The Questions Behind Every Analysis

Every analysis is the same article read under a fixed set of instructions. This page is the plain-English version of those instructions: the expert role we give the AI model and the questions it sets out to answer for each analysis type. It’s a more detailed companion to our methodology overview.

Two limits hold across every analysis. It never asks the model for a numeric score — our own code computes those from what was found. And it never accuses a real person: an automated reading of text can describe the writing and the conduct it reports, but it cannot establish anyone's guilt or motives, so it does not claim them.

Ground rules every analysis shares

The same handful of rules sits behind all the analyses below, so we state them once here rather than repeat them under each one:

Just the article text. We read the article's written text — the menus, ads and cookie banners around it are ignored. Images, video and audio aren't extracted either, so anything carried only by them falls outside the analysis.
It describes the writing — it can't judge people. The analysis states plainly what the writing does, but an automated reading of text has no way to establish a real person's guilt, motives, or inner state — so it never claims them. Findings describe the article and the conduct it reports, stay conditional about anything unverified, and are never accusations against the journalist or anyone named.
Verbatim, verifiable evidence. Findings must quote the article exactly (short fragments), and we check every quote against the source text so you can judge it yourself.
It trusts the article on recent events. The model's training has a cut-off date, so when the article describes something more recent, the analysis goes with the article rather than "correcting" it from possibly outdated knowledge.

What each analysis asks

One per analysis type. Each gives the model a different expert role and a different set of questions to answer about the same article.

Intent Analysis Intent

Surface vs intended meaning

We ask the model to read the article like an editor and separate what the text literally says from what it steers you to believe or do — treating intent as a property of the writing, never a guess about the author’s private motives.

The questions it asks

What does the text literally claim or report — its surface message?
What conclusion, belief, or action is the reader actually being guided toward?
Which techniques — word choice, emphasis, structure, whose voice gets heard — do that steering?
Does the headline match what the body actually supports — or is the gap between a punchy headline and a more careful body itself part of the steer? In the extreme, is the headline a lure for a body about something else entirely?
Is there anything buried in the text that quietly undercuts the steered conclusion?
What is one simple thing you could check in the text yourself to test the read?

Persuasion Persuasion

8 persuasion mechanics + 8 logical fallacies

We ask the model to examine how the writing persuades you — including where it tips into manipulation — by surfacing the hidden assumptions it asks you to accept and the conclusion it leads you toward.

The questions it asks

What is the reader expected to take as true without proof?
What conclusion is the piece leading the reader toward?
Which mechanical tactics are active — burying key facts, scattering context out of order, or normalising a fringe view so it sounds ordinary?
Which classic logical fallacies appear in the article’s own narration — false dilemma, post hoc, guilt by association, cherry-picking and the like?

The analysis scans for the same fixed set of mechanical tactics and logical fallacies every time, flagging only the ones the article’s own narration genuinely performs. The article’s overall lean — loaded language, framing direction, selective emphasis — is deliberately not in this catalogue: that is the Bias analysis’s job:

Information placement — how the text arranges what it includes

Buried information — Key facts or qualifications placed late, after the bold claims.
Contextual scattering — Related information spread apart so it’s hard to connect.
Selective chronology — Convenient gaps in the timeline; cherry-picked time periods.
Timeline manipulation — Events presented out of order to suggest a false cause.

Normalisation & mainstreaming — how the text shifts what counts as acceptable

Euphemism — A sanitised or technical label that makes an extreme action sound routine.
False consensus — Presenting a fringe view as something “everyone” already believes.
Extremism mainstreaming — Framing a radical position as reasonable, moderate, or inevitable.
Legitimising association — Lending an extreme idea credibility by placing it beside respectable ones.

Logical fallacies

Post hoc — Treating “B came after A” as proof that A caused B.
Hasty generalization — Broad conclusions from a few unrepresentative examples.
False dilemma — Forcing a choice between only two options when more exist.
Cherry-picking — Using only the evidence that supports the conclusion.
Appeal to emotion — Fear, pity, or outrage standing in for an argument.
Guilt by association — Discrediting an idea by linking it to something disliked.
Begging the question — Assuming the conclusion in the premise; circular reasoning.
False equivalence — Treating fundamentally unequal things as equivalent.

