‘Human in the Loop’ Does Not Mean a Person Beside a Green Button

It is 5.58 pm. A system has generated fifty local stories from public data. The editor has two minutes before notifications go out automatically. His screen shows a headline, a short summary and a large green button marked ‘approve all’. It does not show the raw table, selection rules or full change history. He knows that delay will damage the newsroom’s speed metric. He clicks.

The following morning, the team discovers that one article merged the records of two people with similar names and wrongly implied a financial default. The project documentation still carries a comforting line: ‘every item is reviewed by a human’. Formally, that is true. Functionally, the human was a seal placed on a box he could not open.

This is a hypothetical scene, but its logic extends well beyond journalism. Human in the loop has become a talisman in AI policies. Add a person to the diagram and the system appears responsible. Yet location in a loop tells us nothing about power. The human may be an expert entitled to stop the process, or an exhausted operator employed to provide the final click.

A loop can become a lead

The simplest picture of oversight is reassuring: the machine proposes; the human decides. A real organisation places many things between proposal and decision — interface design, deadlines, performance targets, hierarchy, training, data access and employment risk. Those conditions determine whether a person is exercising judgement or merely authenticating a result.

An algorithm can direct the operator’s attention towards selected information even when the last move remains technically human. Nicholas Diakopoulos explains how systems classify, rank, associate and filter, while embedding human choices about criteria, data and interpretation. An operator who makes the final decision still sees a field shaped by earlier computational selections (Diakopoulos, 2015, pp. 2–6).

The green button may therefore sit at the end of a narrow tunnel. If the system withholds alternatives, uncertainty and source material, the person reviews a presentation of the result rather than the result itself. If hundreds of cases await approval, review becomes ritual. If refusal requires a report while agreement needs one click, the interface has voted before the operator.

The loop can function like a lead: the organisation retains a human face, while the person follows a route fixed by the system and the pace of work.

Who acts in a hybrid arrangement?

Human–Machine Communication moves beyond treating technology solely as a channel between people. A conversational system can perform as a communicator: producing utterances, responding and shaping a relationship. Aleksandra Skrzypiec presents functional, relational and metaphysical frames for HMC and shows how AI in mass communication unsettles the roles of sender, recipient and author (Skrzypiec, 2025, pp. 35–38).

This does not make a model a morally responsible person. It does help us see distributed agency. An output emerges through data, model design, prompts, organisational policy, interface, reviewer and distribution system. No single part tells the entire causal story. ‘AI did it’ compresses the arrangement precisely where accountability requires expansion.

Distributed agency must not become diluted responsibility. Because several components shape the outcome, an organisation must assign human duties more precisely at each stage. A model cannot hear an appeal, compensate a victim, telephone the subject of a story or alter a budget. Responsibility needs an address outside the interface.

Responsibility, accountability and an Italian line in the sand

Michał Chlebowski distinguishes responsibility — duties attached to a role — from accountability, the mechanisms through which media are called to answer for commitments to society. Journalism needs standards and good intentions, but also routes by which a newsroom can be required to explain and repair its work (Chlebowski, 2024, pp. 115–119).

‘A human checks’ describes, at best, a duty. It says nothing about evidence that the check occurred, the quality threshold or the procedure after failure. An organisation may own an elegant AI policy without measuring review time, registering objections or allowing an operator to halt publication. Responsibility then lives in the document while accountability disappears from practice.

The Italian journalists’ code draws a strong boundary: ‘artificial intelligence may in no way replace journalistic activity’ (author’s translation; Consiglio Nazionale dell’Ordine dei Giornalisti, 2024, art. 19, p. 8). The same article requires journalists to disclose AI use in producing or modifying text, images and sound, retain control and responsibility, explain the nature of the contribution, and verify sources and data.

That is a direction rather than an operating manual. What counts as control when fifty items arrive at once? Which AI contribution is material enough to disclose? Who can evaluate a statistical analysis? Can a reviewer delay publication without punishment? Ethics becomes operational only after these unglamorous questions receive answers.

