Our Methodology
Every BiasChecker.ai analysis is an AI-assisted close reading of an article, run under a fixed set of instructions we have refined over thousands of analyses. The instructions are the same whichever AI model you choose — model choice changes the judge, never the rules.
This page introduces the thirteen analysis types and how to read the verdict each one returns. For the exact questions behind every analysis — and the ground rules they all share — see the questions behind every analysis.
We read the page, not the person
Every analysis describes the writing — the words on the page and the patterns in them. We tell you what a piece is built to make you see, feel, and conclude, and how it goes about it. That is a property of the text, and you can check it against the quotes we cite.
What we deliberately do not do is claim to know the author’s private intentions or state of mind. Naming who is biased, or asserting why someone wrote something, would be mind-reading — and unfair to the writer. So every lens is written about “the article” or “the text,” never the journalist or publisher: “the piece steers you toward…”, not “the author wants to hide…”. The same rule is why a finding is always anchored to a quote you can read for yourself, rather than to a motive we have guessed.
The thirteen analyses
Each analysis type gives the model a different expert role and a different question to answer about the same article. They fall into four families: Slant (how the piece leans and steers), Substance (whether what it argues holds up), Context & Extras (orientation, parallels, and two alternative renderings of the article), and Synthesis (an analyst's report written from the rest). For a closer look at the exact questions behind each one, see the questions behind every analysis.
Slant— Spot the slant
Intent Analysis
Identifies what the text says versus what it steers readers to believe or do. Intent is treated as a property of the writing — the direction the text pushes — never as a claim about the author's private motives. The result names the text's primary purpose and direction, and gives you a quick verification question you can check for yourself in seconds.
Persuasion
Looks for recognised persuasion techniques — emotional appeals, framing devices, logical fallacies — and explains how each one works on the reader, quoting the exact passage where it appears. How serious each technique is comes from our fixed catalogue of techniques, not from the model's mood on the day.
Bias Analysis
Describes how the text leans, with reference to specific words and structural choices, across our catalogue of bias categories — framing, omission, loaded language, source imbalance, political slant and more. Each finding cites the passages that drive it, and the overall bias level is computed by our code from the categories found.
Omissions
A dedicated "what's missing" audit: what a well-informed reader would expect to see but the article leaves out, why each gap changes the impression the article leaves, and what kind of source would fill it. Only consequential, established context counts — and a thorough, balanced article comes back with an empty list. When a piece has a serious flaw that can't responsibly be stated outright, the scan may add one or two neutral "questions worth asking" that let you weigh it yourself.
Substance— Check the substance
Critique
The "does it hold up?" lens. Instead of cataloguing rhetorical tricks, it evaluates the substance of the approach, argument or plan: are the goals right, is the approach sound, is it feasible, and what is being missed — engaging the strongest version of the argument before critiquing it.
Legal Risk
Assesses whether actions or proposals described in the text raise questions under applicable law, working out the relevant jurisdiction from the text itself. It stays strictly conditional about people and unverified facts: it flags what would need checking, it does not pronounce anyone guilty.
Moral Lens
An applied-ethics read of both the writing and the conduct it reports: how the text treats its subjects and its readers, and how the actions described measure against widely shared moral principles — keeping the two judgements clearly separate.
Scientific Assessment
Checks the scientific rigor of the text: whether claims are supported by the evidence offered, how methodology and statistics are used, what is established versus genuinely contested, and where the text overclaims beyond what its own sources can carry.
Context & Extras— Get the context & extras
Background
A neutral briefing for someone landing mid-story: what this topic is, how it developed, and where it stands now. Pure orientation — factual, concise, and without taking sides — rather than a critique of the article. Where the public record allows, it may add an optional "what was promised versus what happened" note: forecasts or pledges made about the topic at the outset, set against how they actually turned out.
Historical Parallels
A historian's lens: finds past events and patterns that genuinely parallel the situation described — analogies, not a timeline of the topic itself — and explains how each parallel resembles the present case and how it played out, ranked by relevance.
