprehension: A construction-integration model. Psychological Review, 95(2), 163–182. Zwaan, R.A. & Radvansky, G.A. (1998). Situation models in language comprehension and memory. Psychological Bulletin, 123(2), 162–185. McNamara, D.S., Kintsch, E., Songer, N.B. & Kintsch, W. (1996). Are good texts always better? Interactions of text coherence, background knowledge, and levels of understanding in learning from text. Cognition and Instruction, 14(1), 1–43. Rozenblit, L. & Keil, F. (2002). The misunderstood limits of folk science: an illusion of explanatory depth. Cognitive Science, 26(5), 521–562. Keysar, B. & Henly, A.S. (2002). Speakers’ overestimation of their effectiveness. Psychological Science, 13(3), 207–212. Newton, E. L. (1990). The Rocky Road from Actions to Intentions. Doctoral dissertation, Stanford University. The results section and Table 1 (pp. 40–41) report 3/120; the abstract (p. ix) reports 2/150, and the later discussion (p. 44) reports 3/150. The source contains inconsistent counts.