How neutral is Alaska News?

We measure our own newsroom the way we wish someone would measure every newsroom: a panel of three AI models from three different companies reads articles with every outlet name and byline stripped out, scores them against a published rubric, and the results ship here, including the weeks we lose. This page is the instrument's first public output and is labeled accordingly: an experiment, with sample sizes shown next to every score.

Latest run: 2026-10-09. Rubric v0.4.0. Panel: anthropic/claude-sonnet-5, openai/gpt-5.6-terra, google/gemini-3.8-flash.

Jump to: the landscape · the board · side by side · fingerprints · same story, two tellings · access · methodology · plain-English terms

The Alaska media landscape, mapped

Two measurements separate these outlets more than any others: which worldview their framing privileges, and whether official claims get tested or relayed. Plotting one against the other, with each bubble labeled in place, puts every archetype in its place: politically clean but deferent sits bottom-center, tilted but independent upper-left, advocacy far right, and the balanced middle is the middle. Positions are this page's own measured numbers and move with the daily data.

How to read this: left and right is ideological tilt; up is source independence (official claims tested rather than relayed). Bubble size is local articles per day, publishing volume rather than audience reach. Dashed outlines mark the known-lean control and the out-of-market reference; the moose marks our own newsroom. Hover or tap any bubble to spotlight it and see its exact numbers; tap again or click away to release it.

The board, and why each outlet is on it

Membership here is an editorial decision with its reason published, never a traffic formula, and no outlet leaves by quiet deletion: a change to this table is a change someone has to explain. Two rows are not peers by design: the known-lean control exists to prove the instrument can detect lean at all, and the out-of-state reference anchors Alaska scores against a newsroom with no stake in Alaska. Reach context is an annotation, never a gate: each entry is a fact the outlet published about itself or an instrument produced, with its source and date, because search interest, domain rank, and circulation are three different instruments that disagree in informative ways.

Every outlet's fingerprint, drawn

The same lean data as a picture, one panel per outlet, every panel on the identical -3 to +3 scale. This includes the control and reference outlets the comparison table deliberately leaves out, because seeing Must Read Alaska's profile beside the mainstream outlets is the fastest way to see both what the instrument detects and how unusual a flat profile is.

Alaska's top outlets, side by side

Since rubric v0.3, lean is measured on five separate axes, because partisanship and ideological tilt are different variables: an outlet can favor no party while still systematically privileging one worldview's institutions, sources, and framings. Alongside party favoritism, the panel scores ideological tilt, deference versus skepticism toward institutions, conservation versus development framing, and establishment versus populist framing, each per article, each in both directions. Every axis reports two numbers: net (the average, which can hide opposite leans that cancel) and strength (the average magnitude per article regardless of direction, which cannot). Read them together: a net near zero with high strength means the outlet swings both ways on that axis story by story, while net and strength nearly equal means a consistent house tendency. Partisan lean also reports the count of articles registering no lean. Lean is direction only, not quality: a lean-free article can still be badly sourced, which is what the quality dimensions (0 to 4, higher is better) measure. The judged sample covers each outlet's own local journalism (news, politics, business, opinion), classified by the same rule for every outlet: national wire stories, sports, and syndicated lifestyle columns are excluded, so an outlet's numbers describe its journalism rather than its content mix. Evidence fairness rewards scrutiny proportional to evidence rather than equal time, and source independence asks whether official framing was tested or relayed: a dimension this newsroom carries deliberately, because its articles are drafted from public meetings and records, which is exactly where institutional framing capture would hide. Judges may mark a dimension not applicable for genres where it has no meaning (a weather brief is not penalized for lacking opposing sources).

Small samples, honestly labeled: these are first-run numbers, not verdicts. The sample grows weekly, and every score above will move as it does. Alaska Public Media cannot be measured through our current open-crawl method: its advertised RSS feed carries no items, and its robots.txt disallows article pages for all crawlers. The door stays open; a working feed or granted access adds it to this table.

How often a lean is detected at all

Averages can converge while behavior differs: a mild lean in a third of articles and no lean in any article can produce similar means. This table drops magnitude entirely and asks one question per axis: in what share of scored articles did the panel detect any lean at all. It is the most differentiating single view of the same data.

Same story, side by side

When two tracked outlets cover the same story and we hold both texts, the panel also ranks the blinded accounts head to head: the most sensitive comparison there is, because differences that absolute scores smooth over surface when the underlying event is identical. Story pairs are matched by an adjudicator whose one-sentence reason is recorded for every match. 50 same-story comparisons so far; a win is counted only when a majority of the three judges agree.

Small numbers, honestly labeled: head-to-head pairs only exist when outlets cover the same story in the same window, which is rarer than it sounds. Counts accumulate daily.