Neutral Rewrite
The article rewritten neutrally and more briefly: the loaded language and persuasion stripped out, the padding cut, and every fact, quote, and attribution kept exactly as the original had them. Strictly subtractive — it never adds a fact, a figure, or context that wasn’t in the article, so it reads as a de-spun, condensed version to check against the source.
Roast
A warm-hearted comedian doing a stand-up bit about the article — a punchy tagline, a playful retelling, and a few standalone jokes about its most roastable details. The bar is whether it’s genuinely funny, not whether there’s a flaw: a dry, clean piece gets nothing to roast, and tragedies or sensitive human stories are never joked about.
Synthesis— Then let it all come together
Synthesis
Synthesis is an analyst's report on the article, written by combining the other analyses' stored results — never a fresh, independent read of the text. When you run it, the four core slant lenses (Intent, Persuasion, Bias, Omissions) run first automatically, joined by any substance lenses a quick relevance check flags as worth running; analyses you already have are reused as-is. A single merge pass then writes the report: a one-sentence "what this article is doing" headline, a short opening read, a few titled sections connecting the findings into one story about how the piece works on its reader, an honest "what checks out" paragraph crediting where the article does its job properly, and a closing note on how to read it. Every section cites the analyses it rests on, and its supporting quotes are reused from those analyses' own evidence (validated by our code, never invented). The report describes the writing and its effect, never a named person's or outlet's motives, and any passage that fails our safety checks is withheld — counted and disclosed, not shown. The concern chip is computed by our code from the contributing analyses' own verdicts: a high chip requires at least two analyses to independently rate the article high, so one alarmed lens alone raises it to at most medium while its full finding stays visible on its own tab. It needs at least two existing analyses to run. Each contributing analysis is one model's stored result: Synthesis prefers results from the model it is itself running as, and falls back to the best other stored result for that article — so different sections of one report can rest on different models' analyses (each input's model is recorded in the result). For the most detailed synthesis, run the analyses you care about with your preferred model before running Synthesis — or run them as Consensus, whose reconciled multi-model results Synthesis picks ahead of any single model's.
The verdicts
Most analyses lead with a verdict — a quick, at-a-glance read of the result. Like every rating on the site it is computed by our code from what the lens found, never asked of the AI model. It comes in two forms.
Concern & Evidence chips
Bias, Persuasion, Scientific Assessment, Omissions and Synthesis lead with two chips:
- Concern — None, Low, Medium or High — how much the lens found worth flagging. It rises with the number of findings (for Bias, weighted by how impactful each one is): broadly, None means nothing notable, Low one isolated point, Medium a handful, and High a piece dense with them. Synthesis's chip is composite: the highest concern any contributing analysis derived, with the extra rule that a High needs two independently high inputs.
- Evidence — Limited, Moderate or Strong — how much verified, quoted support backs those findings.
What a Low versus a High actually tells you differs by lens:
Bias Analysis
A proxy for how much the framing — rather than the bare facts — shapes the impression you come away with. Low: a few minor or easy-to-spot leanings. High: slant runs through the piece across several high-impact techniques (framing, omission, loaded language), so it reads more like a particular take than neutral reporting. A high level is about emphasis and selection — it does not mean any individual fact is false.
Persuasion
A proxy for how hard the piece works to move you. Low: a technique or two. High: many persuasion techniques stacked together — often a sign that the conclusion the piece steers you toward isn't the one its own facts (or its headline) actually support. It flags the presence of technique, not dishonesty or intent.
Scientific Assessment
A proxy for how much to trust the science as presented. Low: a minor issue such as overstated certainty. High: several rigor problems — unsupported causal claims, misused or cherry-picked evidence — so the scientific claims warrant independent checking before you rely on them.
Omissions
A proxy for how complete the picture is. Low: one notable gap. High: several consequential things a well-informed reader would expect are missing — so treat the article as a partial view and seek out the missing context. Only omissions tagged as distorting move the level; useful-background items are listed without raising it.
