One sentence, three small operations

At a press conference, an institute spokesperson says: ‘The programme may launch in May, provided that the safety tests conclude without further concerns.’ This is a hypothetical sentence for our experiment. It contains a date, a condition and a moderate degree of certainty.

The first user asks a chatbot for a summary and receives: ‘The programme will launch in May.’ The second asks for plain language: ‘If the tests go well, the programme may launch in May.’ The third requests an optimistic version for residents: ‘Good news: the programme is likely to launch in May.’

The second transformation preserves the essential meaning while widening access. The first removes the condition and turns a possibility into a fact. The third retains uncertainty but adds a judgement the spokesperson did not make. Three operations took seconds. The original speaker saw none of them. A newsroom published one sentence; its audience may encounter three different messages.

This is meaning drift during conversational rewriting. Summarisation, simplification, translation, tonal change and personalisation do not mechanically move identical bricks. They can rearrange the relationship between fact and possibility, remove a caveat, change the agent or polish

away a conflict. The more natural the answer sounds, the less visible the transformation becomes.

Not every summary betrays its source. Without summaries, translation and plain language, much public communication would remain inaccessible. The problem begins when the adaptation impersonates the original and the recipient cannot see what changed. Imagine an interpreter who sometimes translates, sometimes summarises and sometimes improves the speaker — all in the same voice, without ever switching on a warning light.

Hall: the audience was never an empty vessel

Stuart Hall’s encoding/decoding model predates chatbots by decades, yet it offers an unusually current instrument. Hall rejected the simple picture of communication as a parcel packed by a sender and unpacked unchanged by a receiver. A message is produced within frameworks of knowledge, social relations and technical conditions. Reception also uses cultural codes and experiences that need not be identical to those of the producer (Hall, 1980, pp. 128–131).

Hall described three hypothetical positions of decoding. In the dominant- hegemonic position, the audience accepts the message’s preferred meaning. In the negotiated position, a person recognises the general frame but adapts it to local circumstances. In the oppositional position, she understands the proposed code yet deliberately reads against it (Hall, 1980, pp. 136–138). These are not three species of person or buttons in the brain. The same reader may occupy different positions towards different parts of one story.

The crucial intuition is that meaning does not sit in a text like a coin in an envelope. It emerges in the relationship between the message’s structure and the recipient’s practice. Robert T. Craig describes the broader constitutive view as ‘communication as a constitutive process that produces and reproduces shared meaning’ (1999, p. 125). Communication does not merely transport a finished substance; it helps make the shared world in which a substance becomes intelligible.

There is an important qualification. Hall did not treat a message as infinitely malleable or interpretation as whim. A text has a structure, a preferred reading

and material conditions of production. An audience can resist precisely because there is something towards which it can take a position. An oppositional reading of a novel does not require the printing press to alter the paragraphs for each purchaser.

When each reader receives a private ‘original’

The conversational interface adds a twist to Hall’s model. The audience no longer only interprets a relatively stable object. A user can ask the system to shorten, translate or illustrate it; answer a counterargument; or adjust the tone to her age and interests. The publication is encoded again at the moment of reception.

We need not pretend that a model has an author’s intentions to call this re- encoding. The term describes a function: the system produces a new sequence of signs from a source, a user instruction and its learned patterns. Research in Human–Machine Communication reminds us that advanced technologies are no longer merely channels between people. They occupy communicative roles as senders and receivers, while users respond to them as participants (Guzman & Lewis, 2020, pp. 72–79; Skrzypiec, 2025, pp. 17– 22, 32–36).

This changes the unit of publication. A newsroom could once point to an article, edition and correction. Today one source text may feed thousands of on-demand answers. Each is generated in a different context and may leave no archive. The writer does not know which version the reader met; the reader may not know that another person received a substantially different one.

Atkinson and colleagues use reception theory to ask whether a person retains sufficient interpretative freedom when generative AI becomes part of the analysis itself (2025, pp. 506–508). The paradox is delicate. A chatbot increases audience activity by permitting questions and formal choices. It may simultaneously reduce autonomy if it first removes everything against which the reader might argue.

Five currents that carry meaning away

Meaning drift is not one error. In editorial practice, at least five mechanisms should be kept separate.

Compression removes elements judged less important. A good summary reduces length while preserving the principal relationships. A bad one removes an exception, the denominator of a statistic or a condition. A rise within one age group becomes a rise across the whole population. The sentence is shorter, but its logical reach has mysteriously expanded.

A shift in modality changes the degree of certainty. May becomes will, indicates becomes proves, and has not been ruled out becomes has been confirmed. These are small verbal parts with the power of a focusing ring. In legal, scientific and political news, a single turn can transform a hypothesis into a verdict.

A shift in voice or attribution obscures who is speaking. An expert’s opinion becomes the assistant’s narration, one party’s position becomes an impersonal fact, or a quotation becomes an elegant paraphrase kept inside quotation marks. The BBC/EBU toolkit stresses that full and partial quotations must accurately reproduce the source’s words (Yezza et al., 2025, p. 6). Improved words may be clearer, but they are no longer a quotation.

Conflict smoothing manufactures consensus. Where sources disagree, a model may join them with a connective that implies compatibility, or select a common but trivial denominator. The reader receives a calm paragraph although the newsworthy fact was precisely that the parties disagreed about the data, the definition or who bears responsibility.

Perspective personalisation chooses what should matter to one person. It may help a student, patient or resident understand the consequences of a complex decision. It may also quietly delete costs borne by others. Explaining a reform solely from the standpoint of a driver, investor or parent is not a fuller truth. It is a useful frame that must acknowledge its boundaries.

