NewsLens Prototype
About this prototype
Reading
A staged-transparency prototype

Just read about it?
See what the other outlets left out.

NewsLens is a second stop, not a news source. Look up an event you have already read about and see how several outlets framed the same facts: which claims everyone reported, and which showed up in only one or two.

No bias ratings No accuracy verdicts

Just the claims, the sources, and a count you can check by hand.

Research lookup

NewsLens only compares events it has already processed and a human has reviewed. The search looks inside that set. It will not fetch or analyze a new article on demand.

{{ a.tag }}
{{ a.title }}

{{ a.body }}

Available comparisons

{{ resultCountLabel }}
{{ s.categoryLabel }} · {{ s.date }}

{{ s.headline }}

{{ s.dek }}

{{ s.statusLabel }} {{ s.outletCountLabel }} {{ s.actionLabel }} →

No comparison matches “{{ lookupValue }}”. NewsLens only covers events it has already processed and reviewed, so it can't analyze a new article on demand. Try a broader term, or clear the search to see everything available.

NewsLens publishes a fixed set of precomputed stories. Each one is run once, ahead of time, and passes a human review before it appears here. Stories still in the extraction queue are shown but not yet open.
← All stories
In extraction queue

{{ story.headline }}

{{ story.dek }}

This story has not been published yet. NewsLens only opens a comparison after the AI has extracted the claims and a person has checked that work against a fidelity checklist, looking for omitted context, altered meaning, and wrong attribution. Nothing reaches the public page before that review.

Queued · {{ story.outletCountLabel }} · review pending
← All stories
◷

Reviewing this alongside the paper? Turn on Research notes (top right) to see how each design choice maps to the literature, or open About this prototype for the evaluation.
{{ story.categoryLabel }} · {{ story.date }}

{{ story.headline }}

{{ story.dek }}

Comparing {{ story.outletCount }} outlets: {{ o.name }}

How the outlets framed it

{{ f.outlet }} “{{ f.title }}”
Paper · Introduction Same numbers, opposite spin: Fox Business led with “steady” while ABC led with “worse than expected.” That contrast opens the paper. NewsLens shows the difference in framing but does not label either outlet, which guards against the hostile media effect (Vallone et al., 1985).

As you read, ask

{{ q.n }} {{ q.text }}

What did everyone agree on?

Claims carried by all {{ story.outletCount }} outlets, listed first. Counted, not judged.

{{ a.countLabel }} {{ a.text }}

Where did they split?

One claim up close. Each square is one outlet: filled means it reported the claim, empty means it did not.

{{ story.split.text }}
{{ story.split.countLabel }}
Reported by {{ story.split.reported }}. Not carried by {{ story.split.notReported }}.
reported the claim did not

Who left what out?

One card per outlet, showing how many of the {{ story.claimTotal }} claims it did not mention. A lower count is not a better outlet: a three-minute radio segment carries fewer claims than a nine-hundred-word report by format, so the article type sits next to every count.

{{ om.name }}
{{ om.countLabel }}
{{ om.type }}
An earlier version showed third-party outlet bias ratings here. A persona walkthrough found that placing a “Lean Right” label next to an omission count broke the tool’s neutrality, so the ratings were removed. NewsLens shows the count, not a verdict.
Paper · Findings This is the section where the cognitive walkthrough found the design’s biggest flaw: a third-party “Lean Right” label sat next to a count of what an outlet left out, which broke the tool’s neutrality promise. Removing the ratings from the expanded toggle is the central finding of the paper (Kizilcec, 2016).

The claims, side by side

{{ story.claimCountLabel }}

Each row is one factual claim and a count of how many of the {{ story.outletCount }} outlets reported it. Open a row to read the exact source sentence, or contest it if the grouping looks wrong.

Paper · Discussion The count is plain arithmetic you can redo from the sources, with no AI-written rationale for the grouping (Rudin, 2019). The fullest explanation appears inside the contest flow, at the point where a reader disagrees (Kizilcec, 2016). One risk to watch is the machine heuristic, where a clean “6 of 6” gets trusted too readily (Sundar, 2020).

{{ c.text }}

Human-review note attached
{{ c.countN }} of {{ c.countTotal }} outlets
{{ c.toggleLabel }} {{ c.chevron }}
Reported by: {{ c.reportedNames }}
{{ src.outlet }}
The original article links are illustrative sample data in this prototype, so they do not open a live page. In a deployed version each source would link to the outlet's published article.
Human-review note{{ c.note }}
Disagree with the grouping or attribution? Flag it for a human reviewer.
Before you flag it, here is how this grouping was made

{{ c.groupingExplanation }}

{{ g.n }} {{ g.text }}
Still looks wrong? Tell us what.

