AI Is Not a Hammer. It Is a Room That Rearranges Our Thinking: McLuhan After ChatGPT

The writer who knows the answer but cannot remember the route

Picture a reporter at a local news site. A council budget lands on her desk: dozens of pages, several spreadsheets and two hours until publication. She gives the documents to an AI system and asks for the main findings, then a lead, three counterarguments and questions for the council’s press officer. The reply is quick, coherent and polished. The story goes live. The next morning, a councillor asks why the article has described planned investment as money already spent. The reporter opens her chat, scrolls upwards and discovers something unsettling. She can find the sentence she published, but she cannot reconstruct the reasoning that produced it.

This is not a parable about a spectacular hallucination. The model may have confused two columns; the reporter may have missed a caveat. The deeper problem is the erosion of independent judgement through invisible delegation. A human still pressed Publish and still has a byline. Yet she cannot say which premises were selected, what was left out or what evidence would have changed the conclusion.

On a deployment chart, everything looks exemplary: AI assisted the research, a person remained ‘in the loop’, and an editor approved the copy. But a human in the loop can resemble a driver holding a steering wheel

disconnected from the wheels. Presence is not agency. Agency also requires knowledge, time, access to evidence and a genuine right to object.

That is why the first newsroom question about AI should not be how many minutes it saves. It should be: what kind of work has disappeared from view? Did we delegate transcription, or the decision about what matters in the transcript? Did we outsource the sorting of data, or the standard of truth? A dishwasher and a judge both save somebody time. Only one should be trusted with a verdict merely because it arranges the plates beautifully.

A hammer sits on the table; a medium changes the room

The tool metaphor is reassuring. A hammer can build a house or break a window; moral responsibility belongs to the hand that swings it. In this story technology waits patiently for instructions, changing neither the purpose nor the person. Generative AI sometimes is such a tool. It may, for example, organise a transcript according to rules a newsroom has already established.

The metaphor begins to fail when the system also proposes the questions, supplies the categories and offers a ready-made shape for the answer. A hammer does not praise your plan, suggest what counts as a wall and produce a plausible architect’s certificate when asked. A conversational model does something closer to that: it participates in how the problem is framed before it helps to solve it.

Medium theory, associated above all with Marshall McLuhan and developed by scholars such as Joshua Meyrowitz, asks us to look beyond individual uses towards changes in the whole communicative environment. A medium rearranges pace, access, social roles and the very definition of competence. Writing altered memory; print altered authority; television altered political visibility; search engines altered the habit of finding knowledge. This happened not because everybody used each medium identically, but because older practices began operating under new conditions (McLuhan, 1964; Meyrowitz, 2008).

Neil Postman compressed this into a sentence that reads like an evacuation notice: ‘Technological change is neither additive nor subtractive. It is ecological’ (1992, PDF p. 19). Introduce a new medium and you do not get the old newsroom plus a chatbot. You get a different newsroom: one with a new

expectation of speed, a different economy of first drafts, an altered status for research and a temptation to mistake verbal fluency for mature judgement.

The AI room has distinctive furniture. In the centre sits a prompt box, implying that a problem can be enclosed within an instruction. Beside it is a Regenerate button, teaching us that an answer can be replaced without looking at the world again. Sources are hidden in a drawer, or absent. A clock hangs on the wall: faster almost always looks better. Nobody needs to order the user about. The room itself suggests how to move.

The lift to the floor marked ‘answer’

Atkinson and colleagues argue that media education should assess generative AI as a medium, not merely as an auxiliary device. They illustrate the difference with a difficult book. Reading forces a person through the sequence of premises, examples, footnotes and method. A model-generated summary takes the reader to the conclusion more quickly, but may detach the result from the process that made it intelligible. The authors name the danger precisely: ‘knowledge is separated from the process of reasoning’ (Atkinson et al., 2025, p. 500).

AI is the lift. A lift is not wicked. For someone with limited mobility it opens a building that the stairs had closed. Translation, captioning, document navigation and explanations of technical terms can genuinely widen access. But a person training to be a firefighter cannot learn the building solely by riding the lift. When the power fails, she must know where the corridors lead.

In journalism, the stairs include formulating a question independently, distinguishing a primary source from commentary, noticing a contradiction, checking a date, comparing explanations and attempting to disprove one’s favourite thesis. These are not bureaucratic packaging around finished knowledge. They produce competence. A reporter learns to judge a document while stumbling over an obscure footnote, ringing the person who made the table and discovering that two similar percentages have different denominators.

Atkinson and colleagues are not arguing that every reduction in labour is harmful. Their question is whether a shortcut extends the learner’s knowledge or replaces the effort through which that knowledge would have developed (2025, pp. 500–503). The same distinction matters in a newsroom. An experienced reporter may ask a model to test alternative leads because she

already understands the evidence. A beginner may accept the same lead as a map of the problem and never notice that the map omits an entire district.

From delegating a step to delegating the measure

Automation usually enters through a side door. First the system removes repetitions. Then it summarises an interview. Next it proposes the central theme, ranks the arguments and constructs the ending. Each move seems sensible because the previous move has become normal. At some point, however, we stop delegating another step and begin delegating the measure of a good answer.

Four symptoms reveal the shift. First, the writer checks only whether the prose sounds plausible. Secondly, she checks a link but not whether its contents actually support the relevant claim. Thirdly, she commissions the counterargument from the same model and treats its production as proof of pluralism. Fourthly, she cannot name a single suggestion she rejected. Collaboration without any trace of disagreement may indicate miraculous alignment. More often it resembles a meeting where the intern technically has a vote, but the boss has already ordered champagne for his preferred decision.