Access and distribution practices

Deterministic facts, checkable by anyone with a terminal, describing each outlet's publishing choices: these are not neutrality or quality scores, and blocking AI crawlers is a legitimate choice, not a journalistic failing. Full-text RSS means the outlet's journalism arrives without a site visit; the AI column counts crawler user-agents robots.txt turns away entirely.

Our own row is not spotless: Alaska News's feed is excerpt-only where Anchorage Daily News and Alaska Beacon publish full text. Publishing that fact here is the point of the instrument. A performance dimension (Google PageSpeed) joins this table once its API access is set up.

Methodology

How the numbers are computed. Each judge scores each article (averaging that judge's repeat runs), the per-article score is the median across the three judges, and an outlet's number is the mean of its per-article scores, shown to one decimal. Per-article scores and judge disagreement are in the raw data download. One boundary, stated plainly: the panel assesses the text, not the world. It scores whether claims are attributed and whether the article shows its checking; it does not verify facts and produces no factuality rating, because a model reading one article about a brand-new local event has no independent knowledge to check it against, and pretending otherwise would be the exact overreach this page exists to avoid.

Partial blinding. Before any model sees an article, outlet names, bylines, self-references, and boilerplate are stripped by a deterministic pass, and a guard refuses to score any article where a known identity string survives. This removes explicit identity, not every clue: house style, subject matter, or distinctive passages could still hint at the source, so we call it partial blinding rather than claiming the judges cannot know.

The panel. Three models from three vendors (anthropic/claude-sonnet-5, openai/gpt-5.6-terra, google/gemini-3.8-flash), each scoring independently at temperature 0. The headline number is the panel median; disagreement between judges is data we keep, not noise we hide. No single human rater sits above the panel: one person's read is one set of unauditable priors, while the panel's disagreement is published.

Validation checks and limitations

Can the instrument detect lean at all? A judge that scores everything "center" is broken. Every run includes a control outlet with a widely documented political lean (Must Read Alaska, rated right-leaning by independent media-bias raters). Under rubric v0.2 all three models, blinded, read the control at a panel partisan lean of +1.4 across 14 articles while reading the outlets above near 0. Rubric v0.3 restarted accumulation on 2026-09-20, so its control sample is currently 44 articles (reading 1.6), consistent with its v0.2 read and its external ratings. The gate stands: if the control stops registering, publication stops until a person has looked. One known-lean outlet is a necessary check, not a sufficient one: the next validation step is article-level controls in both directions, including neutral articles from partisan outlets and slanted ones from mainstream outlets.

External cross-check. The panel never sees outlet names, yet its first-run reads match the established rater ecosystem: Ground News aggregates Anchorage Daily News as Center and Alaska Beacon as Lean Left, directions our panel's partisan axis reproduced blind under rubric v0.2, with the raters underneath splitting on Beacon the same way our seats did. Convergence with independent human raters, reached blind, is the strongest external evidence this instrument measures something real.

An outside read of this newsroom. In August 2026, Media Bias/Fact Check independently rated Alaska News Least Biased with High factual reporting and High credibility. They credited neutral language and primary-source reporting, the two things these charts measure, and found no failed fact checks. Their one reservation is the model itself: a high-volume, AI-assisted newsroom is new and has not yet shown over time that it avoids the errors automation can introduce, a concern they judged substantially mitigated by our primary records and transparent sourcing. We agree it has to be shown, and this page, scoring a fresh sample of every outlet's coverage every day and publishing the weeks we lose, is how we intend to show it. Human readers and a blinded model panel, two instruments that share nothing, reached the same reading.

Rubric v0.3 (2026-09-20), and why it changed. Story-by-story human audits found that a single left-right axis under-reads outlets whose house style is institutionally neutral: the language carries no partisan tells while the framing still systematically privileges particular institutions and worldviews. v0.3 therefore separates party favoritism from ideological tilt and scores the worldview axes directly. The same version scoped the judged sample to each outlet's local journalism, because a feed that is one-third syndicated advice columns and national wire dilutes every score toward neutral. The axes were added to measure tilt the old instrument was structurally blind to, not tuned toward any expected result: every outlet is scored by the same prompt, still blinded, and scores from different rubric versions are never mixed, so each version's aggregates accumulate fresh from its start date.

Rubric v0.4 (2026-09-29). Two instrument changes, both from measured gaps: the judged text now includes each article's headline (feed-collected outlets were scored on bodies alone while page-fetched ones included theirs, an asymmetry that surfaced when one outlet's partisan headlines sat atop more restrained bodies), and the lean axes gained a magnitude ladder defining what separates a 1 from a 2 from a 3, because the v0.3 score distribution showed judges almost never using the scale's ends without a definition. Every version's final numbers stay published below: a version bump restarts accumulation, never history.

