The Interface Between Minds

A registrar finishes a night shift and hands over a ward of twenty patients to the day team. She has perhaps ten minutes. In her head sits a rich model: which patient is stable, which one worried her at 3 a.m. without quite triggering an alarm, which family needs a careful conversation, which blood result is pending and would change everything. What she can transfer is far thinner: a few sentences per patient, some numbers, a gesture toward the man in bed seven: “keep an eye on him”. The day team will reconstruct what they can and get on with it. Mostly this works. Sometimes the thing she could not quite encode is the thing that mattered.

That gap, between what one mind holds and what another mind reconstructs, is the subject of this book. More precisely, the book asks what happens to that gap when intelligence enters the channel between them. Here “intelligence” covers two things: the human judgement already present in every exchange, and the computational systems that increasingly select, transform or generate messages inside it.

What form does communication take when intelligence becomes part of the medium through which meaning moves?

Everything else follows from taking this seriously. But the gap is not one thing, and it does not sit in one place. A gap that is one thing can only be described. A gap that is several things can be diagnosed, redesigned and audited. The first job of this chapter is to divide communication into its parts, to name the question that must be asked of each, and to show why an intelligent channel turns a familiar fact, that understanding is work, into an engineering problem with obligations and checks.

Three joints

Look closely at what happened in the handover, because the vocabulary of the whole book is drawn from it.

The registrar held beliefs, worries and priorities. Those were not themselves the message she intended to send. Her communicative intent was to leave the day team able to recognise the risks and act on them. The representation was the material that travelled: a handful of spoken sentences, a printed list and a glance. The receiving team performed reconstruction, building a working model of the ward from those materials plus what they already knew. What she knew, what she intended them to grasp, and what they actually recovered are related, but they are not identical. Ordinary speech reports this as a single event, “she told them”, and three things were in fact happening.

Each joint can be checked on its own terms. At the first, the question is whether the registrar understood what the day team needed in order to act. She may have assumed they knew the history of bed seven, when they did not. At the second, the question is whether the representation could carry the load. Ten minutes and a few sentences cannot carry a whole working model, however careful the speaker. At the third, the question is whether the receiver rebuilt what was intended. A nurse who hears “keep an eye on him” as a routine instruction has reconstructed something real, but not the worry behind it.

The representation is not the meaning. It is something the sender produces, which the receiver uses, with effort and background knowledge, to rebuild something close enough to act on. We speak as though meaning travels inside the words, like water in a pipe, and say “the message says”. It does not travel that way. When the rebuild succeeds we call it understanding, and we quickly forget how much work the receiver did. That forgetting is why the rest of the book is needed. A system that intervened in the channel while sharing the pipe picture would have no reason to check anything: if meaning were delivered intact, a fluent summary would be harmless by construction. It is not harmless. It is a representation, and a representation can fail.

Two further distinctions sit beneath the first, and they will recur.

expression  ≠  intent
word  ≠  concept

What someone says need not contain everything they meant, and a word does not carry a concept independently of the people and situation using it. The same expression can support different interpretations in different contexts. Shared vocabulary and established practice can also make an expression remarkably reliable, but the words alone do not guarantee either outcome. That is why reconstruction is work, and why successful communication has to be assessed against what a particular receiver needed to recover, not against some property of the sentence.

Here, concept is a working name for a distinction or relation that an expression may carry. It does not claim that minds store wordless tokens.

When communication fails, whether through the missed allergy, the misread tone, or the instruction followed literally but wrongly, the failure can sit at any of the three joints. The sender may have misjudged what the receiver already knew. The representation may have been unable to carry the load. The receiver may have reconstructed confidently and incorrectly. These are different failures with different remedies. A great deal of confusion about language, and about AI language in particular, comes from merging them. If you cannot say which joint failed, you cannot say what to change.

One question can be asked at every joint, and it returns throughout this book: what had to survive? Not everything. The registrar cannot transfer her whole working model of the ward, and should not try, since a complete transfer would bury the day team in detail it does not need. For this handover, the necessary content might be the concern about bed seven, the pending result, the uncertainty attached to each, and who must follow up. These are proposed preservation obligations: distinctions the receiving team must be able to recover for this task. Stating them is a human and operational judgement. A similarity score cannot infer them from the words alone. This is a boundary the book keeps, not a temporary limitation. A machine can compare two texts. Only a person with a purpose can say which differences would change the decision.

What the past shows, and what it doesn’t

Human beings coordinated in speech long before anyone wrote anything down. Speech carries far more than vocabulary: rhythm, emphasis, pause, gesture, gaze and shared surroundings. Writing made a different bargain. It let a message travel without its speaker, at the cost of stripping away most of that surrounding context. What remained had to be rebuilt by a reader who was not there.