The audience cannot see the diagram

In a six-country Reuters Institute study, only 12% of respondents said they were comfortable with news made entirely by AI. The figure rose to 21% when a ‘human in the loop’ was added, to 43% when a human led with some AI assistance, and to 62% for entirely human-made news (Simon et al., 2025, p. 8). Only 33% thought journalists always or often checked AI output before publication (p. 9).

These attitudes do not prove that a particular newsroom is irresponsible. They do show that oversight language does not automatically produce trust. Readers cannot see the training programme, workflow diagram or administration panel. They see the resulting journalism, its label and the newsroom’s behaviour after an error.

Nguyen proposes a framework for responsible AI integration involving providers, journalists and media organisations, with transparency, training, policies and human oversight (Nguyen, 2026, pp. 2–7). The next step is to treat oversight as a sequence, not a point. A human must be present where data, output, publication and consequences can still be changed.

A human reviewer needs five things

Real oversight has five conditions. Remove one and review may become an alibi.

Knowledge. The reviewer understands the subject, the tool’s limitations and the editorial standard. Fluency with buttons is not expertise in a legal claim, a war image or a statistical model.

Context. The reviewer sees source material, relevant prompts and transformations, uncertainty and previous decisions. A summary without its record is like checking a passport from the colour of the suitcase.

Time. Case volume and deadline permit the promised standard. If verification takes two minutes, allocating twenty seconds is not an ambitious target. It is the mathematical cancellation of the procedure.

Authority. The person can refuse, halt, demand changes and escalate without disproportionate penalty. An operator liable for the outcome but unable to alter it is the organisation’s bumper.

Answerability. After publication, somebody must explain the decision, hear a challenge and repair harm. Oversight that ends at the click does not cover the life of the story.

The person affected belongs in the loop too

‘The human’ in oversight discussions usually means an employee of the organisation. The subject of a story or recipient of a decision should also stand at the centre. It is their reputation, right, access to a service or place in public debate that may be damaged.

Procedural-justice research shows that people judge institutions not only by outcomes but by process: opportunity to be heard, neutrality, respectful treatment and trustworthiness of the authority (Tyler, 2006). An automated process can perform well on average while denying a particular person a voice. ‘The algorithm made the decision’ closes the conversation where it should begin.

Someone wrongly identified in an AI-assisted story must be able to reach a person who sees the evidence dossier and can correct the publication. They should not be sent to the model provider or an unbounded form. Dignity means not treating them as an exception that spoils the metric. Freedom means a genuine chance to challenge. Accountability means a named decision-maker inside the institution.

Five gates of accountability

Replace the single green button with five gates. Each gate needs an owner, required evidence, escalation threshold, right to stop and route of appeal.

The process owner defines the purpose, permitted data, legal or ethical basis, sensitivity, retention and prohibited uses. They assess representativeness and risks to sources and subjects. Evidence takes the form of a data-use record approved by somebody competent in the domain.

Escalation is triggered by confidential or personal data, information about children, health, violence or protected sources, and unclear provider terms. The responsible editor and relevant security, legal or data-protection lead can stop the use. At this gate, ‘we will not use AI for this task’ remains the cheapest decision.

The output is an unverified proposal, not publishable copy. A qualified reviewer compares it with approved inputs and primary sources, checking facts, quotations, figures, omissions, tone, bias and uncertainty. They see the full context rather than a preview card.

Evidence is a verification log: what was checked, against what, what changed and by whom. Escalation follows when there is no source trail, evidence conflicts, a claim carries high stakes, or the reviewer lacks expertise to assess it. ‘I do not understand this’ is a valid stop signal, not a confession of weakness.

An editor assesses the whole item: public interest, proportionality, foreseeable harm, AI disclosure, headline, imagery and the subject’s opportunity to respond. They ask whether deadline pressure degraded the standard. High- risk work requires a second pair of eyes or independent domain review.

Evidence is a named publication decision attached to a stable version. Escalation is required for potentially irreversible harm, missing comment from an affected party or a material last-minute change. The right to stop must be stronger than the publishing clock.

Oversight continues after release. The newsroom monitors credible error signals, unintended effects, responses from subjects and anomalies in distribution or personalisation. AI labels and corrections should travel with the work as far as the publisher can reasonably ensure.