A tactic is only flagged when the article’s own voice performs it — rhetoric inside a quote the article reports belongs to the speaker, not to the article, and quoted emotional testimony from named people is reporting, never the article’s own appeal to emotion. Openly visible registers are not covert persuasion: a comedy piece, a warm profile, or plainly promotional copy (a press release, an organisation’s own announcement) wears its slant on its sleeve, so its favourable selection of facts is the disclosed form at work — not cherry-picking — and the usual correct output for such pieces is an empty list.

Bias Analysis Bias

29 bias categories

We ask the model to measure how much, and in which direction, the piece leans — pointing to the specific words and structural choices behind the tilt, across our catalogue of 29 bias categories.

The questions it asks

Which perspective does the coverage favour, and how strongly?
Where do word choice, emphasis, or structure tilt the framing?
Is the sourcing one-sided, or are obvious perspectives left out?
When casualty or other figures for more than one side are reported, is the depth of coverage proportional?

The analysis scans for all 29 categories below — grouped by how much impact each tends to have — and reports only the ones the article’s own narration genuinely shows:

High impact

Framing — How information is presented to steer interpretation (includes implicit and agenda-setting bias).
Political — Political leaning or partisanship (includes nationalistic framing).
Racial — Race-related bias in coverage or language.
Gender — Gender-related bias in coverage or language.
Religious — Bias for or against a religion; faith-based assumptions or anti-religious framing.
Commercial — Undisclosed sponsorships, affiliate links, or financial interests.

Moderate impact

Source — Over-reliance on a narrow set of sources (includes authority bias).
Confirmation — Reinforcing an existing belief or narrative.
Selection — Cherry-picking data, quotes, or examples — including a slant-creating gap, where a missing perspective tilts the story (a full what’s-missing audit is the separate Omissions analysis).
Emotional — Loaded language and appeals to emotion.
Sensationalism — Exaggeration, clickbait, hyperbole.
Stereotyping — Generalisations about groups (includes socioeconomic).
Headline — Misleading or sensational headlines and titles.
False balance — Giving equal weight to unequal positions (bothsidesism).
Citation — Selective citation of supporting studies while ignoring contradicting ones.
Survivorship — Focusing on successes while ignoring failures.
Recency — Overweighting recent events over historical context.
Anchoring — Over-reliance on the first information presented.
Cultural — Ethnocentric or Western-centric perspectives.
Statistical — Misleading use of statistics; reading correlation as causation.
Age — Ageism and generational stereotypes.
Negativity — Systematic overemphasis on threats, failures, and bad outcomes.
Geographic — Urban-vs-rural or regional dismissal and stereotyping.
Disability — Ableism and problematic portrayals of disabled people.
Narrative — Forcing events into a hero/villain or rise/fall arc that distorts reality.
Access — Only covering stories journalists can easily reach.
Automation — Over-trusting AI or algorithmic outputs as fact.
Bandwagon — “Everyone believes X” appeals to popularity.
Just-world — Implying victims deserved their fate.

We assess only the article’s own narration — not inflammatory quotes it reports — and treat factual context, skeptical framing, and “contacted for comment” as signs of balance, not bias. The analysis first identifies the article’s register: visible comic tone and openly promotional copy (a press release, an organisation’s announcement on its own site) are disclosed forms, not hidden leans — a promotional piece that reports its facts accurately has its lean named at most once, never restated under several categories. The same dedupe rule applies generally: one lean gets one primary category, not the same tilt filed under framing, narrative, emotional, and sensationalism at once.

Omissions Omissions

What a well-informed reader would expect but the article leaves out

We ask the model to act as an investigative editor and run a dedicated “what’s missing” audit — the relevant, consequential, knowable context the article leaves out.

The questions it asks

What fact, context, perspective, counterpoint, or caveat would a well-informed reader expect to see?
How does each absence change the impression the article leaves?
Is each gap distorting — its absence changes the conclusion a reader draws — or merely useful background a fuller piece would add?
Does the headline fairly reflect the article — or does it overstate, contradict, or assert a certainty the body itself doesn’t support?
Where would a reader look to fill the gap — what kind of source, rather than a specific link?