Walter Ong’s Orality and Literacy (1982) is an influential account of how writing, and later print, changed not just storage but habits of thought: lists, definitions, and texts that could be compared word against word. It is worth reading, and worth reading sceptically. An oral–literate division can obscure the complexity of oral traditions and the persistence of speech within literate societies. The historical relationship also runs through schooling, institutions and power. The more defensible point is that different representations make different kinds of checking and coordination easier or harder. They do not determine what a society will think.

Printing extended the same logic. In The Printing Press as an Agent of Change (1979–1980), Elizabeth Eisenstein argued that standardisation, preservation and cumulative correction (identical maps, comparable editions, errata that could actually propagate) changed the conditions for the Renaissance, the Reformation and the sciences. Adrian Johns and others replied that early print was often corrupt, that careful manuscripts were common, and that fixity had to be built rather than assumed. Eisenstein answered that duplication still changed the conditions of accumulation even if every copy was imperfect. The dispute is more useful than either side alone. It shows a pattern this book will meet repeatedly. A new communication technology changes the constraints on what can be preserved, compared and corrected, but the consequences depend on institutions, incentives and use. Communication is one explanatory factor among several, never the single engine of history.

Read this way, the past teaches something narrower and more useful than a grand story of progress, and narrower than technological determinism in either direction. Writing can preserve a formulation while losing tone and immediate context. Print can make editions comparable while multiplying errors as well as corrections. Electronic media can carry voice and image while introducing new forms of selection and distortion. Each is a change in what can travel, how it can be inspected, and what a receiver must supply. None makes communication automatically faithful. The historical lesson, stated at the resolution the evidence supports, is that a medium changes the conditions of reconstruction, and the outcome still depends on people, purposes and institutions. Held loosely, the past checks one-sided claims: that a richer channel is automatically better, or that a new channel is automatically worse. It is not a prediction about the present.

Intelligent mediation, made explicit

Intelligent tools already take part in communication. Writing aids suggest replies and alter tone. Reading aids summarise and answer questions. Systems also select, translate and reorder what people encounter. Hancock, Naaman and Levy (2020) called one important class AI-mediated communication: computational agents that modify, augment or generate messages on behalf of communicators. That definition is about mediation, not merely about assistance at the endpoints. The proposal in this book therefore has to be more specific than saying that AI has entered a channel that was previously untouched.

The older channel was never wholly passive. Human editors and translators have long reshaped messages in transit. Search engines, filters and recommendation systems have long changed what reaches a reader. These mechanisms differ greatly from one another, and none of them establishes that it preserved the sender’s intent. What is new is whether context-sensitive transformations can be composed into an inspectable mediation process: one that records which source it used, why it chose a representation, what it was obliged to preserve, and when it should leave the source alone. That is a design hypothesis, not a claim to have invented mediation.

The step this book takes is to treat mediation as an explicit engineering layer between a source and a recipient. Given a source, a purpose and permitted context, that layer might select information, change its form, attach evidence, or decline to transform it. Summarising is one instance, and not a trivial one. The condition a fluent summary removes may be the very condition that changes the decision. Later, a system might handle bounded exchanges on someone’s behalf. But permission to transform a representation is not permission to speak or act for its owner. Each new capability needs its own preservation and authority conditions.

Note carefully what this claim is and is not. It does not say that AI “understands” in the human sense, that generated text speaks for a person, or that algorithms began transforming messages with generative AI. It says that software can now propose varied, task-conditioned representations: prose turned into a diagram, an hour of video turned into checkable claims, forty papers narrowed to the few relevant to a question. These may be worth investigating for their flexibility. Every proposal may be wrong, and a plausible rendering does not establish that the right information survived. Whether the proposed layer improves reconstruction over an unchanged source is the question to measure. It is not an outcome this chapter assumes.

An inspectable channel could propose a reading before transformation, then examine what its delivered representation communicates. Neither of these is the sender’s intended meaning, and it helps to keep three things apart. Intended meaning is what the sender sought to convey. Expressed meaning is what the communication makes available in context. Proposed interpreted meaning is the system’s fallible account of the other two. The delivered representation gives the recipient evidence from which to reconstruct. Answers, explanations, decisions and behaviour offer observable evidence about that reconstruction, without exposing a mind. The Understanding Compiler develops an instrument for keeping these readings open to challenge.

