Seven Theories, One Dispute: Craig’s Map for Teams Implementing AI

The meeting begins well. Everyone agrees that the proposed AI news summariser must be ‘accurate’, ‘transparent’ and ‘human-centred’. The editor nods. The engineer nods. The audience researcher, lawyer and inclusion lead nod too. The minutes record consensus.

Forty minutes later, they are arguing.

For the engineer, accuracy means that the summary preserves the central facts. For the editor, it also means preserving uncertainty and news judgement. For the audience researcher, an accurate answer that nobody understands has failed. For the inclusion lead, a fluent summary that systematically flattens minority voices is not neutral. The lawyer wants a traceable duty. The product manager wants something that can ship on Tuesday.

They are not merely disagreeing about the answer. They are using different definitions of the problem. Robert T. Craig’s map of communication theory can help — not by appointing a winner, but by showing which question each person has brought into the room.

Theory is a toolkit for naming trouble

Communication theory is often taught as a museum: one display case for rhetoric, another for semiotics, a slightly humming cabinet for cybernetics. Students learn the labels, pass an exam and leave the exhibits safely behind.

Craig proposed something more practical. The field could be understood as a conversation among traditions that define communication, its failures and its remedies differently (Craig, 1999, pp. 119–161). A theory is useful not only because it describes a phenomenon, but because it gives participants a vocabulary for discussing problems they already experience.

Emanuel Kulczycki develops this practical reading of Craig in detail. His formulation is wonderfully direct: “The process of theorising is a practical response to experienced communication problems” (author’s translation, 2012, p. 148). Theory, then, is not parsley sprinkled over a decision once the serious work is complete. It helps reveal what the serious work is.

AI is never one object. It is code, data, interface, institution, labour arrangement and encounter. Seven lenses do not guarantee wisdom, but they make premature agreement harder.

One summariser, seven definitions of success

Take a hypothetical but ordinary proposal. A newsroom wants an AI assistant to turn full reports into short mobile summaries. Readers can ask follow-up questions. The system is described as a way to improve access and reach younger audiences.

Nothing about the idea is inherently responsible or irresponsible. The consequences depend on design, editorial practice and power. Craig’s seven traditions make those dependencies visible.

1. The rhetorical lens: communication as the practical art of discourse.

Rhetoric asks how a message addresses an audience, establishes credibility and moves people towards judgement or action. For the summariser, the question is not only whether sentences are factually correct. Who is speaking? What kind of authority does the synthetic voice claim? Does the

opening frame uncertainty as settled? Does a call to ‘learn more’ genuinely invite inquiry, or merely push the newsroom’s preferred path?

The rhetorical failure is inept or manipulative discourse. The remedy is not to remove persuasion — impossible in any act of selection — but to make purpose, audience and available counterarguments explicit.

2. The semiotic lens: communication as mediation through signs.

Semiotics examines how signs acquire meaning and how codes differ. A model may replace ‘undocumented migrant’ with ‘illegal migrant’ or compress ‘alleged’ out of a sentence. Each change looks small at token level. Culturally, it can move a person from one moral category to another.

This lens asks the team to compare source and summary for connotation, metaphor, labels, visual symbols and cultural codes. The problem is not noise in transmission but a mismatch or transformation of meaning. ‘Same topic’ is not the same as ‘same signification’.

3. The phenomenological lens: communication as encountering another.

The phenomenological tradition brings attention to lived experience, dialogue and the difficulty of genuinely meeting someone who is not oneself. A summariser can efficiently process testimony while making the person who testified feel erased. A reader can receive a correct answer yet have no way to ask, ‘But what did this mean for the family involved?’

The relevant failure is estrangement: people become objects to be processed rather than subjects who can answer back. The remedy involves presence, listening, context and a route for correction. When the system summarises a first-person account, the person’s dignity may require more than semantic similarity; it may require preserving the terms in which they chose to be known.

4. The cybernetic lens: communication as information processing and feedback.

Cybernetics sees systems, flows, control, noise and feedback. Here the engineer’s questions become central: Where does the source enter? What information is lost at each stage? Which signal triggers escalation? How is a correction fed back into future outputs? What happens when the system receives contradictory sources or when user behaviour becomes the optimisation signal?

The cybernetic problem is a broken or badly regulated loop. A summariser may score well in a static test but fail in operation because corrections do not propagate, feedback rewards clicks rather than understanding, or nobody notices drift. Iwona Hofman’s account of the tradition highlights noise, information overload and mismatch between structure and function as characteristic communication problems (2025, pp. 25–26).

5. The sociopsychological lens: communication as expression, interaction and influence.

This tradition asks what messages do to individuals: attention, attitudes, trust, learning, memory and behaviour. Does the short format improve comprehension? Do confident synthetic answers create over-reliance? Does disclosure of AI involvement change trust? Which users notice an error, and who carries it into a decision?

The failure is not necessarily bad intent; it can be an unmeasured effect. The remedy is empirical: user testing, comparison groups, behavioural measures and analysis across different audiences. ‘People liked it’ cannot stand in for ‘people understood it’, and neither guarantees that they retained the caveat printed in line six.

6. The sociocultural lens: communication as the production of social order.

Communication does not merely travel through a culture; it helps reproduce culture. Repeated formats teach people what a news item is, which sources sound authoritative and whose grammar counts as professional. If every summary adopts the same institutional register, local speech and contested meanings may be treated as imperfections to be cleaned.

The sociocultural lens therefore asks which norms the system repeats and which practices it changes inside the newsroom. Do reporters begin writing for the summariser? Are sources selected because their documents are machine- readable? Does the ‘short version’ become the real public record while the nuance survives only in an archive few people open? The failure is the reproduction of a narrow order under the label of convenience.