All five mechanisms existed before generative AI. Editors condensed, translators interpreted and websites personalised. What changes is their scale, speed and invisibility. A separate transformation can occur for each

recipient, without the author’s approval, an editor’s sign-off or a retrievable version history.

The speaker’s dignity and the reader’s autonomy

The first ethical stake is the dignity of the person whose words are transformed. An interviewee does not lose the right to the meaning of her statement because a system can make it smoother. The point becomes acute with testimony about harm, minority speech, scientific disputes and legal material. Removing hesitation can make a witness more certain than she was. Correcting grammar may improve legibility while erasing origin, age or emotion that formed part of the testimony.

The second stake is the audience’s autonomy. Interpretative freedom does not mean that every reading is equally sound. It means being able to encounter the material, recognise its frame and take a position towards it. If personalisation adjusts meaning to our beliefs before we see it, negotiation becomes a confirmation service. The oppositional reading disappears not through censorship but because the user never meets the text to which she might object.

Fidelity must not become a cult of difficulty. Plain language, audio description, translation and summarisation can enlarge the freedom of people previously excluded. A human-centred system does not force everybody to read a statute in a legislator’s prose. The ethical aim is accessibility joined to verifiability: the adaptation should open a door to the source rather than brick it up behind the visitor.

The Four Transformation Traces

DBMoJo’s practical proposal is the Four Transformation Traces. Use it whenever a publication is automatically shortened, translated, simplified, personalised or turned into dialogue. The newsroom or conversational-service provider owns the process; for high-stakes material, a named human editor approves the traces.

1. Source trace. Every variant leads to a stable primary object: the document, recording or article, with its author, date and version. For a key quotation, the link should open the relevant passage rather than the publisher’s homepage. Pass condition: one action lets the reader see what the answer was made from.

2. Change trace. The system names its operation: summary, translation, plain language, tonal adaptation, synthesis of sources or answer based on an extract. For matters of public importance, it highlights alterations that affect meaning, particularly omitted conditions, figures and caveats. Pass condition: the user knows not only that the material changed but what kind of change occurred.

3. Certainty trace. Statements are distinguished as source fact, attributed claim, opinion, system inference or absence of evidence. Labels need not make the page resemble a Christmas tree; details can expand on demand. Uncertainty that matters to a decision, however, must be visible immediately. Pass condition: a possibility is not displayed as a certain outcome.

4. Alternative trace. The answer retains at least one material caveat, source disagreement or reading that personalisation might otherwise smooth away. This is not a ritual weak counterargument beside every date. An alternative is required when it changes the judgement or shows who bears the cost. Pass condition: a user-specific version does not erase a publicly significant standpoint.

The overall test is simple: with one action, a person can compare the generated version with the source and see the differences that matter. The accountability trail retains the transformation, time, model version and team accepting corrections. Nobody needs the complete engineering log. Everybody needs evidence that adaptation has not become anonymous editing.

A publication with a memory of origin

Imagine a report about a city’s heat-warning system. The source article contains an interview with an official, a neighbourhood map, five years of data

and criticism from a charity working with older residents. A reader asks: tell me in five points what I need to do.

A good conversational version provides practical steps and retains all four traces. Source links lead to the article and the official guidance. Change identifies the output as a task-focused summary and says that most analysis was omitted. Certainty distinguishes current procedures from a planned app feature. Alternative preserves the warning that the scheme may not reach people without smartphones, even though this fact is not necessary for the user to enable notifications.

The last element protects public communication. Personalisation may say, ‘Here is what matters for your immediate purpose.’ It must not silently turn that into, ‘Only this matters.’ The reader keeps the convenience and can still see a social cost that would otherwise disappear from her private version of the world.

The model has limits. A textual comparison will not reveal every cultural shift in translation. Stable links can decay, sources can be wrong and witness safety can prohibit access to the full material. Too many labels may repel the very people an adaptation was intended to include. Traces should therefore be layered, while safety exceptions must be recorded and approved by a human.

Meaning cannot be frozen. Hall never asked for that, and responsible AI should not either. Interpretation will remain alive, negotiated and contested. What can be preserved are the conditions of an honest contest: a common point of reference, knowledge of the alteration, visible uncertainty and the practical right to an alternative reading.

A text may answer, but it should remember

Conversational publication offers a remarkable advantage. A reader can ask a follow-up question, translate difficult prose, request another format or locate what matters in her circumstances. This can make media substantially more accessible. The same mechanism can create a world where everybody receives a courteous version of meaning and nobody knows what the author said.

Hall taught us that the audience is not the end of a transmission pipe. The chatbot adds that the pipe can rewrite the water on its journey — then assure us that it always tasted that way. The new audience right is not a right to one neutral interpretation. Neutrality of that kind is a promise no system can honour. It is the right to see the transformation.

The speaker’s dignity requires that an adaptation not invent certainty, judgement or intention. The reader’s autonomy requires access to material against which she can negotiate or oppose a meaning. Editorial responsibility requires memory: who transformed the text, what changed, and how somebody can return to the source.

A text may answer everyone differently. It should not suffer from amnesia. With four visible traces, conversation need not replace interpretation. It can make interpretation more conscious.

References

  1. Hall, S. (1980). Encoding/decoding. In S. Hall, D. Hobson, A. Lowe, & P. Willis (Eds.), Culture, media, language: Working papers in cultural studies, 1972–79 (pp. 128–138). Hutchinson.
  2. Guzman, A. L., & Lewis, S. C. (2020). Artificial intelligence and communication: A Human– Machine Communication research agenda. New Media & Society, 22(1), 70–86. https://doi.org/10.1177/1461444819858691
  3. Yezza, H., Fletcher, J., & Verckist, D. (2025). News integrity in AI assistants: Toolkit. BBC & European Broadcasting Union.