The case worked through

An example fixes the idea and shows what the audit is for. A colleague sends you a twenty-page incident report on a database outage: timeline, logs, five candidate causes, and a disputed conclusion. You could read it, search it, or ask someone else to interpret it. A mediating system proposes another route: derive a view for your current purpose, which is deciding whether to approve the fix. It might produce a causal diagram of the five hypotheses and their evidence, a comparison of predictions against logs, or a two-paragraph brief that links each consequential claim to its passage. Each is a candidate interface to the report. None is a substitute for consulting the source.

None of these is the report. Each is a derived view, useful only if its interpretation remains faithful and its form helps with the task. Now suppose the report says that hypothesis three can be excluded only if a replication-lag measurement falls below a specified bound, and also says that this measurement is missing. A clean diagram that labels hypothesis three “excluded” has inverted the practical conclusion by deleting a condition. Follow the deletion through. The report’s own conclusion is conditional: nothing about hypothesis three is settled until the measurement arrives. The diagram’s conclusion is unconditional: hypothesis three is out. A reader acting on the diagram might approve a fix that the report’s author, and the missing measurement, would have warned against.

The source passage, the conditional relation, the missing evidence and the disputed status all had to survive. A link back to the report makes the failure inspectable. It does not, by itself, prevent the failure. That distinction matters, because “it cites its sources” is often offered as if it were a safety property. It is not. A link lets a careful reader catch the error after the fact. It does not stop the error from being made fluently, and it does not restore the lost condition for the reader who never follows the link.

The system must therefore keep three things distinct: the source, its own interpretation and representation, and what the recipient actually notices, concludes and does. Each stage needs its own check, because each can succeed or fail on its own. A view without provenance is a rumour with good typography. A view with provenance can still be wrong. Provenance is valuable because it keeps the failure inspectable. It preserves a way back from representation to source, which is a precondition for correction, though not a guarantee of it.

Loss, and gain

Return to the ward and suppose, as a constructed example, that a system reads the overnight notes and observations to prepare a draft handover. It might surface something the registrar overlooked, such as a result flagged at 03:10 or a downward trend in the observations. That would be a gain. It might also turn her hesitant “keep an eye on him” into the tidy word “stable”, deleting the doubt that mattered. That would be a dangerous loss. The two cases look alike in form: both are well-formed, source-linked sentences produced by the same machinery. In consequence they could hardly differ more. Neither output becomes clinically trustworthy because it is well written or source-linked. The relevant records, uncertainty and proposed actions would still need professional review. The same transformation machinery can assist or mislead. What separates the outcomes is what survives inspection and whether the team can correctly act on it.

So every proposed transformation in this book faces a two-sided audit. The preservation check asks whether the required claims, relations, conditions, uncertainty and provenance survived without unsupported additions. The usefulness check asks whether a receiver, working toward a declared goal, recovers something or does something better than they could with the unchanged source. The two checks are kept separate because their failure modes are separate. A source-linked output can pass the first and fail the second: faithful, and useless for the task. A preferred output can fail the first: helpful-looking, and quietly wrong. When the system cannot justify a transformation, leaving the source alone is a legitimate result. It is not a fallback to be apologised for.

SOURCE → TRANSFORMATION → REPRESENTATION → RECONSTRUCTION → SURVIVED / FAILED

Once named, the failures are familiar: a missing fact, a lost qualification, a reversed relation, an unsupported addition, confidence that moved, a trail back to the source that vanished, an intent that changed. Later chapters add failures that appear when a system starts to adapt to a person or act for one: personalisation that goes too far, inference nobody authorised, action nobody permitted. The gains are just as real: structure made visible, context supplied, effort saved, and access widened to someone who could not otherwise use the material. A book that counted only losses would conclude that the safest channel is the dullest one. A book that counted only gains would be selling something. The aim is to keep both in view and to ask the same question of each: what had to survive?

Four questions, in order

If the channel can perform context-sensitive mediation, four questions open in sequence. They organise the four Parts of the book. Each is a progressively harder engineering question, not a capability already shown to work.

Representation. What form should carry this meaning? A page need not always remain prose. A diagram, table, timeline or source-linked explanation may suit a task better, and may also conceal the decisive qualification, as the incident report shows. The first Part develops the proposed Semantic Browser, which would compare derived views with declared preservation obligations and keep an unchanged-source fallback. Whether it beats the ordinary page remains to be tested.

Attention. What deserves to reach me at all? Subscriptions offer forty hours of conference video while papers, podcasts and threads pile up faster than they can be read. Herbert Simon’s point that a wealth of information creates a poverty of attention now applies at personal scale. The second Part develops a Personal Information Radar, which decides which information units to surface, defer or leave alone, and records the basis for those choices. Finding something relevant does not yet show that interrupting a person with it was useful.

