Skip to content

Limits

What an AI Fake News Checker Can and Cannot Do

An honest look at what an AI fake news checker like WizeUp can do, what it cannot, and how AI-assisted claim review differs from human fact checking.

What an AI Fake News Checker Can and Cannot Do preview

Written by the Fulldive product engineering team, based on direct inspection of the repositories listed below.

An AI fake news checker is a useful tool for slowing down before you share a claim. It is not a truth machine. This post explains, specifically and without hype, what an AI-assisted claim checking app like WizeUp can do today, what it cannot do, and why framing matters more than most product copy in this category admits.

What the feature actually is, in code

In the fulldiveVR/fulldive-android-apps repository, the AI feature that ships inside WizeUp is called AI Chat. Commit 3e90b710c3 refactored it in February 2026. In the codebase, AI Chat is not a standalone chatbot. It is a surface attached to trends and article context, so a reader who has opened a trending story can ask questions about the claim in front of them.

That scoping matters, because it constrains both the strengths and the limits of the feature. AI Chat is good at:

  • Summarising a long article or thread into the specific claim being made.
  • Pointing out which parts of a claim are testable and which are framing.
  • Suggesting angles or counter-questions a reader might not have thought of.

It is not good at:

  • Returning a binary “true” or “false” verdict with confidence.
  • Independently verifying recent events outside its training or retrieval window.
  • Replacing a human editor or a primary source.

This post is the “Limits” entry of our WizeUp series. The companion posts (how Fulldive Browser became WizeUp, the checking-news-before-sharing workflow, and personalized feeds, trends, and daily digest) cover the features positively. This one is deliberately cautious.

What a human fact-check actually is

Before talking about what AI can do, it helps to remember what professional fact-checkers do. The International Fact-Checking Network (IFCN) publishes a code of principles that signatories commit to. The headline commitments are:

  1. Non-partisanship and fairness.
  2. Transparency of sources.
  3. Transparency of funding and organisation.
  4. Transparency of methodology.
  5. An open and honest corrections policy.

A human fact-checker reads the claim, locates primary sources, asks the original parties for comment where possible, publishes a rating with evidence, and corrects the record when new information arrives. Google’s Fact Check Tools index those published checks in ClaimReview structured data so other surfaces can cite them.

An AI claim-review feature running inside a mobile app is not that. It is a reading aid sitting on top of a language model, on a phone, talking to one user. Confusing the two is the single biggest risk in this category, and it is the reason our in-repo publishing notes explicitly warn: do not claim fake-news detection is definitive.

What AI-assisted claim review can honestly do

Within those limits, AI assistance is genuinely useful. Here are the capabilities we are comfortable claiming for WizeUp:

  1. Surface the specific claim. A trending social post or article often mixes a factual claim with opinion and framing. AI Chat can extract “the person in this video says X happened on Y date” so the reader knows what to check.
  2. Suggest search angles. Given a claim, the model can propose queries, primary-source types (court records, company filings, official statements, original video), and publications likely to have covered the topic.
  3. Spot inconsistency. If a claim contradicts the article body, or if two trending posts on the same event disagree on a specific detail, AI can surface the discrepancy.
  4. Encourage slowing down. Interaction itself — reading the model’s summary, writing a follow-up — adds friction before sharing. That friction is the point. It is the job-to-be-done described in checking news before sharing.

None of these require the app to “know the truth”. They work by making a reader’s own review faster and more structured.

What it cannot do, and why

It cannot verify recent events it did not retrieve. A language model has a training cut-off and whatever retrieval context the app gives it at runtime. A breaking claim made an hour ago is very likely outside that context. The right behaviour is to say so, not to guess.

It cannot rate sources reliably on its own. Source reputation changes, and models can be confidently wrong about which outlets are careful on which topics.

It cannot produce a stable verdict across runs. Temperature, retrieval variance, and prompt drift mean the same claim can get different answers in different sessions. Professional fact-checks, in contrast, are published with a static rating and a correction log.

It cannot replace a primary source. If a claim is important enough to share, it is important enough to check against the original document, video, or statement. AI can help the reader find that source. It cannot substitute for it.

Limits

To be explicit, since this entire post is a limits post:

  • WizeUp is not a guaranteed hoax checker. Avoid treating any app — ours or anyone else’s — that way.
  • Analytics around trend pages and pushes (commits 0536b97ea8, 569681c230 elsewhere in the repo) are product instrumentation. They tell us whether people are using the feature. They do not tell us the feature was correct.
  • We do not currently publish ClaimReview structured data for AI Chat outputs, because those outputs are not editorially reviewed the way a professional fact-check is. Treating them as if they were would be the same overclaim we are warning against.
  • AI Chat works best when the underlying article or trend has primary sources already in the reader’s locale. It works less well on very new claims, very local claims, and claims that exist only in screenshots.

How to use WizeUp responsibly

If you have WizeUp installed (see /project/fulldive-browser/ for install notes), a responsible workflow looks like this:

  1. Open the trend or article that made you pause.
  2. Use AI Chat to summarise the specific claim — not the framing, the claim.
  3. Ask AI Chat where you might verify it. Treat those suggestions as starting points.
  4. Open at least one primary source directly. If the claim is about an official statement, read the statement. If it is about a number, read the filing or dataset.
  5. Only then decide whether to share, and what caveats to add if you do.

That is the workflow described at more length in checking news before sharing with WizeUp. It is also the workflow that lets you enjoy a claim checking app without overtrusting it.

Where this sits in the broader Fulldive ecosystem

WizeUp is part of a longer Fulldive pattern of shipping small, focused Android apps rather than one monolithic app. The Fulldive company story describes that philosophy. In the context of claim review, it means WizeUp is the AI-assisted reading surface, not a full newsroom. That scope is a feature, not a bug. A small tool with honest limits is more useful than a large tool that overpromises.

If you have questions about our approach — especially about what we are not claiming — we read mail at support@fulldive.com and the FAQ covers common concerns.

Sources

Last updated: 2026-04-16. Commit hashes and version numbers are drawn from Fulldive repositories inspected on 2026-04-13.