Omissions are judged against the article’s own topic and claims; we don’t invent context we can’t stand behind, and a thorough, balanced article legitimately comes back with few or none. A fact stated anywhere in the article — including inside a quoted statement or rebuttal — counts as present (the model must re-check the full text before flagging each item), events that happened after publication are never omissions, and where the article clearly attributes a one-sided account (“prosecutors allege…”), the missing other side is usually background rather than distorting. Each omission is tagged distorting or background, and the concern badge is computed from the distorting ones only — so useful-context items still show without inflating the verdict. A headline that merely rounds or simplifies without misleading isn’t flagged — only a genuine gap between what the headline claims and what the body supports, the strongest form being a headline that promises content the body never delivers at all.

Critique Critique

Feasibility, assumptions, and alternative approaches

We ask the model to act as a domain expert and advisor and judge whether the approach, argument, plan, or decision actually holds up in the real world — focusing on substance, not rhetoric.

The questions it asks

When the text advocates goals of its own — are they the right goals, well-defined and well-justified?
Is the approach sound, and what are its strengths, gaps, and blind spots?
Is what’s proposed realistic and achievable?
What load-bearing unstated assumptions does it rely on — and are there better alternatives?
Is any figure, cost, or benefit genuinely out of proportion to the problem it addresses?

The analysis states the strongest honest version of the argument before critiquing it, and acts as a logic checker rather than a fact-checker. Each evaluation question applies only when the text genuinely raises it — plain event reporting isn’t graded against goals or proportionality it never advances, and an explainer describing a third party’s plan is reporting, not advocacy: the critique targets the article’s own reasoning, never the reported plan’s design, and no advice is aimed at actors the reader cannot influence. Each finding is made once rather than repeated across sections, “could also mention X” padding is pruned, and a piece that holds up gets a short verdict instead of a full evaluation scaffold. The model also tags every suggested fix as either a load-bearing gap (the conclusion doesn’t survive it) or a way to strengthen an argument that already stands; the headline verdict — holds up, holds up partly, or doesn’t hold up as presented — is then computed from the count of load-bearing gaps alone, never asked of the model directly.

Legal Risk Legal

Check legal compliance

We ask the model to act as a legal analyst and assess whether the conduct or proposals described could raise questions under applicable law — as disclosed interpretation, never a verdict on a person.

The questions it asks

Whose described conduct could raise legal questions, and why?
Which area of law is implicated — international, criminal, constitutional, regulatory, civil, or procedural?
Which jurisdiction’s law applies, inferred from the text itself?
What is the strongest plausible defence or mitigating factor for the conduct?

It can surface where the conduct described might raise legal questions, but it cannot establish that anyone actually broke the law — so it only points to what a qualified professional would want to examine, never a verdict on a person. Two things are explicitly not conduct: proposed legislation being debated (the lawful legislative process at work), and harsh political rhetoric, which is protected expression rather than a criminal threat unless the text describes a specific, credible intent to harm. Each finding is also tagged as either a red flag (a specific party’s concrete conduct of the kind courts or regulators actually act on) or background legal context (inter-state law framing of a conflict — even at grave scale — abstract questions about a law’s validity, protected advocacy); the headline — no red flags, context noted, or legal questions raised — is computed from whether any red-flag finding is present, never from how many findings a verbose model happened to enumerate.

Moral Lens Moral

Assess against moral principles

We ask the model to act as an applied-ethics analyst and judge whether the conduct is wrong — even where it is perfectly legal — looking separately at how the text presents events and at the events themselves.

The questions it asks

Does the text itself cross an ethical line — dehumanising language, mocking suffering, dishonesty?
Are the actions or events described morally problematic — proportionality, unnecessary harm, rights, exploitation?
Are the official justifications actually morally sufficient, or merely asserted?
Which specific moral principles are genuinely at stake here?

