Does this article — or video — contain bias or manipulation?

The model is trained on documented manipulation techniques — patterns identified by media literacy researchers — not on political positions. A sentence that uses loaded language is flagged whether it comes from the left or right. The model does have stronger coverage of geopolitical conflicts (Israel-Palestine, Ukraine-Russia) where historical context manipulation is especially common.

The model uses per-technique decision thresholds calibrated on a held-out evaluation set to balance precision and recall. In real-world use, you'll see both false positives (strong legitimate opinion flagged) and false negatives (subtle manipulation missed). Treat results as a starting point for critical thinking, not a final verdict.

Fact-checkers verify specific claims against evidence. This tool identifies techniques — ways language is used to manipulate rather than inform. A sentence can be factually accurate and still use manipulative framing. Both approaches are complementary.

Those tools rate publications as a whole — "left," "center," "right" — based on editorial patterns and ownership. Biasly is sentence-level: it ignores who published the article and looks at the language inside it, flagging specific manipulation techniques wherever they appear. The two approaches are complementary — a "centrist" outlet can run a piece full of loaded language, and a "biased" outlet can run a clean factual report.

For articles, videos, and pasted text, Biasly identifies who's being quoted — by category (government, military, analyst, civilian, organization) and by national affiliation — and flags imbalances. An article quoting 14 Israeli officials and 0 Palestinians will say so, regardless of whether the prose itself is flagged for manipulation techniques. Quoted balance is one of the strongest signals of editorial framing, and it's often invisible until you count it.

When an article has a documented backstory — a geopolitical conflict, a financial crisis, a long-running policy fight (healthcare, immigration, voting, guns, drugs), a science or public-health debate, or a civil-rights movement — Biasly surfaces the historical background mainstream coverage tends to omit. Two tabs: a neutral encyclopedic overview, and a "what this article doesn't tell you" view that compares what the piece covers against commonly-skipped events — the 1953 CIA-backed coup, the Budapest Memorandum, the 1999 Glass–Steagall repeal, Shelby County v. Holder, and similar. Every factual claim is cited to a Wikipedia article and the URL is HEAD-validated before it reaches you — no hallucinated links. Routine single-company earnings, sports, and celebrity stories return nothing — no scope creep.

Claude Sonnet 4.6 generates the background as the article is analyzed. A scope test runs first: the article qualifies only if it has a consequential, well-documented, multi-actor history a reader needs to judge the framing — recency or controversy alone isn't enough, so routine business, sports, and celebrity pieces hide the panel entirely. For in-scope articles, Sonnet writes a neutral encyclopedic summary plus a delta-mode "what this article doesn't tell you" section comparing what the piece covers against commonly-omitted events. Responses stream in over about 20 seconds, with the topic and citations appearing in the first 3–5 seconds. Cached for 30 days — re-opening the same article is instant.

The model flags techniques regardless of intent. A satirical column may use loaded language deliberately, and the model will flag it accordingly — that's a feature, not a bug. Treat the article's format as context the reader supplies; the tool tells you what's in the text, not what genre it is.

General-purpose LLMs work for one-off analysis but are inconsistent at scale — the same sentence can get different verdicts on different runs, and per-call cost makes streaming sentence-level scoring impractical. Biasly uses a calibrated transformer for the binary verdict (fast and consistent) and uses Claude only for the human-readable explanation of why a flagged sentence is manipulative.

Yes — like any classifier. Paraphrasing, formal-sounding euphemisms, and unusual sentence structure can all suppress detection. The model is also stronger on news-format prose than on raw social media or transcribed speech with repair disfluencies. Think of it as a critical-reading assistant, not a verdict.

Major newspapers block automated access from server IPs. For paywalled or protected sites, copy and paste the article text directly into the text analysis mode.

A two-stage AI pipeline. First, a fine-tuned transformer model (trained on thousands of labeled examples of biased and neutral text) classifies each sentence. Then, for flagged sentences, Claude AI identifies the specific manipulation technique and explains the reasoning — so you get a precise, readable explanation rather than just a label. The transformer has particularly strong coverage of geopolitical framing, historical revisionism, military euphemisms, and economic populist rhetoric.

Yes. Paste a video URL into the Video tab — YouTube, TikTok, Instagram, Facebook, and X are all supported. We fetch the transcript and analyze it in 1-minute segments, showing timestamps alongside each verdict so you can jump straight to the flagged moments. Videos without captions or transcripts aren't supported yet.

Yes. For URL analysis, we automatically follow pagination links and analyze up to 500 paragraphs across up to 5 pages — including live blogs like the Guardian that use block-based pagination.

English only. Non-English input will produce noisy results.

For each analysis we store a 200-character preview of the input, the verdict, confidence score, list of flagged techniques, and (for URLs and videos) the source link. We don't store full article bodies, and we don't sell or share this data — it's used to monitor model performance and find regressions.

Not yet. The endpoints are public and SSE-streamed if you want to script against them, but there's no documented API contract or rate limits — which means we may change the response shape without notice. A proper API is on the roadmap.

Not currently. Some components may be released over time; the core scoring stack stays proprietary for now.

Free during beta, no account required. Long-term plans aren't fixed — likely a free tier for standard use plus a paid tier for higher volume or API access. The goal is for the basic analyze-an-article case to remain free.