Why MBFC was never enough to assess the credibility of news

The structural limits of rating outlets

Start with the most basic one: it judges the outlet rather than the individual article. Every story from a high-credibility wire service inherits the same rating, whether it’s a 1,500-word investigation or a two-sentence stub generated overnight. Quality varies enormously within a trusted outlet, and a source-level label cannot see that variance. The reader’s real question, “should I trust this piece?”, goes unanswered.

Coverage is also incomplete. MBFC rates a few thousand outlets, mostly Western and English-language. The long tail of regional press, niche trade publications, foreign-language sources, and the endless supply of blogs is simply unrated. When a domain isn’t in the table, a source-level system gives you nothing. In practice, roughly half the domains flowing through a real ingest pipeline have no rating at all. Treating all of them as “unknown” isn’t a credibility assessment. It’s a shrug.

Then there’s the timing problem. Ratings are reviewed by hand and updated infrequently. An outlet that gets quietly acquired, rebranded, or hollowed out into a content farm keeps its old, flattering rating until a human gets around to re-reviewing it. Manipulation moves in real time. A label updated once a year can’t.

A rating also says nothing about how a story spread. A “high factuality” outlet can be the unwitting amplifier in a coordinated copy-paste campaign. Source ratings judge each outlet in isolation, so they’re blind to the relationships between sources: the synchronized timing, the shared citations, the eight “independent” outlets that all roll up to the same parent company.

And a rating can’t see synthetic text. No source-level score has any view into whether a specific article was written by a human or generated by a language model. That’s a property of the bytes on the page, not the masthead above them.

Finally, leaning on one external authority makes it a single point of failure. You inherit its blind spots and its politics, and any rating you can’t override or supplement becomes a liability the moment it’s wrong.

The pattern across all six is the same. MBFC answers “is this outlet generally reliable?” That is an important question, but it isn’t the operational one a reader has, which is “should I trust this specific article, right now, given how it reads and how it spread?”

How Trust Signals fixes this

The solution isn’t to throw out the baby with the bathwater, e.g. source ratings. It’s to use them only for what they can realistically measure, usually the reputation of the outlet, and to assess the article itself and how it spread with separate signals. That’s exactly what we built Trust Signals to do.

Applied to a single article, the three layers produce five signals.

Trust Signals on a single article. The five readings map onto the three layers: Source and Owner cover the outlet, Claims and Origin cover the article itself, and Spread covers how it traveled.

The first layer is the source: is the outlet generally reliable? This is where MBFC lives, alongside ownership rollups that expose when nominally separate outlets share a parent, a numeric reputation tier, and flags for state-affiliated media. When a domain has no rating, we don’t stop. We infer its bias from where the outlet’s articles sit relative to rated neighbors, and we attach a confidence score so low-confidence guesses get discounted rather than trusted blindly. That alone keeps the unrated half of the web from being treated as a void.

The second layer is the content: is this article substantive and authentic? Every article gets a quality score at ingest time, computed by a fast, inspectable heuristic that penalizes empty stubs, syndicated boilerplate, and link-soup aggregator pages, without punishing concise wire copy for being short but complete. On top of that runs an AI-tells detector, dozens of structural and stylistic checks that flag synthetic text and escalate to a language-model verdict only when the heuristics cross a threshold. This is exactly the layer a source rating structurally cannot provide.

The third layer is the network: how did this story spread? This is the failure mode source ratings are blind to, where reputable-looking outlets behave as a coordinated bloc. We detect near-duplicate copy and shared-citation patterns within tight time windows, roll them up per article, and score amplification by blending republication rate, ownership concentration, and state affiliation. The clearest illustration is what we call a non-independent verdict. An article can have plenty of apparent “corroboration” and still come from a high-MBFC outlet, yet be untrustworthy because the corroboration is just coordinated copies. You only catch that by combining the source, corroboration, and network signals.

Reputation as a weight, not a verdict

The key difference is where the source rating sits in the decision. Instead of being the full and final answer, MBFC becomes one input among several. A low reputation tier nudges the confidence score down, but it doesn’t override what the article’s text and spread pattern are telling us. Analyst overrides sit on top of everything, so the platform can correct or escalate beyond any external source.

The label itself comes from a verdict ladder that weighs corroboration against coordination, with a veto reserved for articles whose claims are actively contradicted:

The principle underneath all of it is defense in depth. Source reputation is static and outlet-wide. The quality score is per-article but shallow. AI-tells catch synthetic text, and coordination and amplification catch network manipulation. Each signal covers a blind spot in the others, so an adversary would have to defeat all of them at once. Deterministic checks do the bulk of the work, since they’re fast to run and easy to audit, and the language model is reserved for the genuinely hard judgment calls.

MBFC was a reasonable starting point for a question the internet, and our polarized world and media landscape, have since outgrown. The credibility of news was never something you could read off a masthead. It’s a property of this article, written this way, spreading this fast, and assessing it honestly takes more than one rating ever could.