Versions are never mixed in any aggregate; this table exists so the discontinuity at each bump is visible rather than hidden. v0.2 scored whole feeds; v0.3 onward scores each outlet's local journalism only, and v0.4 adds headlines to the judged text, so differences between columns reflect instrument changes as well as the outlets.

Calibration baseline, and where it came from. The panel was first run against a frozen 14-article reference set. Full provenance, because it matters: the reference audit was produced by an AI assistant and reviewed by a human in conversation, and the rubric's scoring anchors were derived from that same audit, so agreement with the seed set shows the judges reproduce a documented read, not independently established truth. Every later week is scored on articles the rubric has never seen, which is the ongoing check that matters. Bias mean absolute error per seat on the seed set (2 runs per seat): anthropic 0.2, openai 0.1, google 0.1. Any change of models or rubric re-scores this set and the discontinuity is published.

Collection ethics. An honestly identified crawler, robots.txt respected, no paywall circumvention, no competitor text republished. Anchorage Daily News and Alaska Beacon are read from the full-text RSS feeds they publish. Scores describe the text, never anyone's intent.

Conflicts, stated plainly. This page grades our own newsroom and is operated by us; the mitigations are the blinding, the cross-vendor panel, and shipping our losses (see the openness table). One panel seat runs a model from Anthropic, whose models also draft Alaska News articles; the other two seats exist so that seat never rules alone.

Raw data. Every number on this page comes from one downloadable JSON (aggregates and sampling frames; no article text). Last updated 2026-10-09; the page and the download update together.

Back to the Alaska News profile · Alaska News

The terms, in plain English

Every dotted-underlined term on this page lands here. One reading rule carries most of it: the lean axes measure direction and the quality scores measure craft, and they must never be mentally combined, because a story can point nowhere and still be built badly, or point somewhere and be built well. On every lean axis the signs are written out below, because a signed number whose direction a reader has to guess is a number the page has not actually published.

Panel

The three AI models (from Anthropic, OpenAI, and Google) that score every article independently. No single vendor's model gets to decide.

Blinded

Outlet names, bylines, and self-references are stripped before scoring, so the panel judges text, not reputations. Partial, because style or subject can still hint at the source.

Rubric

The published scoring rules. When the rubric changes, scores from different versions are never mixed; each version's final numbers stay published in the lineage table.

Sample size (n)

How many articles stand behind a number. Bigger n means more confidence; a small n is an early read, not a verdict.

Net

The average direction across articles. Articles leaning opposite ways cancel, so a net near zero can mean level or can mean both-ways; strength tells you which.

Strength

The average per-article magnitude of lean, ignoring direction. High strength with a near-zero net means the outlet swings both ways; low strength means articles are individually level.

Partisan lean

Party favoritism. Negative = Democratic-coded framing, positive = Republican-coded, 0 = none detected. Requires actual party-relevant treatment, not vibes.

Ideological tilt

Worldview rather than party: whose assumptions and policy logic the framing privileges. Negative = progressive, positive = conservative.

Institutional posture

Whether official and institutional accounts are relayed or tested. NEGATIVE = deference (official accounts carried as fact), POSITIVE = skepticism (institutions treated as interested parties). Alaska News's negative score here means we center and trust official records more than anyone on the board, which we publish as the thing we are working to change.

Conservation-development framing

On resource and land stories: whose costs get quantified. Negative = conservation framing privileged, positive = development framing. Scored only where the topic applies, which is why its n is smaller.

Establishment-populist framing

Whose voice is the default lens on power. Negative = establishment (expert, official, incumbent perspectives), positive = populist (outsider and ordinary-person grievance). Independent of left-right: populism comes in both flavors.

Evidence fairness

0 to 4. Whether claims get scrutiny proportional to the evidence: no manufactured balance for unsupported positions, no suppression of supported ones.

Sourcing

0 to 4. Named sources, linked records, checked numbers, affected stakeholders heard. A single-perspective account of a contested matter caps at 2.

Source independence

0 to 4. Whether interested parties' claims are tested rather than relayed. An article can sound perfectly neutral while adopting an official's framing wholesale; that scores low here. This is Alaska News's weakest score and our active work.

Analysis labeling

0 to 4. Whether interpretation is marked as interpretation. News, analysis, and opinion should never blur inside news copy.

Language neutrality

0 to 4. Whether the wording itself stays out of the way: no loaded adjectives, partisan cues, or smuggled judgment.

Known-lean control

Must Read Alaska's role: a widely known leaning outlet scored by the same blinded instrument. If the panel could not detect its lean, the instrument would be broken; detecting it is what makes everyone's zeros credible.

Out-of-state reference

Nieman Lab's role: a national journalism-trade newsroom with no stake in Alaska, scored identically, so Alaska scores can be read against a non-Alaska baseline.