The alert appeared on a Tuesday afternoon in the app many people in the district used to read local news. A headline said that the council had allocated €2.4 million to repair lifts in public housing. Beside the story shone a green label: “high reliability.” Marta, who lived in one of the affected buildings, opened the link with relief. Yet the figure came from a sentence attributed to a municipal source who had not published a budget, agreement or identifiable statement. The story named a councillor, but included neither a direct quotation nor a document. The green mark did not say that the money existed. It said something more confusing: that the article, taken as a whole, seemed safe.
That visual cue can help when it guides a first reading, but it becomes dangerous when it appears to decide for someone. Marta needed to know whether work would happen, when it would begin and how urgent repairs could be requested. The label answered none of those questions. The problem was not that a tool pointed to uncertainty; it was that it hid where the uncertainty sat. A responsible rating does not deliver a sealed verdict on a complete news story. It explains which claims it checked, what evidence it found, which part it could not verify and how its conclusion can be challenged. A colour can open a conversation. It should not close one.
A story contains claims with different reach#
Calling a story reliable as though it were an indivisible object mixes facts, attributions, forecasts and opinions. In the lift case there were at least five separable claims: that a €2.4 million allocation exists; that it will go to those buildings; that the agreement has been approved; that work will start on a particular date; and that current breakdowns will receive an interim response. Each needs different evidence. An approved budget may support the first. A list of properties may support the second. A procurement timetable may illuminate the third and fourth. A service protocol may clarify the fifth. No single indication makes them all true automatically.
Calling a story reliable as though it were an indivisible object mixes facts, attributions, forecasts and opinions.
An anonymous source creates a specific limit. It may provide a valuable lead, especially when it draws attention to a decision not yet announced, but readers cannot assess that person’s position, incentives or exact words. Nor can they tell whether the figure was an estimate, an internal request or money already voted for. Marking the entire story green because it fits earlier conversations turns a lead into a guarantee. The alternative is not to call it false merely because a document is absent. It is to say: “The principal figure has no verifiable published source; the attribution is recorded as an unidentified statement.” That sentence preserves usefulness without pretending to offer certainty.
Breaking the story apart also protects people who did solid work. The report might contain dated photographs of broken lifts, residents’ testimony and an email response from the maintenance office. Those materials could support the fact that breakdowns existed and affected a building. On their own, they would not support the announced amount or the start of construction. When a label distinguishes these elements, it recognises evidence where it exists and stops one weak point from either discrediting or laundering the whole text. Reading stops being “believe it” or “do not believe it” and becomes “this is documented, this is attributed, this awaits evidence.”
The evidence card must be readable#
An interface does not need to turn every story into an endless case file. It can offer a short card alongside every important claim. For the €2.4 million figure, the card could state the precise claim, the state of the evidence and its date: “Not confirmed by a public document; attributed to an unidentified municipal source; reviewed 21 September.” For the existence of breakdowns, it might say: “Confirmed through dated photographs, three testimonies and a maintenance response.” For a projected start date: “Attributed estimate; contract not published.” Labels are not decoration. They describe the kind of relationship between a sentence and the material available.
The categories must not conceal the difference between absence and conflict. “Not reviewed” means nobody has examined a claim yet. “No published source” means accessible support was sought and not found. “Conflicting evidence” means materials point in different directions. “Partly confirmed” means some components withstand scrutiny and others do not. If all of these become the same green, amber or red, the system reduces an explanation to a decoration. It may retain colour as a summary, provided the summary leads to plain language and does not replace it.
Readability is a matter of fairness, not cosmetic design. Marta should not need newsroom vocabulary to understand the reach of a source. The card should avoid statements such as “elevated confidence model” when what happened is that an algorithm found many citations while verifying none. It should also separate the person who supplied material from the person who evaluated a claim. When the app uses a database, it should show the version consulted and the date. When a human reviewer is involved, their role should be identifiable without exposing unnecessary personal data. An understandable explanation lets a person provisionally accept a rating while also knowing what to ask next.
Confidence is distributed, not averaged#
An average is convenient for a product maker: it gathers scattered signals and returns one number. But an affected person does not live inside an average. For a resident who needs a lift, the decisive fact is whether repair will happen and under what conditions. A respected outlet, an authentic photograph and a recent date do not compensate for the principal figure lacking published support. Nor does an administrative source prove that a plan answers the needs of people unable to leave their homes. The tool should show where its confidence is high, moderate or low, and why that difference matters for the decision at hand.
A simple claim-by-claim scale could do that work. The claim about breakdowns would have high confidence where there is direct, verifiable evidence. The existence of budget conversations might have medium confidence if several people agree but no minutes are supplied. The exact amount would have low confidence until a document or attributable statement appears. The construction schedule could remain undetermined. None of these descriptions requires dismissing the story. They help people avoid turning a promise into a fact or discarding a reasonable warning merely because it has not reached the highest level of proof.
This distribution requires criteria to be examined before they are automated. If a platform gives an institutional statement more weight than neighbourhood reporting, it should explain that it is assessing the ability to attribute a decision, not necessarily the experience of people who will live with its effects. If it penalises an unpublished source, it should say that this limits verifiability, not that it proves bad faith. Words matter because a label can affect the reputation of a journalist, an association or someone reporting a problem. Caution is not softness. It is precision about what we know and what we still do not know.
A useful appeal is part of the evidence#
The rating cannot depend on a single chance to provide context. Imagine that the following day the residents’ association receives a copy of the governing board’s agreement. Or the councillor publishes the complete minutes and clarifies that €2.4 million covers only a first phase. The story and its label should be updateable promptly, but the change cannot be a silent substitution. The record should show what changed, what new evidence was added, which conclusion was revised and when it happened. Then everyone who saw the original green mark can understand whether the amount changed, its scope changed or the missing source simply became public.
For this process to be real, the appeal button cannot be a hidden route or an inbox that never answers. It should state what material is relevant, allow someone to explain why a rating affects them and offer a reasonable review timeframe. A resident with no technical knowledge should be able to attach a letter, point to a public page or describe a contextual error. The platform may reject insufficient evidence, but it should give its reason by reference to the criterion applied. When a disagreement cannot be resolved, it can remain visible too: “The source maintains X; available documentation confirms Y; Z remains unresolved.”
Appeal protects both publishers and readers. Without it, a reliability tool can entrench an error precisely because its appearance of neutrality discourages response. With it, the system acknowledges that classifying information is revisable work. That requires human resources and rules against abuse, of course. Not every objection deserves identical treatment, and a coordinated campaign should not prevail through volume. Yet filtering harassment does not justify removing the possibility of correction. The legitimacy of a label depends on being able to explain not only how it began, but also how it changes.
Green should invite attention, not replace it#
Marta could use a well-designed label as a reading map. She would see that the breakdowns are documented, the amount requires a public agreement and the construction date remains open. She would then know where to ask: the housing office, the council meeting or the journalist who signed the story. She could also share the article without repeating as fact an amount not yet demonstrated. This is a modest but decisive improvement. Technology does not spare her judgement; it returns a judgement that is better situated.
Newsrooms and platforms have a particular responsibility when their visual signals resemble permission. A vivid green may lead one person not to open the link, another to dismiss criticism or an institution to feel confirmed before it has accounted for itself. That is why a system must reserve colour for orientation and accompany it with visible limits. It must record what it reviewed, accept new evidence, separate claims and prevent the absence of a published source from disappearing behind a global score. Trust is not manufactured by a light. It is built when reasons remain in view and can be disputed.