Research on automated news shows that systems do not eliminate editorial decisions. They relocate them into choices about data, categories, rules and thresholds (Diakopoulos, 2019). A model may count more quickly, but somebody determines what deserves counting. It can identify a pattern, but it cannot independently establish that pattern’s public significance. When those choices become invisible, responsibility dissolves among the supplier, the prompt, the interface and the person whose name appears on the story.

The greatest risks therefore arise in tasks with high stakes and weak verifiability: interpreting a source’s intentions, explaining a conflict, diagnosing a cause, assigning blame or choosing between several reasonable accounts. The harder a result is to compare with a simple answer key, the less useful a generic instruction to ‘think critically’ becomes. Critical judgement is not a red button installed inside a human being. It is a disciplinary practice, exercised against particular evidence (Atkinson et al., 2025, pp. 505–508).

Freedom of judgement is not a tick box

The ethical stake is larger than the originality of a paragraph. It is freedom of judgement: a person’s capacity to understand the route to a conclusion, challenge it and accept responsibility for the choice. Writers retain dignity not when their names decorate automated output, but when they remain subjects of the decision.

The audience has a corresponding interest. A reader should be able to ask why a newsroom considered one fact decisive. If the only answer is ‘the system selected it’, technology has become an alibi. The model will not attend the editorial meeting, apologise to the person harmed by an error or explain a correction. Responsibility remains human; therefore the capacity to trace and refuse the reasoning must remain human too.

Freedom does not mean doing everything by hand. A pilot does not lose agency because a computer maintains altitude. She loses it when the airline no longer trains her to recognise the conditions in which automation fails, while its schedule denies her the time to intervene. Likewise, a newsroom can automate transcription, formatting and comparisons across large datasets while protecting the skills needed to judge meaning, risk and truth.

A human-centred approach consequently demands more than final approval. The person needs access to evidence, suitable knowledge, a clear field of responsibility and the right to stop the process without being punished for slowing production. Otherwise ‘human in the loop’ is decorative language: a plastic plant in a room that ceased to have windows years ago.

The 5P Audit for an AI environment

DBMoJo’s practical response is the 5P Audit: Pace, Passage, Provenance, Pushback and Personal Sign-off. The writer performs it when planning the assignment and again before publication; for high-risk work, an editor co-signs it. The audit does not ask vaguely whether AI was used. It examines what AI’s presence did to the process of knowing.

1. Pace. Name the activity that the system accelerated. Transcription, format conversion and duplicate detection usually reduce mechanical labour. Selecting the thesis, assessing a witness and interpreting evidence involve judgement. Test: the writer can explain why the

accelerated stage was not where the crucial competence needed to be acquired.

2. Passage. Record the step that disappeared inside the interface. Did the model choose extracts from a document, turn correlation into cause or remove an exception while summarising? Test: every material conclusion has a short map of its premises, including meaningful alternatives that were rejected.

3. Provenance. Travel from sentence to primary evidence. A link to an article repeating somebody else’s result is not enough. Check the author, date, scope of the data and whether the cited passage supports this precise claim. Test: another editor can repeat the journey without access to a private chatbot conversation.

4. Pushback. Identify who can genuinely challenge the output. This may be another reporter, a specialist, a source with a different perspective or a deliberately designed falsification test. Asking the same model again is useful assistance, not independent scrutiny. Test: before settling the thesis, record one piece of evidence that would change it.

5. Personal sign-off. Identify the recognisably human contribution: original reporting, an ethical choice, interpretation, field observation or a reasoned rejection of a suggestion. A byline is not a prize for clicking. It is a promise. Test: without the model, the writer can summarise the argument, locate its evidence and name at least one AI proposal that was declined.

The audit trail can be a compact card attached to the story in the content system: AI task, sources, risks, checker, rejected suggestion and final decision. There is no need to publish every prompt. There is a need to preserve the route of responsibility.

One budget story, tested — and the limits of the test

Return to the council budget. Pace: AI compared the tables and listed changes. Passage: it did not reveal that one column described a plan while the other recorded actual expenditure. Provenance: the reporter returned to the resolution and its methodological notes. Pushback: she checked her interpretation with an independent local-government economist. Personal sign-off: she rejected the dramatic claim that spending had been cut because

the figures did not support it, replacing it with the more accurate finding that the timetable had shifted.

This workflow is slower than copying an answer and faster than checking every cell manually. That is the point. We need not preserve old difficulty for difficulty’s sake. We should preserve resistance where judgement is formed. AI can be the lift if the reporter still knows the plan of the building.

The 5P Audit has limits. It cannot reveal unknown training data, repair a defective source or substitute for specialist knowledge. It may itself become a ritual completed after the fact. Its organisational preconditions are therefore a right to halt publication and time for verification. Without both, even an excellent checklist becomes a seat belt made of paper.

Nor does every use of AI weaken judgement. A well-designed system can expose contradictions, make documents accessible, suggest questions and assist somebody working across languages. Risk rises when the output is fluent, the task lacks an easy answer key, the user is still acquiring expertise and the production schedule rewards acceptance more than objection. That is not an argument against the technology. It is an argument for arranging the room so that a person can still reach the door unaided.

The McLuhanesque question, then, is not only what AI does for the communicator. It is what kind of communicator AI makes. A serious answer will not be a declaration of faith or fear. It will leave evidence in the craft: a verifiable source, a visible moment of dissent and a writer who remembers not only the floor marked ‘answer’, but the stairs.