About NewsWatch
We read US election campaign coverage three times a day, check each story against the outlets reporting it, and publish what we find — including how confident we are, and where we might be wrong.
What this is
NewsWatch collects election reporting from 36 sources every eight hours. Each story is grouped with the other outlets carrying it, assessed, and sorted into one of three buckets: accurate reporting, fabricated or machine-generated content, or a mixture of the two. The percentages on the homepage are simply the shape of what we read that day.
Every verdict is published with its evidence — the signal scores, the claims examined, the confidence, and the caveats. You can disagree with any of them on specifics. That is the point: a judgement you cannot inspect is just an assertion.
Why it exists
Synthetic and low-sourcing political content is cheap to produce and increasingly hard to distinguish from reporting at a glance. Most people encounter it in a feed, mixed in with everything else, with no signal about where it came from.
NewsWatch does not try to be the arbiter of that. It does something narrower and more defensible: it shows you how widely a story was corroborated, how specific its sourcing is, and what an assessment of its claims turns up — then leaves the conclusion to you.
Who runs it
NewsWatch is an independent, automated project. It takes no funding from any campaign, party, PAC or advocacy group, and it endorses no candidate. It is not a newsroom and does not do original reporting — it reads other people's.
The pipeline is non-partisan by construction, not just by intention. Story selection is penalised for over-representing one party, the writing rubric forbids partisan vocabulary and imputed motive, and a separate review pass checks each day's output for asymmetry before a human sees it. Those checks are described below.
How a verdict is produced
The short version: collect, group, assess, publish — with the limits stated at every step. The detail follows.
The pipeline
1. Collection. 36 feeds, every eight hours. Most are direct publisher feeds spanning wire services, legacy newsrooms, and openly partisan outlets on both sides. The rest are keyword searches that deliberately reach past the known-publisher list into the long tail — which is where synthetic and low-sourcing campaign content actually lives. A registry of only reputable feeds would make the “fake” bucket empty by construction and this whole site a tautology.
2. Relevance. Articles are kept if they match US election-campaign vocabulary. The filter is deliberately loose: a precision filter here would quietly drop the strange, low-quality campaign content that is the point.
3. Clustering. Stories are grouped across outlets by headline similarity. This produces the two numbers the site leans on hardest: how many independent domains carried a story, and how widely it was picked up.
4. Extraction. Article bodies are fetched and reduced to prose. When the full text cannot be retrieved, the feed summary is used instead and the article page says so — a verdict formed on two sentences of RSS blurb should not look like one formed on the full article, so confidence is capped when that happens.
5. Signals. Six transparent scores, all shown on every article page. None is a verdict on its own.
6. Assessment. claude-opus-5-5 reads the article alongside the signals and the corroboration context, and returns a label, a confidence, a summary, the claims it examined, and its caveats. The model is explicitly instructed that its training data predates these events, that failing to recognise something is not evidence of falsity, and to default to mixed rather than fake under uncertainty.
7. Debrief. A short briefing written from the stored verdicts.
What the three labels mean
| Real & factually accurate | Central claims check out, sourcing is identifiable, and nothing in the text suggests fabrication. |
| A mix of accurate and construed | A real event reported with significant distortion: misleading framing, missing context, or invented detail around a true core. |
| AI-generated, fake or fabricated | Central claims are fabricated or materially false, or the piece is generated filler with no reporting behind it. |
An article is not “fake” for having a viewpoint, for being written badly, or for being unflattering to someone. Opinion, analysis and openly partisan commentary are assessed as real when their factual premises hold.
The signals
Source trust is a reputation prior by domain, in five tiers. It is a starting point, not a conclusion: a tier-1 wire can publish something distorted and an unknown domain can carry accurate local reporting. Where a verdict rests mostly on reputation rather than on the text, the assessment says so.
Corroboration counts independent domains carrying a matching story, on a saturating scale — the jump from one outlet to three is meaningful; nine to eleven is not. It is discounted when every carrying outlet is low-tier, because content farms copy each other and that is not independence. Note the asymmetry: corroboration is good evidence that something happened, but its absence is weak evidence of anything. Stories are often local, early or narrow.
Specificity measures the density of named people, direct quotes, attributed statements, figures and dates. This is the most load-bearing heuristic in the set, because fabricated campaign copy is overwhelmingly generic about who said what, and when.
AI text markers combines sentence-length variation, lexical diversity, and the density of phrases that cluster in unedited model output. Read this one with suspicion. It is a weak prior that may support a verdict already grounded in missing sourcing or false claims; it may never carry one alone.
Emotional loading flags clickbait, outrage and hedging constructions. Byline records whether a named human author is credited.
Known limitations
We mostly cannot see the content people worry about. This is the single biggest limitation, and it explains why the accurate share runs high. Our corpus is news published to open RSS feeds and news-search indexes — which is, almost by definition, the work of organisations that operate like publishers. Fabricated and synthetic political content overwhelmingly circulates somewhere else: social platforms, paid ad networks, messaging apps, and sites that neither publish a feed nor get indexed. A high accurate percentage here is a statement about published news feeds. It is not a statement about what reaches you, and it should not be read as reassurance.
Coverage is partial. Only what appears in these feeds is assessed, and only up to a per-refresh budget. The percentages describe this corpus, not the information environment.
Source tiers lag reality. Reputation starts from a hand-maintained list, extended by structural rules for government, university and public-media domains. American local news has a very long tail, so a legitimate newsroom can still be scored as an unknown domain until someone notices. Where that happens the article page says the tier was inferred rather than verified.
Selection is not neutral. The classifier budget is weighted toward widely-carried stories and toward unknown domains. That is a deliberate choice — it finds more of what the site exists to find — but it means the “fake” share is not an unbiased estimate of how much fabricated content exists.
Clustering is approximate. Headline-similarity grouping will occasionally merge two distinct stories or split one, which moves corroboration and ranking.
Paywalls and blocks degrade verdicts. When a site refuses the fetch, the assessment falls back to the feed summary at reduced confidence. This systematically affects subscription outlets more than free ones.
Verdicts are not re-examined. An article is assessed once, when first seen. Corrections published afterwards will not change its label. Only its popularity and corroboration are kept current.
The model can be wrong, and confidently. Confidence is calibrated by instruction and capped by the quality of the evidence available, but it is still the model's own estimate of itself.
If you think a verdict is wrong
It may well be. Open the article page: it shows every signal score, the claims the model examined, what evidence it had, and what it flagged about its own limits. That is deliberately enough information to tell you why the system reached its conclusion — and therefore enough to disagree with it on specifics rather than in general.