Framecheck
Thirty-two recurring ways an article, a headline, or a fact-check can be built entirely from true statements and still leave you believing something false. This isn't a truth detector — nothing here tells you what's true. It shows you the move, with the exact evidence, so you start catching it yourself.
Paste an article, see what gets flagged
Checked against all 32 patterns above. Mechanical flags are direct structural comparisons; judgment flags are candidates, not verdicts — check the quotes yourself either way.
The 32 Patterns
Attacking the Wrong Target
The rebuttal is true, but it isn't answering the claim that was actually made.
Fact-checking the wrong assumption
The response debunks the weakest, most literal reading of a claim, then treats that as having addressed the substantive point.
Example: "printing money" answered as a literal-banknotes claim, never as deficit spending financed via central-bank bond purchases.
Non-answer fact
A true, verifiable fact is offered as if it resolves a different question than the one actually asked.
Example: "inflation was global" doesn’t address whether a country’s own policy added to its own inflation on top of global factors.
Debunking the exaggerated version
A claim stated with hyperbole or a stereotype gets debunked at its most extreme reading, and that's treated as debunking the underlying pattern too.
Example: "that’s a stereotype" answers the overstatement, while the actual frequency of the underlying pattern is never examined.
Magnitude as truth-value
Disputing how widespread something is functions as an answer to whether it happens at all — without addressing the other reported instances.
Example: "not widespread" used to answer "does this happen," while independent reports from unconnected locations go unaddressed.
Cherry-picked anecdote as representative
The mirror of magnitude-as-truth-value: one dramatic, atypical case is presented as if it represents a whole category, with no base rate given either way.
Example: a single extreme incident cited as proof of a widespread trend, with no data on how common it actually is.
Sourcing & Verification
Where a claim actually comes from, and whether that's disclosed.
Citogenesis / churnalism
An outlet's claim traces back only to other outlets repeating each other, with no primary source — document, quote, record — ever reached at the root.
Example: outlet A cites outlet B, which cites outlet C, which turns out to be citing outlet A.
Single anonymous source, multi-source confidence
A claim resting on one unnamed source is delivered with the same tone of certainty as an independently corroborated one.
Example: "sources familiar with the matter" carrying a central claim, with no second, independent source ever named.
Undisclosed funding for "independent" experts
A quoted expert or think tank is funded by the industry, company, or political actor the story concerns, without that relationship disclosed.
Example: an "independent analyst" quoted on a policy, funded by the exact side that benefits from it.
Public funding, unstated independence assumption
Applies evenly to every publicly or state-funded outlet — BBC, CBC, NPR, PBS, RT, CGTN, Voice of America, Al Jazeera, DW — regardless of which currently reports well.
Example: coverage citing a state-funded broadcaster without noting who funds it — true of outlets across every alignment.
AI-targeted content flooding
Large volumes of content, engineered by structure and sheer volume rather than for human readership, are published specifically to be scraped into AI training or retrieval data.
Example: a network publishing millions of articles never meant for human readers, later repeated as fact by AI chatbots roughly a third of the time in independent audits; separately, a documented multi-million-dollar campaign built sites styled as independent sources to shape chatbot answers on a live conflict.
Timing & Correction
When a claim lands, and whether the correction ever catches up.
Silent correction
A fast, wrong story is quietly edited or updated later with no notice, while the version already shared is the one everyone saw.
Example: a headline changed days later with no correction notice, while the original screenshot is still circulating.
Asymmetric prominence
The original claim gets full amplification and repetition by other outlets before it's verified; the correction, days later, gets a fraction of that reach.
Example: a front-page claim, a correction buried and unlinked from the original.
Deadline-locked timing
A claim lands before an irreversible decision point — an election, a vote, a verdict. Any correction afterward cannot undo the decision already made.
Example: a story suppressed or amplified in the days before a vote, corrected only after the vote closes.
Severe claim stated as fact ahead of verification
A severe, alarming claim — an atrocity, a deliberate attack on civilians — is reported as settled fact before independent verification exists, often reinforced with emotionally intense imagery. The emotional reaction lands before any check on the claim does, and a later correction (see “deadline” and “asymmetric reach” above) doesn’t undo an emotional reaction someone’s already had.
Example: a specific atrocity claim reported as fact the same day, with no named source or independent confirmation, alongside a genuinely distressing photo.
Language & Framing
The words chosen instead of the argument made.
"Lied" asserted as fact
Lying requires provable intent — knowing something is false and saying it anyway — not just being wrong or disputed. "Lied about X" stated in the outlet’s own narrative voice skips that proof.
Example: a headline stating someone "lied," when the claim is disputed or simply incorrect, not shown to be knowingly false.
Persuasive definition
A term's meaning is redefined so a contested question gets settled by the definition itself, while the word keeps its original emotional charge.