It weighs the article against a fixed set of moral principles in two groups — how the writing treats its subjects and readers, and whether the conduct it describes is justified — reporting only the ones genuinely at stake:

In the writing — how the article presents events

Dignity — Degrading or dehumanising language in the article’s own voice.
Truthfulness — Dishonesty or deception.
Compassion — Mocking suffering or celebrating tragedy.
Equality — Portraying groups as unequal in worth.
Wellbeing — Needlessly nihilistic; crushing hope without cause.

In the conduct described — the events themselves

Proportionality — A response out of all proportion; collective punishment.
Harm — Actions causing unnecessary harm without justification.
Justice — Unfair consequences or lawlessness.
Sanctity of life — Treating life as disposable.
Rights — Infringement of human rights.
Autonomy — Coercion, or restricting freedom without cause.
Responsibility — Avoiding accountability for the actions taken.
Privacy — Invasion of personal boundaries.
Property — Theft or destruction of possessions.
Fairness — Unfair treatment or distribution.
Exploitation — Taking advantage of vulnerable people.

It can weigh the conduct and the writing, but it cannot judge a person’s character — so its judgements stay conditional, aimed at what was done rather than who did it. A finding must also tell you something the article doesn’t already: natural events with no moral agent are not condemned, conduct that appears only in a passing background clause is scene-setting rather than this article’s events, and conduct the article itself presents as settled wrongdoing is treated as the story, not a discovery — though where the text contests, defends, or minimises the conduct, the lens’s own judgement does add information and is given. The headline sentence — for example “the reporting is fair; the events it describes raise ethical concerns” — is computed from the two axes, so neutral coverage of grave events is never presented as a flawed article.

Scientific Assessment Science

Check scientific accuracy and rigor

We ask the model to act as an expert in scientific method and assess the rigor of both direct scientific claims and any reporting on research studies.

The questions it asks

Are the claims factually correct against settled scientific consensus?
Is the evidence appropriate and strong enough for the claims being made?
Are cause-and-effect claims actually supported, or is correlation being read as causation?
For science journalism: does it represent the study’s findings, limitations, and methods accurately — or overstate them?
What evidence would actually settle a genuinely disputed claim?

Recent findings are not flagged as errors merely because they fall past the model’s knowledge cutoff, and clearly-labelled religious, moral, or philosophical claims are not judged as failed science. The scope test cuts both ways: an article making no scientific claims at all is marked not-applicable rather than “scientifically sound”, while a press release or op-ed asserting an empirical or causal claim in its own voice is fully in scope — genre never exempts the claim.

Background Background

Topic context: what this is, how it developed, and where it stands now

We ask the model to act as a reference editor and write a neutral briefing for someone who has landed on the story mid-stream — the factual history of the topic itself.

The questions it asks

What is this topic, and where does it stand today?
How did it develop — the key milestones, oldest to newest?
Who are the main parties or roles involved?
Which parts are settled history, and which are recent or genuinely contested?
Where the public record allows: what was originally forecast, promised, or projected about this topic — and how did that actually turn out?

It stays neutral — describing how each side characterises a dispute rather than taking a side — and flags low confidence rather than inventing dates, quotes, or statistics. Relative time (“recently”, “on Friday”) is resolved against the article’s own dateline, not today’s date. The timeline stays strictly anchored to the article; the one exception is an optional “what was promised versus what happened” note, where the model may add well-documented forecasts or pledges from the public record that the article itself omits — included only when it genuinely recalls concrete specifics (a named institution, a pledged figure, a dated forecast), never vague “critics predicted” filler.

Historical Parallels Parallels

Find historical precedents

We ask the model to act as a historian and surface two or three past events or patterns that genuinely parallel the situation — structural analogies, not a timeline of the topic itself.

The questions it asks

What historical event or period resembles this situation, and how closely?
What patterns match — and what is importantly different?
How did the historical case actually play out, and what does that suggest here?
What lesson does it offer?

A few strong parallels beat many weak ones — and when no genuine parallel exists (a celebrity item, a routine announcement, a how-to), the analysis declines outright rather than inflating a keyword-level match into one. It may close with a real, fittingly-chosen quotation that speaks to the situation; the quotation must genuinely exist and carry a named source, with any uncertainty about wording or authorship declared rather than hidden.