The person. How should communication reach me now? The same material serves different needs at a desk and on a train, in five minutes and in fifty. The third Part asks which declared preferences and temporary conditions a mediator may use, how inferred beliefs stay distinguishable from facts, and how a person can inspect or correct them. Adaptation is not permission to build a permanent profile. Sometimes the right output is less, or nothing at all.

Delegation. What happens when humans stop writing and reading every intermediate message? A system might predict a bounded choice, propose a reply or coordinate a meeting. Systems could exchange representations without a person inspecting each step. But prediction is not identity, and an accurate guess is not authority. The fourth Part examines explicit grants, the scope of permitted speech and action, independent gates, revocation, and what recipients are entitled to know. A system able to communicate on someone’s behalf is not thereby authorised to do so.

Each question exposes a limit of the one before it, and that sequence is the argument, not an accident of ordering. Representation without attention drowns the reader in beautiful views of everything. Attention needs context to distinguish the relevant from the merely new, or the urgent from the merely interesting. Personal mediation then raises a further question: which exchanges, if any, does the person want to hand over? Delegation without the earlier layers would manufacture unaccountable action. Without provenance, without a way to check what survived, and without policies a person can inspect, delegation has nothing to stand on. Handing over more is not the measure of success. Surviving more is.

How the book speaks

A short note on method, so the reader knows what kind of statement they are reading. Three kinds appear.

Established background is supported by prior research, and the chapter notes its relevant populations and limits. Argument is what this book proposes: a distinction or mechanism that the research motivates but does not settle. Proposal is a test or design that has not yet been run.

The ward handover and the incident report are constructed cases. This book reports no book-specific experiments with readers or delegated systems. A limited implementation uses a scripted stand-in for a model. Such checks can establish what a program accepts or refuses under supplied inputs. They cannot establish whether a real model will preserve meaning, or whether mediation will improve a person’s decisions.

One more orientation. The book proposes a single experimental layer between people and AI systems as they communicate, and it shares its name with the book: Language. The Semantic Browser, Personal Information Radar, personal mediator and proxy are successive views and tests of that layer. They are not four established products. A small implementation may ultimately be enough, and some proposed stages may prove unnecessary. Whether the layer makes communication measurably better, at an acceptable cost and within legitimate authority, is the question the book ends on. It is not a result assumed at the start.

Preference goes to peer-reviewed work, proceedings, datasets and official documentation. Preprints are labelled, and vendor claims are attributed rather than adopted as findings. Correlation is not causation, preference is not performance, and a scripted test of a component is not evidence that a person benefited from it. Where research disagrees, the disagreement stays visible, as it does in the Eisenstein–Johns exchange above.

The word meaning is used provisionally in these opening chapters, for the distinctions a sender seeks to communicate and a receiver reconstructs. Neither can be read directly from a person’s mind. A test must therefore treat declared intentions, identifiable source obligations and observable receiver outcomes as imperfect operational evidence. It must not pretend to measure complete inner meaning.

Where this leaves us

Language is an interface through which people try to make meaning recoverable by other minds. Earlier media changed the conditions of that work, and human and computational intermediaries already reshape messages. What is new is not mediation itself. It is the prospect of composing selection, transformation, adaptation and bounded delegation into one inspectable layer.

The central question remains what had to survive? Preservation, however, is only half the test. The other half is whether the recipient can use what survived, with less error or effort than before, and without granting the mediator authority it never received.

Before asking what a channel with intelligence can do, then, it is worth asking how much the receiver was always doing. If meaning is rebuilt and not delivered, then what a message means depends on the person receiving it as much as on the words. What, exactly, does a receiver rebuild?

References

• Ong, W.J. (1982). Orality and Literacy: The Technologizing of the Word. London: Methuen. (Routledge reprints, including the 30th anniversary edition, 2012.)
• Eisenstein, E.L. (1979–1980). The Printing Press as an Agent of Change: Communications and Cultural Transformations in Early-Modern Europe. Cambridge: Cambridge University Press. (Paperback edition 1982.)
• Johns, A. (1998). The Nature of the Book: Print and Knowledge in the Making. Chicago: University of Chicago Press.
• Eisenstein, E.L. (2002). “An Unacknowledged Revolution Revisited”. American Historical Review, 107(1), 87–105.
• Hancock, J.T., Naaman, M. & Levy, K. (2020). AI-Mediated Communication: Definition, Research Agenda, and Ethical Considerations. Journal of Computer-Mediated Communication, 25(1), 89–100. DOI 10.1093/jcmc/zmz022.
• Simon, H.A. (1971). Designing Organizations for an Information-Rich World. In M. Greenberger (ed.), Computers, Communications, and the Public Interest. Baltimore: Johns Hopkins Press, pp. 37–72.