Example: a law restricting mandatory union agreements, named for the positive right it claims to protect rather than what it restricts.
Disagreement made taboo
The claim is protected not by evidence surviving scrutiny, but by making its negation costly to say out loud.
Example: a label chosen for its association with a historical atrocity, applied to policy skeptics on an unrelated question.
Severity label inflation
An extremity label is applied to a position that's simply on one side of center from the labeler's own frame.
Example: "extremist" applied to a mainstream position that merely disagrees with the outlet’s own frame.
Double standard
The same underlying situation gets a different label depending on which side is being described.
Example: a leader without a current election called one term in one country's coverage, a different term in a structurally identical case elsewhere.
Regime-label vs. nation-label switching
Choosing between the ruling party's name and the country's name controls who's implicated — the government specifically, or the whole population.
Example: the same action described as done by "the regime" in one case and "the country" in a comparable one.
Technical vs. plain register
The identical mechanism gets a clinical, expert-sounding name or a plain, alarming one, depending on who’s speaking and what they want the reader to feel.
Example: "quantitative easing" and "printing money" describe the same central-bank action; one register reads as neutral policy, the other as alarming.
Most emotionally potent detail leads
Among several equally accurate ways to describe the same event, the single most emotionally resonant detail is foregrounded ahead of a more general, equally true description — shaping emotional impact independent of whether either description is accurate.
Example: "killing a 14-year-old boy and injuring several others" instead of "killing several people," when both are equally true — the child’s age leads.
Numbers & Evidence
Real data, framed to produce a feeling rather than to inform.
Impossible or moving evidentiary bar
"No evidence of X" conflates "hasn’t cleared this specific threshold" with "false" — and the threshold is sometimes set, or kept moving, so it can never be cleared.
Example: a standard of proof restated each time existing evidence approaches it, so the claim stays permanently "unproven."
Correlation headlined as causation
A study’s "X is linked to Y" finding becomes "X causes Y" by the time it reaches the headline.
Example: "Study shows X causes Y" atop a paper that only ever measured correlation.
Leading poll-question wording
A survey question is phrased to produce the desired answer, and the resulting number is reported as neutral public opinion without the question text disclosed.
Example: a "most people believe X" statistic, with the actual question wording never printed alongside it.
Relative vs. absolute risk
The same underlying numbers get reported as whichever ratio produces the desired emotional scale — a dramatic relative percentage, or a reassuring tiny absolute one.
Example: risk cut from 2-in-1,000 to 1-in-1,000 reported as "50% safer" rather than "0.1 percentage points safer" — both true, same data.
Figure without its comparison baseline
A raw number — dollars, headcount, article count — is reported with no comparison baseline: no rate of change over time, and no sense of what’s typical for the specific group being cited. Without either, a lay reader has no reference point to judge the scale.
Example: a farmer’s $91,000 fuel bill reported alone tells a reader nothing — they have no idea what a farm normally spends on diesel, unlike their own experience filling up a car once a month.
One-sided coverage of a mutual relationship
The article covers one direction of a situation that’s inherently two-directional — both sides of a conflict, dispute, or exchange — without addressing the other side’s corresponding actions, even briefly. No specific counter-figure is needed to notice the gap; the imbalance itself is the tell.
Example: an article on one side’s strikes in an ongoing exchange of strikes, with no mention that the other side is doing the same thing around the same time.
Distribution & Presentation
The layer most people actually see — headline, image, byline — versus what's underneath it.
Consensus ≠ truth
"Most experts believe X" is cited as if it settles the matter, without engaging the specific dissenting argument actually being made.
Example: a majority-view citation used to close a debate, with the minority position’s actual reasoning never addressed.
Headline mismatch
The headline makes a stronger or different claim than the article body actually supports — the layer that's most shared and least read.
Example: a headline stating something as settled that the body itself hedges, disputes, or never mentions again.
Weasel-worded collective will
"[Group] want/believe X" is stated as fact for a large, diffuse, unmeasured population, with no poll or survey behind it. Doesn’t apply to a small number of specific, identified parties (e.g., “the two sides” in a bilateral meeting) stating their own agreed position — no poll is needed when the parties speaking for themselves are the entire population being described.
Example: "Canadians want more of X" printed with no citation, on a question where actual polling often runs the other way.
Misleading image or video pairing
An old, unrelated, or out-of-context image runs alongside a story, implying it's from the current event, without a single sentence in the text being false.
Example: file footage from a different year or location run under a current-events headline with no date or location caption.
What's actually buildable
Sourcing, Timing, and Numbers mostly reduce to a structural comparison between two texts or two numbers — checkable by anyone, low room for any single model's own bias to leak in. Language & Framing, and parts of Attacking the Wrong Target, need real interpretive judgment, which is exactly where a model's own training leaves fingerprints — so the honest design there is to surface the candidate pattern and its evidence, never a verdict.