Verdict headlines
Legal Risk, Moral Lens and Critique lead with a short headline from a fixed vocabulary instead of the chips — a count of findings would mislead for these lenses, because one load-bearing problem matters more than several background notes:
Legal Risk
Leads with one of no red flags, context noted, or legal questions raised. Each finding is 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, and the headline turns on whether any red flag is present. It is emphatically not a finding that anyone broke the law or is liable — only that, if accurately reported, these points are the kind a qualified professional might want to examine.
Moral Lens
Leads with a two-part sentence that keeps the axes separate — for example “the reporting is fair; the events it describes raise ethical concerns”. The writing and the conduct it reports are judged independently, so neutral coverage of grave events is never presented as a flawed article. The judgements are conditional and aimed at the text and the actions described — never at the character or motives of a named person.
Critique
Leads with one of holds up, holds up partly, or doesn't hold up as presented. Every suggested fix is tagged as either a load-bearing gap (the conclusion doesn't survive it) or a way to strengthen an argument that already stands, and the headline is computed from the load-bearing gaps alone — so a piece that holds up with a few polish suggestions still reads as holding up.
The remaining lenses (Intent, Background, Historical Parallels, Neutral Rewrite, Roast) carry no verdict at all — they exist to inform, reframe, or entertain, not to flag concerns.
How to read the verdicts. These are automated, heuristic indicators from an AI close-reading of the text — not factual verdicts, accusations, or professional legal, medical, financial or scientific advice. A higher level reflects how much a lens flagged in the writing, which you can always inspect through the cited evidence and judge for yourself; it is never a statement that any person did something wrong. Different AI models may surface slightly different findings on the same article. See our Terms for the full disclaimer.
Synthesis — one report, a panel of analyses
Every other analysis reads the article. Synthesis reads the analyses: it is a meta-analysis that takes the results this article already has and writes them up as one analyst's report — the read, how the piece works on a reader, what checks out, and how to treat it. It never re-reads the article to form new opinions of its own.
- Built from stored results, never from nothing. Synthesis reads the article's existing analyses straight from the result cache. Below two usable inputs it declines and says so, costing nothing — a synthesis invented from one lens would be a summary wearing a report's clothes.
- Which analysis of each lens it uses. It prefers the result produced by the model it is itself running as; failing that it takes the best other stored result, and a Consensus result is preferred over any single model's. So different sections of one report can rest on different models' work — each input's model is recorded in the result.
- Nothing invented. Every section cites the analyses it rests on, and its supporting quotes are reused from those analyses' own evidence, matched by our code: a reference that cannot be resolved back to a real input is dropped rather than printed.
- The concern level is computed, not written. It comes from the contributing analyses' own server-derived verdicts, and a high reading requires at least two of them to rate the article high independently — so one alarmed lens raises the report to at most medium while its own finding stays fully visible on its own tab.
Synthesis and Consensus are different axes, and they compose. Consensus takes ONE analysis and runs it across several models, reporting where they agree. Synthesis takes ONE model and runs it across several analyses, reporting how they fit together. Run the lenses you care about as Consensus first and the two stack: Synthesis picks those reconciled multi-model results ahead of any single model's, so the report rests on what a panel agreed rather than on one model's reading.
Consensus — one analysis, a panel of models
Any single AI model has blind spots and quirks. For the analyses whose findings come from a fixed catalogue (Bias, Persuasion, Legal Risk, Moral Lens, Scientific Assessment) we offer Consensus: instead of one judge, a panel of models from different AI providers each analyses the article independently, and an editor model then merges their findings into one result — under the same fixed instructions as any single-model run.
- A group is a maker. Models are grouped by the company that built them — Anthropic, OpenAI, Google, DeepSeek, Z.ai (GLM), Moonshot AI (Kimi), xAI — and the panel seats one model per group. The grouping is what makes "two models agreed" mean something: two models from the same maker share training data, tuning choices and house style, so their agreement is closer to one model agreeing with itself than to a second opinion. A group is deliberately not a price tier or an open-weights/closed-weights split — those cut across makers and would let siblings vote twice.