Pick the closest issue. Your flag goes to a person, and it does not change the public page on its own.
✓
Flag received. Queued for review as {{ contestTicket }}

A human reviewer will check this claim against the fidelity checklist. Nothing on the public page changes until they act. Thanks for keeping the box honest.
Glass box

How NewsLens compared these articles

{{ explanationToggleLabel }}

These five sections cover how the comparison was put together, written for a reader rather than an engineer.

{{ e.label }} {{ e.body }}

This panel explains what the system did and why. It cannot answer "what would change if I gave it a different article," since NewsLens runs on a fixed, precomputed set.

These method pages are not published yet, so the links do not go anywhere. This is a known gap in the current build, not a broken feature we are hiding. Turn on Research notes for the paper reference.
Paper · Methodology + Findings The disclosure aims for what a reader can actually act on, not only a technical account of the model (Shin, 2021). Staged mode keeps this panel closed; expanded mode opens it. Known gap: the “documented methods” links do not resolve yet, so Amershi et al. (2019) Guideline 11 is only partly met.

Sources & method

Which outlets, and why

{{ story.outletRationale }}

Reading the counts: “{{ story.outletCount }} of {{ story.outletCount }}” means all {{ story.outletCount }} of these outlets. It says nothing about the rest of the press. The sample is intentionally small.
Model card · limitations
{{ l }}
Out of scope by design
✕ {{ x }}
Pending
Correction log not yet published

The first human review pass on this story was not logged, so there is no public error rate yet. Future stories will record every correction from the start and publish the rate here. This gap is disclosed rather than hidden.
← All stories
About this prototype

A design-based capstone, shown as a working product

NewsLens is the artifact for an HCAI 4304 capstone on staged transparency in AI journalism. This page is the research layer. It explains what the prototype is for, how it was built and evaluated, and where it still falls short. The story pages themselves stay in product voice and keep this material out of the reader's way.

This live build is a design iteration completed after the July 19, 2026 paper. The artifact evaluated in the paper is the prior build; the visual system and story set here evolved under the same research-through-design method. The core findings, staged disclosure and the removal of third-party bias ratings from the expanded toggle, are preserved. The jobs-report story uses the verified corpus with verbatim source sentences; the other sample stories are illustrative.

Two layers, one artifact

Layer 1 · the product

Every story page reads like a shipping product. Plain language, no citations, no jargon. A reader never sees an academic reference.

Layer 2 · research notes

The Research notes toggle on any story page adds margin annotations that tie each design choice back to the literature. Off by default. This is the view for reviewing the work against the paper.

The idea being tested

The research question: can staging AI transparency earn a reader's trust without overloading them? The design rests on two working principles.

Path to Transparency

Readers grant basic trust through clean design first, then transparency is revealed in stages rather than dumped at once (Kizilcec, 2016; Shin, 2021).

10-80-10 division of labor

A human sets direction (first 10%), AI does the volume of extraction and code (middle 80%), and a human reviews and takes responsibility (final 10%) (Mollick, 2024).

Three iterations

The build itself was the research method (research through design). The story page moved through three versions.

{{ it.v }}
{{ it.title }}

{{ it.body }}

How it was evaluated

Three analytical methods, all run by the designer. No live user study was conducted, so no measured effects on real users are claimed.

{{ e.stat }}
{{ e.label }}

{{ e.body }}
Persona · cognitive walkthrough
Skeptical partisan reader

Chosen because prior trust level changes how a reader reacts to transparency cues (Shin & Park, 2019). This walkthrough exposed the AllSides bias-label flaw described below.

Persona · cognitive walkthrough
News-literacy educator

A teacher looking for a classroom artifact (Tully et al., 2020). This walkthrough validated the guiding questions and the always-open explanation panel.

The finding that changed the design

Expanded mode turned out to bundle three separate changes. Opening the source panels and the explanation panel both helped. But it also switched on third-party AllSides bias labels, placing a "Lean Right" tag next to a count of what an outlet left out. Walking the skeptical reader through that screen showed it broke the tool's neutrality promise and made the whole site look untrustworthy (Kizilcec, 2016). The ratings were removed from the toggle in the final version.

Known gaps in this build

Disclosed rather than hidden, consistent with the glass-box approach.

{{ g }}

All figures in the story pages are illustrative sample data for the prototype, not a live extraction. Citations here refer to the accompanying paper's reference list.