Evidence is a monitoring record and a named duty owner. A credible complaint, rapid spread, harm to an identifiable person or discrepancy with the source triggers escalation. The owner can pause promotion, add a warning or temporarily restrict an item while preserving the decision trail.

Every item offers an appeal route understandable to an ordinary person. A complaint receives a reference, timeframe and responsible human. The newsroom preserves the dossier, assesses the extent of harm, contacts the affected person and issues a correction proportionate to the original reach.

Evidence is a correction record, reasoned decision and process change where the fault was systemic. Appeal cannot depend on a complainant’s follower count. A person without a platform has the same right to a fair hearing as a public figure.

The oversight dossier

To stop the gates becoming another poster, each must answer five questions. Who? Name, role and deputy. On what basis? Sources, standard and required evidence. What triggers escalation? Specific thresholds, not ‘where necessary’. What may the human do? Stop, amend, seek expertise, withdraw or repair. How can an outsider appeal? Channel, deadline and decision level.

The dossier need not expose confidential data or security controls. It must let an auditor reconstruct whether meaningful oversight occurred or whether a person merely stood near the button.

Metrics need similar care. The number of approved items rewards speed. Measure review time, escalations, error types, appeals, repair time and use of

stop authority. Zero objections may indicate a flawless system. It may also indicate that objection is punished.

Where human oversight fails

Humans are not magic quality filters. They tire, carry biases, lack training and over-trust fluent outputs. Automation can weaken expertise when staff spend months approving rather than practising the underlying task. Excessive alerts produce routine dismissal.

Nor does every decision need identical review. A punctuation correction and an accusation against an identifiable person require different gates. Scale control according to potential harm, reversibility, data sensitivity, audience size and access to appeal.

Even an excellent internal process cannot cure every weakness in an external model or infrastructure. Sometimes the responsible choice is not to deploy. A person in the loop cannot legitimise an undignified purpose or a system that the organisation cannot adequately control.

Do not ask whether a human clicked

Ask whether they could see, understand, challenge and stop. Ask whether the organisation retained evidence of that work. Ask whether the person affected can reach a human with the power to change something.

The green button is a convenient symbol because it draws responsibility as the final point. In reality, accountability runs through the story’s whole life, from data to correction. A human should not decorate the end of the loop. They should hold the keys to five gates — and be required to open them from the public’s side as well.

The future rarely arrives with a date stamped on the box. It enters as a default setting, a faster workflow, a helpful summary or a green button marked Approve. By the time we call it transformation, many of its assumptions have already become ordinary.

The essays in this book have followed those assumptions into newsrooms, phones, answer engines, synthetic images and automated systems. Again and again, the central problem has not been that technology can do nothing useful. It is that usefulness can hide a transfer of power. A system selects before it answers, edits before it speaks and classifies before a person can object. Convenience is real, but so are the people who supply the data, absorb the errors, repair the outputs and disappear from the sales presentation.

This is why human-centred communication cannot mean placing a decorative person at the end of an automated chain. A human signature is not a click. It is the capacity to explain a decision, challenge the frame, protect a source, preserve uncertainty, refuse publication, correct harm and remain answerable afterwards. It also includes the freedom of the person affected to understand, contest and appeal.

The protocols proposed here are deliberately practical, but none is a moral vending machine. Checklists cannot manufacture courage, and provenance records cannot decide what deserves attention. They can, however, slow down the moment in which responsibility is most likely to vanish. They turn an invisible assumption into a question that colleagues, audiences and communities can examine together.

To frame tomorrow is to decide what enters the picture, what remains outside it and who is allowed to hold the camera. Artificial intelligence can widen

access, reveal patterns and remove needless labour. It can also reproduce exclusion at extraordinary speed. The difference will not be determined by the machine alone.

Tomorrow will be shaped through thousands of small editorial, technical and institutional choices. Each choice needs an owner, an explanation and a route back to the people who carry its consequences. The future may be synthetic in parts. Responsibility cannot be. It still requires a human signature.