- Three makers, and cached seats are free. A run seats three groups. Any analysis of the article already cached from a group's model takes that group's seat for free (the highest-tier cached model wins the seat), and every covered group joins — so a well-analysed article can seat four or five makers at no extra cost, while a run never buys a fourth. When the cache already covers three, the only thing a consensus pays for is the merge.
- Who fills an empty seat. Seats still empty are filled by fixed representatives, in order: Claude Haiku 4.5 for Anthropic, GPT-5.4 Mini for OpenAI, Gemini 3 Flash for Google and DeepSeek V3.2 for DeepSeek. The merge is performed by DeepSeek V4 Pro, the same editor model every time, and one that never holds a panel seat — a judge merging its own analysis would favour its own reading. This is the current line-up; we review it as models change, so the representatives are likely to differ over time. Once a consensus is built for an article it is kept as is and records the models that actually sat on it; analyses cached later do not rewrite it (see Strengthening a panel below).
- Two halves: what the panel agreed, and what one model saw. A finding enters the consensus proper only when at least two panel members independently reported it — our code counts the agreement and enforces the rule before the result is accepted; it is not left to the editor model's judgement. Each agreed finding carries how much of the panel backed it ("3 of 4 models"). Findings only one model raised are not discarded: they are set aside under Raised by one model, with the model named and its own evidence intact. Models attend to different parts of an article, so a single voice usually means the others were reading elsewhere — not that they checked and disagreed. An empty consensus is still a valid outcome: if the panel agreed on nothing, the result says so, and where the panel split on the headline verdict the result reports the split rather than picking a winner.
- Why some analyses have no Consensus. Counting agreement needs findings our code can match across models, which means a fixed catalogue. Omissions and Critique produce free-form claims and suggestions instead, so there is nothing to match: two models can describe the same gap in different words, or different gaps in similar words, and no amount of merging makes that countable. Rather than let an editor model decide what "the panel agreed" means, we do not offer a consensus for those two.
- Same rules, same scoring. The merged result flows through the lens's ordinary pipeline — verified quotes, code-computed severities and verdicts — exactly like a single-model result.
- Full provenance. The result records which models sat on the panel, which of them were reused from cache, and which model performed the merge. A consensus never ships on fewer than three members: if a panel member fails to respond on a well-covered article, the remaining members still produce a consensus and it is clearly flagged as a partial panel, but on a three-seat run there is nothing to fall back on and the analysis reports the failure so you can run it again.
- Strengthening a panel. Because a cached consensus is never silently rewritten, an article can outgrow the panel it was judged by — a later analysis may add a maker the panel never had, or a stronger model in a maker it did. When that happens the result offers a Strengthen control naming exactly what it would add and what it costs — in practice the merge pass alone, because a rebuild is only offered on coverage that is already cached, so nothing new has to be run. The offer appears only when the new panel would be strictly stronger than the stored one.
Where the panel agrees is the strongest read an analysis can give; where a single model diverges from it, that tells you something about the model. Pricing and when to choose Consensus over a single model are covered on the models page and in the FAQ.
Publisher content & community analyses
An analysis is read in your own browser, at your request, from the page you are looking at. What we keep and share afterwards is deliberately minimal:
- We store our analysis, not the article. Community-shared results contain our AI-generated commentary plus short positional markers that let your browser re-locate the quoted passages on the live page — not the article text itself. Short excerpts shown alongside findings are quoted for analysis and review, always attributed and linked to the source.
- The community feed is curated and signed-in only. Shared analyses surface only inside the product to signed-in users, and only for outlets on a hand-maintained list. Every outlet on that list has been checked against its machine-readable text-and-data-mining preferences — the TDM Reservation Protocol (
tdmrep.json), ai.txt, andnoaidirectives — and we honour a reservation by leaving the outlet out of the shared feed; the same check applies before any outlet is added. - Disputes are heard. Every shared analysis can be rated by readers, and analyses enough readers flag as inaccurate are hidden automatically. Publishers can reach us via the contact page to dispute or request removal of an analysis of their content.
Analyses are AI-interpreted patterns and opinions about how a text argues — never verified facts about the publisher, and never a fact-check. See the terms of use for the full framing.