Neutral Rewrite Rewrite

The article rewritten neutrally and condensed — loaded language and persuasion stripped out

We ask the model to act as a plain-spoken sub-editor and rewrite the article neutrally and more briefly: keep the material facts and each side’s position, strip out the loaded language, emotional padding, and persuasion — including the persuasive structure (ordering, emphasis, framing) — and add nothing new.

The questions it asks

Which words are doing persuasion rather than reporting — and what is the neutral way to say the same thing?
What is emotional staging, repetition, or filler that can be cut without losing a material fact?
Is the article’s architecture itself steering the reader — a countervailing fact buried at the end, one side given far more room, or the narration framing events as threats or reassurances — and how should the rewrite reorder and rebalance to undo that?
Is every claim still attributed exactly as the original attributed it, rather than promoted into a flat statement of fact?

This lens is strictly subtractive: it never adds a fact, a figure, context, or a fact-check — anything not in the original article is left out. It may reorder and rebalance (leading with the most consequential facts, compressing an over-weighted side harder, moving a buried caveat next to the claim it qualifies), and when it strips a slanted sentence it must keep the facts that sentence carried — but it cannot restore perspectives or sources the original left out. It is also a rewrite, not a summary: a lean, factual piece should come back near full length, and meta-voice (“The article notes that…”) is banned and mechanically stripped — the output is the article, not a description of it. Direct quotations are used sparingly and only word-for-word; speech is otherwise reported indirectly. The model also names the story’s main actors, each verified verbatim against the article before display. A mechanical check then compares the rewrite against the original: a quotation or accusation that is not in the source blocks the rewrite outright, and a number that cannot be verified against the original is shown highlighted rather than silently trusted. It is a neutralized, condensed version of the article’s own reporting, so check it against the source before quoting.

Roast Roast

A warm, witty comedic take on the article

We ask the model to act as a warm, sharp stand-up comedian doing a bit ABOUT the article — its one job is to be genuinely funny about it, not to grade the journalism.

The questions it asks

What punchy one-line tagline captures the article — the comedic hook?
How would you playfully retell what the article actually says?
Which specific details — a phrase, a claim, an absurd number stated with a straight face — are most worth a standalone joke?
What is the one honest, humour-free takeaway that grounds the bit?

The humour only ever targets the writing and framing — never the people in the story or what they are going through. Real tragedies (deaths, serious harm, disasters) and sensitive human hardships (illness, disability, grief, addiction) are declined outright, and a dry, cleanly-written piece with no comedic angle gets nothing to roast — a forced joke is worse than none.

Synthesis Synthesis

One combined read: what this article is doing and the few things that matter, synthesized from the other analyses

We ask the model to act as an editor merging the article’s existing analyses — never adding analysis of its own — into one short trust read: what the piece is doing to its reader, and the few things that actually matter.

The questions it asks

In one sentence of 25 words or fewer, what is this article doing to its reader — leaning, steering, or informing straight?
Across all the analyses, which one to three points genuinely change how a reader should treat the piece? When several analyses flag the same issue, say it once.
Which specific findings from the input analyses support each point? Every takeaway must reference real findings by id — our code checks the references and drops anything unsupported, so quotes are always reused from the underlying analyses, never invented.
What does the piece do fairly — restraint, a steelman, clean sourcing — worth crediting honestly?

Synthesis only: the model recombines the other analyses’ stored findings and may not introduce any new claim or example. When the analyses found little, a calm, unalarmed overview with zero takeaways is the correct outcome — it never pads a takeaway just to fill the tab. The concern chip is not asked of the model at all — our code sets it to the highest concern any contributing analysis derived.

A living method

These instructions are refined continuously — sharper guardrails, clearer questions, new checks — so the wording shifts over time. This page reflects a review on 2026-07-17; the commitments above (no model-emitted scores, verbatim evidence, and never an accusation about a person) are stable and survive every revision.

Once an analysis runs, the result leads with a verdict — a quick read of what it found. The methodology overview explains the verdicts and what their levels mean.