7. The critical lens: communication as reflection on power and domination.

The critical tradition asks who benefits, who defines rationality and whose interests disappear behind the apparently natural operation of a system. Who

owns the model? Whose data trained it? Which jobs are intensified or removed? Can communities challenge a harmful representation? Does personalisation turn public knowledge into a proprietary service?

The problem is not only bias in an output but power over the conditions of speaking and being heard. The remedy can include redistribution, worker voice, public accountability, alternative infrastructure and the right to refuse. A technically accurate summary may still be critically unacceptable if it locks a newsroom into a supplier that can alter access, price or policy without meaningful contest.

Why the lenses collide — and why that is useful

The seven traditions can recommend different actions.

A cybernetic team may want more personalisation because feedback improves relevance. A critical analysis may warn that the same feedback concentrates surveillance and platform power. A rhetorical editor may favour a strong, confident lead; a phenomenological colleague may want the ambiguity of testimony preserved. A sociopsychological test may show that simple labels aid recall; a semiotic analysis may show that those labels carry a stigmatising history.

This collision is the value of the exercise. Craig’s map makes visible that teams are balancing goods that cannot always be converted into the same unit.

Human–Machine Communication adds a further complication. AI can be perceived as a tool, a mediator or a communicator. Functional, relational and metaphysical questions therefore overlap: what the system does, what relationship people form with it and what kind of actor they believe it to be (Skrzypiec, 2025, pp. 35–38; Guzman & Lewis, 2020, pp. 70–83). A friendly synthetic voice is an interface choice, a rhetorical performance, a social cue and a claim about agency at once.

The person who pays for a theoretical blind spot

Abstract disagreement becomes concrete at the edge of the system.

Suppose — again hypothetically — that the summariser condenses a council report about closing a disability service. The facts are present, and the

readability score is excellent. Yet the summary drops the testimony of service users because it appears repetitive. The affected reader loses the only part explaining what closure means for daily independence. The cybernetic pipeline worked; the phenomenological and critical tests failed.

Or imagine a correction submitted in a minority language. Customer support accepts it, but the feedback system cannot link it to the English-language source. The newsroom announces that a human review process exists. For the complainant, it does not.

A human-centred project must identify not an abstract ‘user’ but the person likely to bear each failure: the source whose words are reframed, the reader who cannot access the long version, the reporter blamed for an automated error, the moderator who absorbs harmful material, the community unable to appeal. Dignity means being represented as a subject, not merely a data point. Freedom means having a viable alternative. Responsibility means that somebody with authority must answer and repair.

The Seven-Lens Card: a meeting that produces decisions

The Seven-Lens Card is a structured ninety-minute review for a specific use case. It is not a general declaration that an organisation ‘uses AI responsibly’.

Step 1: Write the decision in one sentence. Not ‘implement a summariser’, but ‘publish AI-generated 120-word summaries of council reports in our mobile app’. Name the affected groups and the human decision-maker. Vague scope produces decorative answers.

Step 2: Bring evidence, not adjectives. The team prepares sample outputs, error cases, workflow diagrams, supplier terms, user research and accounts from affected people. ‘Transparent’ is banned unless followed by ‘to whom, about what, and with what consequence’.

Step 3: Ask one demanding question through each lens.

• Rhetorical: What action or judgement does this message invite, and whose credibility does it borrow?

• Semiotic: Which labels, metaphors or cultural meanings change between source and output?

• Phenomenological: Whose experience is converted into an object, and can that person answer back?

• Cybernetic: What is lost, what feedback returns and where can the loop fail silently?

• Sociopsychological: What effect do we expect on understanding, trust and behaviour, and how will we test it?

• Sociocultural: Which norms, roles and routines will the system reproduce or change?

• Critical: Who gains control, who carries the cost and who can refuse or appeal?

Step 4: Record collisions. Do not force consensus. Write paired tensions: personal relevance versus common agenda; fluent compression versus cultural fidelity; speed versus meaningful review. A collision without a name will reappear later as a ‘surprising’ incident.

Step 5: Convert each material risk into a test and an owner. For example: compare summaries across three language varieties; require retention of uncertainty markers; test comprehension with disabled users; give the editor a stop right; provide sources with a correction route. Add a threshold and a date. ‘Monitor bias’ is not an action.

Step 6: Preserve dissent and appeal. If a lens reveals an unresolved threat to dignity, safety or democratic participation, record who objected and what evidence would reopen the decision. A product deadline does not transform disagreement into consent.

Step 7: Re-run the card after contact with reality. Review at launch, after the first serious correction and whenever the model, supplier or use changes. Communication practices evolve; a one-off assessment becomes historical fiction surprisingly quickly.

The limits of seven windows

Craig’s map is a Western metatheoretical construction, not a universal atlas. Kulczycki notes the problem of incorporating non-Western traditions without simply reducing them to Western categories (2012, pp. 184–186). Teams should therefore add knowledge grounded in the cultures and communities affected, rather than treating seven as a sacred maximum.

The card can also become bureaucratic theatre. Seven shallow answers are not better than one serious investigation. Small, low-risk tools may not justify a ninety-minute review, while high-stakes systems require legal, security, accessibility and domain-specific methods beyond communication theory.

Nor do the lenses make moral choices for us. They expose trade-offs, but people must still decide which harms are unacceptable and who has authority. Power can enter the exercise itself: a junior researcher may ‘own’ a risk without the budget to change it.

The meeting can still end in disagreement. That is healthier than false consensus. The engineer can show the feedback loop; the editor can defend uncertainty; the inclusion lead can identify the voice that compression erases. The team can then decide in full view of the conflict.

Theory has done its job when it changes what becomes sayable before the system changes what becomes possible.

Seven lenses do not give us seven excuses to delay. They give us seven chances to notice the human being hidden by the word ‘implementation’.