VeriAI

Inspiration

AI tools can generate answers in seconds, but a convincing answer can still contain incorrect, unsupported, or outdated factual claims.

We were inspired by a simple question: “Instead of asking whether an entire AI answer is trustworthy, why not check its individual claims against evidence?”

That idea led us to build VeriAI, an evidence-grounded AI claim auditor designed to make AI verification more transparent and easier to understand.

What it does

VeriAI takes an AI-generated answer and breaks it into individual factual claims.

For each claim, it:

1. Extracts an atomic factual assertion.
2. Retrieves relevant evidence from the web using Google Search grounding.
3. Compares the claim against the retrieved evidence.
4. Classifies the claim as SUPPORTED, CONTRADICTED, or UNCERTAIN.
5. Shows the evidence, source, and explanation behind the decision.

A key principle of VeriAI is that UNCERTAIN does not mean FALSE. If sufficient reliable evidence cannot be found, the system does not automatically call the claim false.

How we built it

We built VeriAI as a lightweight full-stack web application.

The frontend uses React, TypeScript, and Tailwind CSS to provide a simple interface for submitting AI-generated answers and reviewing claim-level results.

The backend uses Node.js, TypeScript, and Express. It connects to Gemini 3.8 Flash with Google Search grounding for claim extraction, evidence retrieval, and claim-versus-evidence analysis.

The verification pipeline is:

AI Answer → Claim Extraction → Evidence Retrieval → Claim Analysis → Verdict + Evidence

We also kept the Gemini API key on the server side instead of exposing it in the browser.

Challenges we ran into

One of the biggest challenges was distinguishing between a claim that is contradicted and a claim for which we simply could not find enough evidence.

Another challenge was claim decomposition. A single sentence can contain multiple factual assertions, so treating the whole sentence as one claim can produce misleading results.

We also had to consider current information, search relevance, conflicting sources, and incomplete web evidence.

These challenges influenced the design of VeriAI's three-way verdict system: Supported, Contradicted, and Uncertain.

Accomplishments that we're proud of

We are proud that VeriAI can analyze mixed AI answers containing correct, incorrect, and unsupported information and evaluate the claims separately.

During testing, the system successfully identified clearly false claims as CONTRADICTED while keeping obscure claims with insufficient evidence as UNCERTAIN.

We are especially proud of building a working end-to-end prototype rather than only a concept: users can submit an AI answer and receive an evidence-grounded verification report with sources.

What we learned

We learned that AI verification is not simply about finding a web page that contains similar words.

Reliable verification requires careful claim decomposition, relevant evidence retrieval, comparison between the claim and evidence, and honest handling of uncertainty.

We also learned that a good AI system should communicate its limitations instead of presenting uncertain information with false confidence.

What's next for VeriAI

Next, we want to improve source ranking, handle conflicting evidence more intelligently, strengthen verification of time-sensitive claims, and evaluate the system against larger factuality benchmarks.

Our long-term goal is to make AI-generated information easier to inspect, question, and verify.

VeriAI: Don't blindly trust an AI answer. Check what the evidence says.

Built With

• ai
• checking
• css
• express.js
• fact
• gemini
• generative
• google
• node.js
• react
• safety
• search
• tailwind
• typescript
• web